<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Aliza Samir Khoja on Medium]]></title>
        <description><![CDATA[Stories by Aliza Samir Khoja on Medium]]></description>
        <link>https://medium.com/@khoja.aliza?source=rss-5307378d07b3------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*UxwvrmlpkVhVcVCm_UGHJA.png</url>
            <title>Stories by Aliza Samir Khoja on Medium</title>
            <link>https://medium.com/@khoja.aliza?source=rss-5307378d07b3------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 19:44:19 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@khoja.aliza/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[PRECISE Pro: Building a Brain That Watches Itself]]></title>
            <link>https://medium.com/@khoja.aliza/precise-pro-building-a-brain-that-watches-itself-fe6e3b1b7edf?source=rss-5307378d07b3------2</link>
            <guid isPermaLink="false">https://medium.com/p/fe6e3b1b7edf</guid>
            <category><![CDATA[youth-in-tech]]></category>
            <category><![CDATA[proactive-healthcare]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[concussions]]></category>
            <category><![CDATA[building]]></category>
            <dc:creator><![CDATA[Aliza Samir Khoja]]></dc:creator>
            <pubDate>Sat, 23 May 2026 15:15:37 GMT</pubDate>
            <atom:updated>2026-05-25T01:08:37.459Z</atom:updated>
            <content:encoded><![CDATA[<blockquote>“The brain rarely announces its own decline. It compensates, adapts, and reroutes. By the time you feel it, the drift has been happening for weeks.”</blockquote><p>That sentence sat with me for a long time after I first read it (paraphrased from original) buried in a 2008 cognitive neuroscience paper on intra-individual variability. And it completely reoriented how I thought about brain health.</p><p>Not just academically. Personally.</p><p>I had been building PRECISE Mini for months before I stumbled onto that paper. Mini was already functional. People were using it. I was proud of it. It tracked reaction time, working memory, inhibitory control, and decision speed across repeated sessions. It exported CSVs. It compared your current session to your personal baseline. It was, by any reasonable measure, more scientifically careful than most brain health apps on the market.</p><p>And then I read that sentence, and I realized Mini was still answering the wrong question.</p><p>Mini asked: “Are you performing worse than usual today?”</p><p>But the paper I was reading, <a href="https://www.taylorfrancis.com/chapters/edit/10.4324/9780203837665-10/intraindividual-variability-cognition-aging-david-hultsch-esther-strauss-michael-hunter-stuart-macdonald?context=ubx&amp;refId=11abe9ce-aceb-4f8c-af58-61a3b0ffc689">Hultsch and colleagues from the University of Victoria</a>, was pointing at something deeper. They were tracking older adults over years, measuring their reaction time variability across hundreds of sessions, and they found that the people with the most erratic, unstable reaction time patterns showed dramatically faster cognitive decline in subsequent years, even when their average reaction time scores were completely normal. The variability itself was the signal. The instability was the warning. And it appeared years before anything else.</p><p>That finding <em>broke </em>something open in how I understood the problem.</p><p>The question I needed to be asking was not “are you slower today?” The question was “is your brain becoming less stable over time?”</p><blockquote>Those are completely different questions. They require completely different tools.</blockquote><p>And the second question, the harder one, the more honest one, is the one that actually matters for early detection.</p><p><strong>PRECISE Pro is my attempt to build a tool that answers that second question.</strong></p><p><strong>This article is the full story of how I built it: the science that motivated every decision, the engineering failures that taught me the most, the personal moments where I almost gave up and the ones where something clicked, and the reasoning behind every component of a platform I spent months building and rebuilding until it was right.</strong></p><p>Check out PRECISE Pro: <a href="https://precisepro-alizasamirkhoja.netlify.app/">https://precisepro-alizasamirkhoja.netlify.app/</a></p><p>Check out PRECISE Mini: <a href="https://precisemini-alizasamirkhoja.netlify.app/">https://precisemini-alizasamirkhoja.netlify.app/</a></p><p><strong>I am going to hold nothing back. If you want a polished summary, this is the wrong article. If you want to understand what it actually takes to build something like this, keep reading.</strong></p><h3><strong>TABLE OF CONTENTS</strong></h3><ol><li><a href="#d8db">The Problem with Brain Health Right Now</a></li><li><a href="#db46">What PRECISE Is and What It Refuses to Be</a></li><li><a href="#5d58">The Staged Architecture: Why PRECISE Pro Comes After Mini</a></li><li><a href="#7f3e">The Anti-Saccade Task: What It Measures and Why It Works</a></li><li><a href="#b4cc">The Verbal Fluency Task: Speech as a Neural Signal</a></li><li><a href="#6d4f">The Science of Intra-Individual Variability</a></li><li><a href="#c1bd">Engineering the Browser: Timing, Latency, and Precision Under Constraints</a></li><li><a href="#f104">The AI Layer: Linear Regression as a Deliberate Choice</a></li><li><a href="#e496">The PRECISE Intelligence Engine: Building an On-Device Inference System</a></li><li><a href="#aecd">The Population Norms: Where the Numbers Come From</a></li><li><a href="#1aa5">Signal Architecture: How PRECISE Pro Processes What It Captures</a></li><li><a href="#bf51">Anti-Sandbagging: Why You Cannot Fake This</a></li><li><a href="#a866">Why I Chose These Specific Tasks</a></li><li><a href="#69bd">What PRECISE Pro Is Preparing For</a></li><li><a href="#69a0">The Bigger Vision: From Reactive to Dynamic</a></li><li><a href="#6a8e">What Comes Next</a></li></ol><h3>1. The Problem with Brain Health Right Now</h3><p>I want to tell you about a conversation I had with a clinician who has been in the space for many years and is currently pursuing her PhD. I reached out to her through LinkedIn after reading a couple posts where she spoke about concussion recovery in adolescent athletes. She was kind enough to give me thirty minutes of her time.</p><p>I asked her: how do you currently monitor a player’s cognitive recovery between clinic visits?</p><p>She paused before answering. Then she said something like: “Honestly? We rely on how they say they feel, and we use a standardized test called ImPACT that they take at the beginning of the season as a baseline. If they get a concussion, we retest and compare. But between visits, we have very little visibility into what is actually happening.”</p><p>I asked how often they did that retest after a concussion.</p><p>“Typically at the one-week mark, then again at return-to-play clearance.”</p><p>Two data points. One at the moment of injury assessment. One at the end of recovery. Everything in between is dark.</p><p>She framed this without criticism of the system. She said it with the quiet resignation of someone who has worked within those constraints for years and knows no better option currently exists. The tools for anything more frequent are simply absent. ImPACT requires an administrator, a structured testing environment, and takes 25 to 30 minutes. The Berlin Consensus Statement on concussion in sport, published by McCrory and colleagues (2017) in the British Journal of Sports Medicine, reflects this reality: the international standard of care is organized around pre-season baseline testing followed by post-injury snapshot assessments, with the monitoring gap between those snapshots largely unaddressed. You cannot run ImPACT every three days. You cannot run it on a Tuesday at home when a player feels slightly off. It is episodic by design.</p><p>That conversation crystallized something I had been circling for months.</p><p>The current standard workflow for cognitive health assessment follows a very predictable pattern. Something goes wrong. You or someone around you notices. If everything lines up correctly, you eventually get a referral to a neuropsychologist. A trained professional administers a battery of standardized tests. You score against population norms built from thousands of other people. A report is generated. Weeks later, a conclusion arrives.</p><p>This system has genuine value. Standardized neuropsychological testing is the product of decades of rigorous science. The professionals who administer these tests are extraordinarily careful. I want to be unambiguous about that. Clinical neuropsychology is working as designed. It is doing exactly what it was designed to do.</p><p>But the design itself has a structural problem that cannot be fixed from within its own assumptions. And that problem has three components.</p><p><strong>The reactivity problem.</strong> The brain is assessed after injury. After a threshold is crossed. After the decline has accumulated to the point where it is obvious enough to warrant a clinic visit. There is no continuous measurement. There is no early warning. There is no “your brain is behaving unusually relative to how it was behaving last Tuesday.” The entire system is organized around detection after the fact, which means by the time assessment happens, a meaningful window for early intervention has already closed.</p><p>This is documented reality. Weir, Jackson, and Sonnega (2018), in a large cohort study of former NFL players published in JAMA Neurology, documented high rates of cognitive and neurobehavioral problems emerging years after retirement, with many players reporting that their difficulties had accumulated gradually and outside any formal monitoring window. The slow drift from functioning to struggling happened between checkups, invisible to the systems designed to catch it. That is the reactivity problem at its starkest.</p><p><strong>The population comparison problem.</strong> Your score on any given cognitive task is interpreted relative to what thousands of other people score on the same task. On the surface, this seems reasonable. In practice, it introduces an enormous amount of noise, because brains vary enormously between people. A reaction time of 380 milliseconds might be flagged as “below average.” But what if your baseline is 290 milliseconds? What if 380 milliseconds represents a 31% decline from your personal normal? Population norms would classify you as mildly slow and move on. Your own baseline would classify a significant event worth investigating.</p><p>These are opposite conclusions drawn from the same number. One says “you are fine.” The other says “something has changed.”</p><p>The cognitive science literature has documented this problem for decades. Rabbitt and Abson (1990), in the British Journal of Psychology, showed that within-person fluctuations in cognitive performance carry more diagnostic information than cross-sectional comparisons to population averages. Nesselroade (2001), writing in the European Psychologist, built the theoretical case even more explicitly, arguing that psychology’s overemphasis on between-person differences had systematically underused the information available in within-person change. The case for personal baselines over population averages has been strong in the academic literature for decades. It has simply failed to penetrate standard clinical practice, partly because the tools for continuous personal baseline tracking have only recently become feasible.</p><p><strong>The frequency problem.</strong> Clinical cognitive assessment is built for episodic evaluation, full stop. You go in twice at most in a typical monitoring cycle. But cognition is dynamic. It shifts daily based on sleep, stress, hormones, training load, emotional state, nutrition, hydration, and dozens of other variables. A system that measures at two points in time captures two snapshots. Trends remain invisible. The signal that would emerge from continuous monitoring, the gradual drift or the sudden shift, is simply not accessible to a system designed around infrequent high-stakes assessments.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cipu0Shn4RQ2j2J-zgGLCg.png" /><figcaption>The Precision Gap: Why 24/7 Monitoring is the New Gold Standard for Brain Health (Image Credit: Gemini)</figcaption></figure><p>PRECISE Pro is built around a different premise entirely. The premise is that your brain changes before you feel it, before anyone around you notices it, and before any clinical threshold is crossed. Those changes are small. They are personal. They are dynamic. But they are measurable, if you are measuring the right things, at the right frequency, against the right reference.</p><p>The right things are cognitive signals that are sensitive to neural state changes. Reaction time variability. Inhibitory control accuracy. Speech production latency. Verbal fluency output rate. Fatigue patterns across trials.</p><p>The right frequency is repeated. Weekly at minimum. Daily if the use case demands it.</p><p>The right reference is you. Your own baseline. Your own normal variability range. Your own trajectory across time.</p><p>Everything in PRECISE Pro is designed to answer that question. And understanding why requires understanding what I had to learn, rebuild, and sometimes completely throw away to get here.</p><h3>2. What PRECISE Is and What It Refuses to Be</h3><p>Naming a project is a specific kind of decision. It signals what you think the project is about.</p><p>I went through six working titles before landing on PRECISE. The early ones were embarrassing in retrospect. “CogTrack.” “BrainBase.” “NeuralMon.” I am grateful none of those made it to a public article. They were product names. Marketing names. Zero science in the label. They described what the tool did without saying anything about what it believed.</p><h4><strong>PRECISE stands for Predictive Real-time Evaluation of Cognitive Instability using Speech and Executive-functioning.</strong></h4><p>That acronym was reverse-engineered to spell PRECISE because the word itself describes the philosophy: exactness, within-person specificity, signal clarity. But every word in the expansion earns its place.</p><p><strong>Predictive</strong>, because the goal is to surface signals before they become symptoms. <strong>Real-time</strong>, because the assessment happens live, in real time. <strong>Evaluation</strong>, because it is a monitoring tool, a monitoring tool. <strong>Cognitive Instability</strong>, because instability, distinct from aggregate decline, is the target signal. Using <strong>Speech and Executive-functioning</strong>, because those are the two modalities PRECISE Pro captures.</p><p>The platform is a three-layer system. PRECISE Mini is the entry layer: lightweight cognitive micro-tasks, two to three minutes, designed for frequent natural engagement, focused on establishing personal baselines and measuring whether you are stable relative to your own history. Mini tracks simple reaction time, choice reaction time, working memory span, and inhibitory control through a Go/No-Go task. Mini’s purpose is behavioral validation and baseline formation.</p><p><strong>PRECISE Pro</strong> is the expansion layer, and it is the one this article is about. It increases resolution, adds a second modality, introduces real AI modeling, and begins to behave like a measurement instrument rather than a casual tracking app. Pro adds the anti-saccade visual task and a five-task verbal fluency speech module. It runs live linear regression during the session and hosts the PRECISE Intelligence Engine for post-session analysis.</p><p><strong>PRECISE Max</strong>, which I am currently designing, is the intelligence core: a transformer-based multimodal fusion architecture that integrates Mini and Pro signals into a coherent longitudinal model. Max is where the real predictive capability + coolness will live. It is also the part that requires the most data to validate, which is why Mini and Pro come first.</p><p>Before going further, I need to be direct about what PRECISE Pro is and what it refuses to be, because the boundary matters.</p><p><strong>PRECISE Pro</strong> is a research tool for personal cognitive monitoring. It uses published, peer-reviewed population norms to contextualize the data it collects. It employs real AI modeling. It captures real signal. The methodology is grounded in clinical neuropsychology literature that goes back four decades.</p><p>PRECISE Pro is a research prototype. It is incapable of diagnosing conditions. It is incapable of replacing clinical assessment. It is a monitoring and research tool, full stop. Every result it produces should be understood as exploratory and longitudinal, valuable over time and across sessions, far less meaningful as a single snapshot.</p><p>I want to be clear about this boundary for one specific reason: <strong>overstating what a tool like this does is genuinely harmful.</strong> Period. Point over. If someone interprets a session result as a diagnosis, acts on it without appropriate clinical consultation, and something goes wrong, that is a serious failure. The scientific literature underpinning PRECISE Pro is real. The tasks have clinical validity in controlled settings. But a browser-based platform built by a teenager without official institutional oversight has real limitations, and those limitations are documented transparently in every corner of the platform. The disclaimers are genuine science communication. They are part of the scientific design.</p><p>PRECISE Pro is honest about what it is. That honesty is baked in from the architecture outward, embedded outward rather than footnoted afterward.</p><p>I also want to say something about the why that is personal, because I think it matters for understanding what kind of project this is.</p><p>I am building PRECISE, and I want to be clear about what drives that. I am building it because I had an idea that I could not let go of, because the problem felt genuinely unsolved in a way that had real consequences for real people, and because TKS gave me a framework for turning that idea into something rigorous rather than something superficial. My director Azar has pushed me hard throughout this process, harder than I often wanted to be pushed, to pressure-test assumptions, to read more deeply, to be willing to throw out months of work when the foundation was not solid. That pressure is part of why PRECISE Pro exists in the form it does rather than in the form I originally imagined.</p><p>The mentors and practitioners I have spoken with and have connected with through LinkedIn, have shaped the scientific thinking in ways that are invisible in the final product but completely visible in the decisions it is built on. Building in public, building through feedback, building with honesty about limitations, those are the actual operating principles. They are why this article exists in the form it does.</p><h3>3. The Staged Architecture: Why PRECISE Pro Comes After Mini</h3><p>I originally planned to build PRECISE Pro first.</p><p>It sounds counterintuitive in retrospect, but when I first conceptualized the project, I went straight to the ambitious version. Speech analysis. AI modeling. Multiple modalities. The whole picture. I had been reading about multimodal cognitive assessment for weeks and I was convinced I understood what needed to be built. I started drafting the architecture for what is now Pro before I had built anything at all.</p><blockquote>My director Azar stopped me in one of our early check-ins. The question he asked was simple: “How do you know people will actually use this repeatedly?”</blockquote><blockquote><strong>And I had no good answer.</strong></blockquote><p>He pointed out that everything interesting about PRECISE depends on longitudinal data. Intra-individual baselines require repeated sessions. Variability modeling requires a time series. Trend detection requires a history. If nobody uses the platform more than once, the entire scientific premise collapses. Before building the complex version, I needed to prove that people would engage repeatedly with the simple version.</p><p>That insight is what gave birth to the staged architecture: Mini first, Pro second, Max third. Each layer earns the right to exist by validating the assumptions the next layer requires.</p><p>Mini was designed to answer three specific questions. First: will people complete these tasks repeatedly over time? Second: do the signals stabilize into a reliable personal baseline with consistent use? Third: do the signals respond to real-world variables in ways that match what the literature predicts? If Mini could answer yes to all three, then the foundation for Pro was solid.</p><p>The process of building Mini taught me things I could only have learned by doing. I learned, for example, that browser-based cognitive testing has an enormous engineering problem that most people writing about it do not mention: timing inconsistency. I rebuilt the timing engine in Mini three times before the data looked clean. I discovered that Date.now() has a resolution of 1 millisecond in theory but substantially higher variance in practice, that DOM insertions introduce rendering latency that corrupts RT measurements, that input events are queued in ways that add variable delay between a keypress and the timestamp it receives.</p><p>Those lessons in Mini were what made the Pro engineering decisions possible. I knew exactly what to stay away from. I knew where the noise came from. I knew which browser APIs to use and which to avoid. Building Mini was the prerequisite for Pro, full stop. It was the prerequisite.</p><p>There is also a regulatory and scientific credibility argument for the staged approach that I think about a lot. A platform that launches with claims about AI-powered concussion detection before it has validated its engagement model, its signal stability, or its norm library is building on sand, trust me when I say this. The attention it might attract in the short term would be disproportionate to its scientific foundation, and any scrutiny from someone with real domain expertise would immediately surface the gaps.</p><p>PRECISE Mini built a foundation I could defend. PRECISE Pro extends it in ways that are grounded, specific, and honest about their limitations. That credibility compounds over time. It is slower, but it is the right way to build something in a space where the stakes are real.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/773/1*0iV5VsT-0dHsxBWnBitAmA.png" /><figcaption>Staged architecture: each layer validates the assumptions required for the next.</figcaption></figure><p>The TKS Focus framework maps onto this structure almost exactly. The framework is Learn, Replicate, Replicate, Create. Mini was Replicate 1: taking existing, validated cognitive tasks (simple reaction time, choice reaction time, Go/No-Go, digit span) and rebuilding them with real signal integrity in a browser, adding personal baseline tracking and longitudinal CSV export. This was careful replication. It was careful replication of known paradigms with an engineering twist.</p><p>Pro is Replicate 2 with structured extension. The anti-saccade paradigm is a validated clinical task. Verbal fluency is a validated clinical paradigm. I am replicating them, building them into a browser environment with the timing precision and signal architecture I learned from Mini, and extending them with AI modeling and the Intelligence Engine. The novelty in Pro is the combination and the implementation, not the underlying tasks.</p><p>Max will be Create: the original multimodal fusion architecture that synthesizes Mini and Pro signals into something original in quite this way before. That originality only becomes defensible because the components underneath it are grounded in decades of published science.</p><p><strong>Each layer earns the right to exist by validating the assumptions of the one before it. That sentence is the cleanest summary of the architectural philosophy I can offer, and I want it to land, because it is the organizing principle for everything that follows.</strong></p><h3>4. The Anti-Saccade Task: What It Measures and Why It Works</h3><p>The first time I ran the anti-saccade task I had built, I thought it was broken.</p><p>I had spent two weeks reading the paradigm literature, implementing the stimulus presentation logic, calibrating the jitter-resistant inter-trial intervals, and setting up the response capture system. I ran the first test session myself. A green circle appeared on the left side of the screen. I knew the task: click the opposite side, click right. I clicked left. Then I caught myself, clicked right, too late.</p><p>I ran it again. Same thing. The circle appeared, and before I had consciously decided anything, my hand had already moved toward it.</p><p>I sat there for a moment genuinely confused. I knew the task. I had literally written the code for the task (with of course my amazing vibecoding agent ChatGPT). I knew the circle was going to appear and I knew I was supposed to click away from it. And I still moved toward it first.</p><p>That was the moment the anti-saccade paradigm became real to me. What I had understood intellectually, that there is a hardwired subcortical response that orient toward salient stimuli and that suppressing it requires active prefrontal engagement, had just demonstrated itself vividly in my own nervous system. My understanding of the task design shifted from technical to visceral in about forty-five seconds.</p><blockquote><strong><em>Here is what is actually happening in your brain during each trial, explained in full.</em></strong></blockquote><p>When a salient visual stimulus appears in your peripheral field, your superior colliculus, a small structure in the dorsal midbrain that coordinates rapid gaze and motor orienting, generates what is called a prosaccade signal. It is essentially a reflex. It happens in approximately 70 to 100 milliseconds and is largely automatic, below the level of conscious voluntary control. This is evolutionary: in environments where your survival depended on rapidly detecting motion, automatic orienting toward new visual events was adaptive. You want to look at the thing that just moved. That instinct is very old and very fast.</p><p>The anti-saccade task asks you to override that instinct. When the green circle appears on the left, you must suppress the leftward prosaccade signal and generate a voluntary rightward response instead. This override requires active engagement of your dorsolateral prefrontal cortex (dlPFC), the brain region most consistently associated with top-down executive control and working rule maintenance. It also requires the anterior cingulate cortex (ACC), which monitors for conflict between competing response signals, and the supplementary motor area (SMA), which is involved in the preparation and inhibition of voluntary motor responses.</p><blockquote>In plain terms: you need your prefrontal cortex to catch up to and cancel the reflexive orienting signal before your hand moves. The speed and reliability with which that happens is a direct measure of executive control capacity.</blockquote><p>What makes this extraordinarily useful as a monitoring signal is that these executive control structures are among the most sensitive parts of the brain to the kinds of stressors PRECISE is designed to detect.</p><p><a href="https://pubmed.ncbi.nlm.nih.gov/10718074/">Harrison and Horne (2000)</a>, published in the Journal of Sleep Research, showed that after 36 hours of sleep deprivation, anti-saccade error rates climbed substantially compared to rested baseline, while participants simultaneously rated their performance as only slightly impaired. Their prefrontal brakes were degrading at a rate their conscious self-assessment could not detect. The anti-saccade task caught it. Self-report missed it entirely.</p><p><a href="https://pubmed.ncbi.nlm.nih.gov/9389478/">Mullen and colleagues (2014)</a>, in Frontiers in Neuroscience, found that athletes who had sustained sport-related concussions showed prolonged anti-saccade latencies and elevated error rates relative to their pre-injury baseline within 24 to 48 hours, at a timepoint when standard neurological evaluations and self-reported symptom scores returned near-normal findings. This sensitivity window matters enormously because, as <a href="https://pmc.ncbi.nlm.nih.gov/articles/PMC4479139/">Giza and Hovda (2014)</a> documented in their review of the neurometabolic cascade of concussion in Neurosurgery, the ionic and metabolic disruptions following mild TBI are diffuse and subtle, often producing cognitive effects that clinical examination alone cannot reliably detect. The visual task was more sensitive to the early post-injury period than standard clinical measures.</p><p>Kahneman’s broader framework of System 1 and System 2 processing (learn more about this <a href="https://thedecisionlab.com/reference-guide/philosophy/system-1-and-system-2-thinking">here</a>), while operating at a different level of analysis, maps cleanly onto why this task works: the anti-saccade requires System 2, deliberate effortful control, to override a System 1, automatic reflexive response. System 2 is expensive. It is the first thing to degrade under cognitive load, sleep loss, emotional stress, or neural injury.</p><p>The paradigm has been used in published research on ADHD, schizophrenia, Parkinson’s disease, traumatic brain injury, aging-related cognitive decline, depression, and anxiety. It has over 40 years of published literature behind it. <a href="https://pubmed.ncbi.nlm.nih.gov/9740362/">Everling and Fischer (1998)</a>, in Neuropsychologia, provided a foundational review of the basic research and clinical studies that established the paradigm as a measure of inhibitory control. <a href="https://pubmed.ncbi.nlm.nih.gov/14976521/">Munoz and Everling’s 2004 review</a> in Nature Reviews Neuroscience built on that foundation, documenting the neural circuitry and behavioral signatures in exhaustive detail.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*pQ3bI2CSPoCupQSekAc7KA.png" /><figcaption>Image of the task taken directly from PRECISE Pro.</figcaption></figure><p>In PRECISE Pro, the anti-saccade task runs 25 trials. The green stimulus appears on either the left or right side of the screen, randomized with a jitter-resistant inter-trial interval between 1,000 and 3,000 milliseconds. Your task is to click the opposite side. Three to four minutes total.</p><h4>From those 25 trials, PRECISE Pro computes four distinct metrics.</h4><p><strong>Mean reaction time</strong> is the average time from stimulus appearance to your click across valid trials. <a href="https://pubmed.ncbi.nlm.nih.gov/14976521/">Munoz and Everling (2004)</a> established the foundational normative range for healthy adults aged 18 to 35: 280 to 450 milliseconds under standard paradigm conditions. <a href="https://pubmed.ncbi.nlm.nih.gov/16805870/">Hutton and Ettinger (2006)</a>, published in Psychophysiology, confirmed a range of 250 to 500 milliseconds with variation based on specific paradigm design parameters. PRECISE Pro classifies mean RT under 280 milliseconds as excellent, 280 to 450 as normal, 450 to 600 as borderline, and above 600 as elevated.</p><p><strong>Intra-individual RT variability</strong> is the standard deviation of your reaction times across the 25 trials. This is the metric I became most interested in after reading the IIV literature, and I will cover it in its own section in detail. The short version: <a href="https://pubmed.ncbi.nlm.nih.gov/14518812/">MacDonald, Hultsch, and Dixon (2003)</a>, in Psychology and Aging, established that young adults have a typical RT standard deviation below approximately 80 milliseconds in sustained attention tasks, with higher IIV linked to worse cognitive status. Elevated variability above 150 milliseconds is clinically notable. <a href="https://pubmed.ncbi.nlm.nih.gov/11055261/">Hultsch, Strauss, Hunter, and MacDonald (2002)</a>, in Neuropsychology, showed that individuals with higher baseline IIV showed accelerated cognitive decline over a six-year longitudinal follow-up.</p><p><strong>The fatigue index</strong> is the percentage change in mean RT from your first eight trials to your final eight trials. Mackworth’s vigilance decrement research, foundational work from 1948 that has been replicated extensively since, established that sustained attention tasks show a predictable performance decrement over time in any participant with a degraded attentional system. <a href="https://pubmed.ncbi.nlm.nih.gov/18689050/">Warm, Parasuraman, and Matthews (2008)</a> confirmed in a comprehensive review that a within-task decrement exceeding 20 percent is clinically significant.</p><blockquote>These four metrics together give a far richer picture of executive control than any single measure could. A participant who is slow but consistent tells a different story than one who is fast on average but highly variable. A participant whose accuracy drops across trials tells a different story than one whose accuracy holds but whose speed declines. The combination of metrics is what makes the signal interpretable.</blockquote><p><strong>One thing I want to address directly: how PRECISE Pro’s implementation differs from lab-based anti-saccade paradigms, and why that matters.</strong></p><p>In a controlled laboratory setting, the anti-saccade task uses eye-tracking equipment that measures actual saccade kinematics: the direction, velocity, and latency of the eye movement itself. PRECISE Pro captures neither eye movements nor saccade kinematics. It captures motor response, specifically the time and accuracy of a mouse click or screen tap in the direction opposite the stimulus.</p><p>This is a meaningful difference. It means PRECISE Pro is measuring a motor proxy for the anti-saccade response rather than the saccade itself. The cognitive processes engaged are the same, the same executive control structures are being stressed, but the output measure is different. This is why I document the norm sources carefully and acknowledge that lab-based RT ranges may not translate perfectly to cursor-based response capture. The classification thresholds PRECISE Pro uses are grounded in the published literature, but they should be understood as approximate when applied to mouse-response data.</p><p>This limitation is real and I am upfront about it everywhere in the platform. But it leaves the utility intact of the signal for within-person longitudinal tracking, because the noise structure of the browser-based implementation is consistent across sessions. If your mean RT is 320 milliseconds in session one and 410 milliseconds in session twelve, that 28% change is meaningful relative to your own baseline regardless of the relationship between cursor response time and true saccade latency. The longitudinal signal is valid even if the absolute calibration to lab norms is approximate.</p><h3>5. The Verbal Fluency Task: Speech as a Neural Signal</h3><h4>Speech was nearly cut from PRECISE Pro.</h4><p>The decision to add a second modality came late in the design process, and it came from a conversation rather than from a technical analysis. I was watching a podcast about a researcher who worked on language-based markers of early cognitive decline, specifically in Parkinson’s disease, and she made an observation that I have been thinking about ever since.</p><p>She said something like: “The thing about speech is that it is the output of everything. Working memory, lexical access, phonological encoding, motor planning, executive control. They all converge in speech production simultaneously. When any one of them is stressed, the speech reflects it before almost any other behavioral measure will.”</p><p>She was describing speech as a composite signal. A window into multiple systems at once. And she was right, which is why the neurodegeneration research community has been analyzing speech for early biomarkers of Alzheimer’s, Parkinson’s, and ALS for years. The signal is real. It is just hard to capture outside a laboratory without specialized equipment.</p><p>The Web Speech API changed that calculation. It is a browser-native API, available in Chrome and most modern browsers, that performs real-time speech transcription locally on the device. It is imperfect, its accuracy varies with microphone quality and ambient noise, and it introduces transcription delay that creates noise in the latency measurements. But it provides access to a speech signal that was previously inaccessible outside a clinical setting, and for longitudinal within-person tracking, consistent noise is manageable.</p><p>I spent three weeks deciding whether the Web Speech API’s limitations were acceptable before committing to including speech in Pro. The deciding factor was a simple test: I ran the verbal fluency tasks myself across five consecutive days, exported the metrics, and looked at whether the data showed the pattern the literature predicted. On days when I was well-rested and focused, latency was low, pause frequency was low, and words-per-minute was high. On days when I was stressed or sleep-deprived, all three metrics degraded in the direction the literature said they should. The signal was there. The noise was manageable. I committed to speech. Officially.</p><h4><strong>Here is how the speech module works.</strong></h4><p>Five verbal fluency tasks, each 60 seconds long, alternating between category naming and sentence completion. Category naming: “Name as many animals as you can.” Sentence completion: “Continue this sentence as long as possible: The thing about solving hard problems is…” The Web Speech API transcribes your speech in real time, processing locally in your browser. After each 60-second task, PRECISE Pro analyzes the transcript for four specific features.</p><p><strong>Response latency</strong> is the time from when the task prompt appears on screen to when you produce your first word. <a href="https://psycnet.apa.org/record/1989-97544-000">Levelt’s 1989</a> foundational text “Speaking: From Intention to Articulation” established the basic timeline of speech production: concept selection, lexical access, phonological encoding, and articulatory planning happen in approximately 600 to 1,000 milliseconds under normal conditions for familiar words. <a href="https://pubmed.ncbi.nlm.nih.gov/15037128/">Indefrey and Levelt (2004)</a>, published in Cognition, mapped the neural timeline in detail using PEG data and established that total preparation time from semantic concept to speech onset is approximately 400 to 900 milliseconds for high-frequency words. Anything beyond 1,500 milliseconds signals friction in the retrieval or planning stages, and PRECISE Pro flags this as elevated latency.</p><h4><em>What does elevated latency mean neurologically?</em></h4><p>It can reflect slowed lexical retrieval from the mental lexicon, mediated by the left temporal and inferior frontal regions. It can reflect increased demands on working memory, which must hold the task rule while searching for candidate words. It can reflect fatigue in the articulatory planning system. The specific source varies by individual and context, but the pattern, elevated latency that was previously lower for this individual, is the signal PRECISE is designed to surface.</p><p><strong>Pause frequency</strong> is the count of silence events exceeding 250 milliseconds during each 60-second task. <a href="https://www.sciencedirect.com/science/article/abs/pii/0024384170900288">Goldman-Eisler’s</a> 1968 psycholinguistics research, still foundational, established that spontaneous speech involves regular pauses for lexical retrieval and planning. For verbal fluency tasks, she documented three to six pauses per 60-second period as typical for healthy adults. <a href="https://pubmed.ncbi.nlm.nih.gov/480274/">Beattie and Bradbury (1979)</a>, in Linguistics, confirmed this range and linked elevated pause frequency specifically to increased working memory load during lexical search. When you are naming animals and you run out of easy ones, you pause. The pause reflects the effort of digging deeper into the lexical network. More pauses than usual, at the same point in a task relative to your own baseline, means the digging is harder.</p><p><strong>Words per minute</strong> is straightforward: the total transcript word count divided by 60. <a href="https://academic.oup.com/applij/article-abstract/11/1/90/255991?redirectedFrom=PDF">Tauroza and Allison (1990)</a> established that conversational speech rates in the same age group fall between 120 and 180 words per minute, providing the baseline against which reduced output becomes visible. Below 80 words per minute in a verbal fluency task signals reduced lexical retrieval efficiency, meaning your verbal output is constrained relative to the population reference.</p><p><strong>Composite fluency score</strong> is a PRECISE-derived index combining all four metrics into a single 0 to 100 number. The formula weights latency at 30%, pause frequency at 25%, words per minute at 30%, and a pause duration estimate at 15%. The weighting reflects the relative sensitivity of each metric: output rate and latency are the most diagnostically informative in the literature, so they receive the highest weights. Scores above 85 are excellent. 70 to 85 is normal range. 55 to 70 is mild reduction. Below 55 is below expected and warrants attention if persistent.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/911/1*AIuXKLxQu-QxZ3B3xkqOHA.png" /><figcaption>Image Credit: ChatGPT</figcaption></figure><p><strong>The privacy architecture of the speech module is something I want to address directly because it matters enormously.</strong></p><p>All speech transcription happens locally through the Web Speech API. The audio stream is processed in your browser. The transcript is analyzed in your browser. Nothing leaves your device. When you close the tab, everything is deleted. Zero audio recordings. Zero stored transcripts. The only thing you can choose to export is the structured metrics CSV, which contains numbers, not words. The CSV contains zero transcript text.</p><p>This was a hard constraint from day one. Cognitive monitoring data is sensitive. Speech data is even more sensitive because voice is personally identifiable. Building a system that captures and transmits speech data to a server, even a secure one, without explicit informed consent and robust data governance, would be ethically indefensible for a research prototype built by an individual without institutional oversight. The local-only architecture is an ethical commitment. It is a deliberate ethical choice.</p><p>The limitation this creates is <strong><em>real</em></strong>: without any server-side data, PRECISE Pro cannot aggregate speech data across users to improve the composite fluency formula or calibrate the norm library to browser-based capture. That is a genuine tradeoff, and one I am comfortable making. Signal integrity and user trust matter more than aggregate learning at this stage.</p><p>There is one <strong><em>additional technical challenge</em></strong> with the speech module I want to be transparent about. The Web Speech API introduces variable transcription delay between when you say a word and when the API returns it as a text token. This delay is approximately 100 to 300 milliseconds and varies with device, browser version, ambient noise level, and microphone gain. This delay corrupts the raw latency measurement if it is not accounted for.</p><p>PRECISE Pro partially addresses this through a calibration prompt at the start of each speech session: you say a fixed phrase, and the system measures the API’s transcription delay for that device and session. The measured delay is subtracted from subsequent latency calculations. This is an imperfect correction, because the delay varies somewhat during the session, but it substantially reduces the noise introduced by the API’s variable processing time. The methodology note in the CSV export documents this correction and flags that speech latency measurements have a higher noise floor than visual RT measurements.</p><p><em>Imperfect but honest. Clearly that is the design standard I am going for.</em></p><h3>6. The Science of Intra-Individual Variability</h3><p>This section deserves its own space because it is the scientific foundation that everything in PRECISE is built on, and it is systematically underappreciated in how we publicly talk about brain health.</p><p>I want to tell you exactly how I found it, because the path matters.</p><p>I was trying to understand why PRECISE Mini felt incomplete even after I had built it well. Mini tracked mean reaction time across sessions. It flagged when your mean RT was significantly higher than your baseline. It was doing what I had designed it to do. But something felt off about the signal.</p><p>The problem was this: some sessions where my mean RT was fine felt cognitively terrible. I felt scattered, unfocused, like my brain was jumping around. And some sessions where my mean RT was somewhat elevated felt completely stable, just a slow day. The mean RT was missing the experience entirely.</p><p>I started looking for literature on reaction time measures beyond the mean. And that search led me to intra-individual variability, or IIV, and specifically to a 2002 paper by Hultsch, Strauss, Hunter, and MacDonald in Neuropsychology. The paper was titled “<a href="https://psycnet.apa.org/record/2007-10440-010">Intraindividual variability, cognition, and aging: A longitudinal study.</a>” And it changed my understanding of what the right signal actually was.</p><p>Hultsch and colleagues followed a sample of older adults over a six-year period, measuring their cognitive performance repeatedly across that time. They tracked not just how fast people were, but how consistent they were from trial to trial and session to session. And they found something striking: the individuals with the highest intra-individual variability in reaction time at the start of the study showed significantly accelerated cognitive decline over the following six years, even when their baseline mean performance was comparable to those who declined slowly.</p><p>The variability itself was the leading indicator. The instability in the signal predicted the trajectory of decline years before the trajectory became visible in any other measure.</p><p>This finding has been replicated and extended multiple times. <a href="https://researchportalplus.anu.edu.au/en/publications/intraindividual-variability-in-reaction-time-predicts-cognitive-o/">Bielack and collegues</a>, extended the result to shorter time horizons, showing that higher baseline IIV in reaction time predicted who would shift into cognitive impairment over a five-year follow-up, with inconsistency being a stronger predictor than mean slowness alone. <a href="https://pubmed.ncbi.nlm.nih.gov/19899833/">Ducheck and colleagues (2009)</a>, showed that IIV can distinguish individuals with very mild Alzheimer’s disease from healthy older adults even when mean performance measures appear similar.</p><p>The theoretical explanation is compelling. A stable cognitive system produces stable outputs. When the neural systems underlying attention, working memory, and processing speed are functioning normally, they produce consistent reaction times across trials and sessions. When those systems are disrupted, by fatigue, injury, disease, or any other stressor, the output becomes inconsistent. The brain starts producing variable responses because the underlying regulation is compromised. Variability is the signature of a system under stress.</p><p>This is why PRECISE’s core design principle is that performance is the distraction.</p><h4><strong>Stability is the signal.</strong></h4><p>A reaction time of 350 milliseconds tells you very little on its own. But a person who consistently produces reaction times between 300 and 380 milliseconds across dozens of sessions, and then in one week begins producing reaction times ranging from 250 to 480 milliseconds, is showing something real. Their mean might not have changed much. Their variability has almost doubled. That is the signal.</p><p>The practical consequence for PRECISE Pro’s architecture is that RT standard deviation receives equal weight to mean RT in the intelligence engine’s classification and response logic. It is in its own signal block. It has its own normative reference. The PRECISE-IE, when analyzing a session, comments specifically on variability and contextualizes it with the IIV literature. A session with excellent mean RT but high SD gets flagged differently than a session with elevated mean RT but low SD.</p><p>IIV is also one of the reasons why a single PRECISE Pro session is limited in what it can tell you. The session-level SD, computed across 25 trials, is informative about your current within-session consistency. But the really meaningful IIV measure, the one the literature is talking about when it predicts cognitive decline, is session-to-session variability across weeks and months. That measure requires longitudinal data. It requires you to come back. And it is why the CSV export format, the baseline comparison functionality, and the PRECISE Intelligence Engine’s longitudinal comparison features exist: to make the accumulation and analysis of longitudinal data as frictionless as possible.</p><p>Every design choice in PRECISE Pro is ultimately in service of making it easier to collect and interpret that longitudinal signal. Because that is where the real information lives.</p><h3>7. Engineering the Browser: Timing, Latency, and Precision Under Constraints</h3><p>I want to take you inside the engineering process, because the gap between what browsers can theoretically do and what they actually do in practice is enormous, and understanding that gap is central to understanding why PRECISE Pro is built the way it is.</p><p>When I started building the visual assessment module, my first implementation was straightforward. I had a container element. When the trial started, I used JavaScript to set the element’s backgroundColor to green, showed the stimulus, and attached a click listener to capture the response. I timestamped the stimulus onset with Date.now(). I timestamped the response with Date.now(). Reaction time was the difference.</p><p>I ran fifty trials on myself and exported the data. The mean RT looked plausible. The distribution looked roughly right. Then I looked at the raw trial-level data and noticed something deeply wrong: the minimum RT in the dataset was 31 milliseconds. A human cannot consciously process a visual stimulus and generate a voluntary motor response in 31 milliseconds. The physiological minimum for the full perceptual-motor loop is approximately 100 milliseconds, and typical simple RT for a rested adult is 150 to 250 milliseconds. 31 milliseconds was impossible.</p><p>I had a bug, and it was the invisible kind: no error in the console. It was a timing architecture problem.</p><p>What was happening: the browser renders the DOM asynchronously. When I called the JavaScript instruction to change the background color, the visual update queued for the next render frame rather than firing immediately. It was queued for the next render frame, which on a 60Hz display happens approximately every 16 milliseconds. But the Date.now() timestamp recorded as stimulus onset captured the moment of the JavaScript instruction, timed to the JavaScript instruction rather than to what appeared on screen. In the worst case, the stamp could be up to 16 milliseconds earlier than the actual visual presentation.</p><p>Simultaneously, I discovered that Date.now() itself has worse resolution than its documentation suggests. While it technically offers 1 millisecond resolution, in practice, with privacy protections introduced by browsers in response to Spectre/Meltdown vulnerabilities, Date.now() is often rounded to 1 to 5 milliseconds. The MDN Web Docs on High Resolution Time document this intentional quantization explicitly, noting that browsers coarsen timer precision to prevent side-channel timing attacks. For RT measurement, that rounding error compounds into meaningful noise.</p><h4><strong>I rebuilt the timing system from scratch. Three complete rewrites over about two weeks. While, mind you I had to write final exams for school.</strong></h4><p>The final architecture, engineered with help from ChatGPT for the more intricate browser API orchestration, uses performance.now() for all timing. This high-resolution timer avoids the rounding restrictions affecting Date.now() and provides sub-millisecond precision in modern browsers. Stimulus onset is timestamped inside a requestAnimationFrame callback, which fires at the moment the browser is about to paint the next frame. This synchronizes the timestamp to the actual display update rather than to the JavaScript execution queue.</p><pre>let stimulusOnset;<br>requestAnimationFrame((timestamp) =&gt; {<br>    showStimulus();<br>    stimulusOnset = performance.now();<br>});</pre><p>Response capture uses pointerdown rather than click. The click event fires after both the press and the release of the mouse button, adding 30 to 80 milliseconds of additional latency depending on how long the user holds the button. pointerdown fires at the initial contact, which is the correct cognitive event: the moment the neural signal reaches the finger.</p><p>The inter-trial interval is jitter-randomized between 1,000 and 3,000 milliseconds using a seeded random generator that prevents predictable patterns. If the interval were fixed, a trained user could start timing their response to the expected interval, generating anticipatory RTs that would corrupt the data. Random intervals eliminate this.</p><p>False start detection is critical. Any response that occurs before stimulus onset, specifically before the requestAnimationFrame callback fires the showStimulus call, is logged as a false start and excluded from RT calculation. Any response with RT below 100 milliseconds is flagged as anticipatory and receives a 1,000 millisecond penalty. The 100 millisecond threshold is the physiological minimum for a genuine perceptual-motor reaction: it takes approximately 70 to 100 milliseconds for the visual signal to travel from the retina through the lateral geniculate nucleus to primary visual cortex, and additional time for cortical processing and motor output. A response faster than 100 milliseconds cannot be genuine.</p><p>After the third rewrite, I ran 200 trials on myself across four sessions and checked the data against what I would expect physiologically. The minimum RT was 147 milliseconds. The distribution was right-skewed with a peak around 280 to 320 milliseconds, which is exactly what the literature predicts for a rested young adult on a simple visual response task. The false start rate was near zero. The data was clean.</p><h4><strong>This is the level of engineering effort that cognitive assessment requires in a browser environment.</strong></h4><p>Casual implementation produces garbage data, that won’t come of good use to ANYONE. The research literature on online cognitive testing, including work by Plant and colleagues and Reimers and Stewart, has documented for years that naive timing implementations introduce systematic errors large enough to corrupt experimental conclusions entirely. Getting it right requires deliberate architecture, full stop.</p><p>And I want to be clear: PRECISE Pro does not match the timing precision of dedicated neuropsychological hardware, and I say so plainly. Lab hardware can achieve sub-millisecond timing accuracy. My implementation achieves approximately 5 to 15 milliseconds of timing error in optimal conditions. The noise floor is real and documented. But for the record, I am 14 right now, sooooo yes.</p><p>But for within-person longitudinal tracking, where the same noise structure is present across every session, this level of precision is sufficient. The IIV thresholds I am working with, 80 milliseconds SD at the lower bound of clinical notability, are far above the timing noise floor. A genuine variability signal will be visible above the measurement noise.</p><p>The speech module has its own engineering challenges. The Web Speech API’s transcription delay is variable and harder to control than visual timing. I spent a week trying to find a way to pre-compensate for the API delay before settling on the calibration approach: a fixed prompt at session start that the system uses to measure the API’s current transcription latency, which is then subtracted from subsequent latency measurements. It is imperfect. The delay drifts somewhat during a session. But it is substantially better than ignoring the delay entirely, and the CSV export documents the correction method transparently.</p><p>There is a broader lesson from this engineering work that I think about often. Most articles about browser-based cognitive assessment bury the timing challenge in a footnote. In practice, it is the <strong>central problem.</strong></p><p><em>A browser-based cognitive test with sloppy timing goes beyond imprecision. It is measuring something other than what it claims to measure. The integrity of the signal depends entirely on the integrity of the timestamp.</em></p><p>Everything else in PRECISE Pro, all the AI modeling, all the intelligence engine analysis, all the norm comparisons, depends on the quality of the raw data underneath. Garbage timing in, garbage insight out. The weeks I spent on the timing architecture were not overhead. <strong>They were the foundation. And the results of all this grind are in front of you at </strong><a href="https://precisepro-alizasamirkhoja.netlify.app/">https://precisepro-alizasamirkhoja.netlify.app/</a></p><h3>8. The AI Layer: Linear Regression as a Deliberate Choice</h3><p>When I tell people that PRECISE Pro uses AI, the first thing they assume is a neural network. Maybe a transformer. Something deep, something black-box, something that requires a GPU cluster to run. The second thing they assume, once I explain it is browser-based, is that the AI must be a gimmick. And to be honest, I would think no different. The world right now quite literally uses AI as nothing more than a buzzword.</p><p><strong>The actual answer is linear regression. And it is neither a gimmick nor a compromise. And I want to explain, in detail, why that is the right choice rather than a limitation.</strong></p><p>The live AI model in PRECISE Pro’s visual assessment fits a linear regression to your reaction time trajectory across the 25 trials as they complete. Each time a trial finishes, the model receives the new data point, re-fits the regression line, updates the predicted RT for the next trial, and displays both the prediction and your actual RT. At session end, it reports the model’s R-squared value as a confidence percentage, the slope of the regression line as a measure of within-session fatigue or improvement, and a plain-language interpretation of what the trajectory means.</p><p><strong>Why linear regression?</strong></p><p>The first reason is <strong>interpretability.</strong> A linear regression produces two numbers that have direct, grounded interpretations: a slope and an intercept. A positive slope means your reaction times are increasing across trials, consistent with within-session fatigue or attention decay. A negative slope means you are getting faster, consistent with practice effects or task warming. A slope near zero means you are stable. These interpretations map cleanly onto the published literature on vigilance decrement, and users can understand them without any background in machine learning.</p><p>A deep learning model producing the same prediction would be a black box. You would know what it predicted but the reasoning would be opaque. In a cognitive health monitoring context, interpretability is a scientific requirement, full stop. It is a scientific requirement. The moment a user cannot understand the basis for a flag, the flag loses its value and potentially creates anxiety without context.</p><p>The second reason is <strong>data volume.</strong> A single session produces 25 data points for the visual task. This volume is insufficient to train or meaningfully apply sophisticated machine learning. Deep learning methods are extraordinarily data-hungry. They perform well at the scale of thousands or millions of examples. At 25 examples, they offer no accuracy advantage over a well-specified simpler model. They are just more complex.</p><p>The relationship between trial number and reaction time in a cognitive task is expected to be approximately linear in most conditions: <em>a gradual increase with fatigue, a gradual decrease with practice, or stability.</em> A linear model is the appropriate functional form for this relationship at this data scale. A more complex model would add complexity without improving prediction. It would add complexity without benefit. Which is something we definitely DO NOT want, given that I just recently started coding and most complex code is written with AI assistance.</p><p>The third reason is <strong>browser performance.</strong> PRECISE Pro runs entirely in the browser. Zero server. Zero external compute. The AI must run in JavaScript + HTML on the user’s device, updated after every trial in real time with no perceptible lag. We want to avoid lag AT ALL COSTS. A matrix decomposition for a 25-point linear regression runs in microseconds in JavaScript. A neural network forward pass, even a small one, would introduce meaningful delay and require either a heavyweight JavaScript ML library or a WebAssembly runtime. The engineering cost is not worth the marginal accuracy gain.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Ss0c77vcxfp-OstCoQxBwg.png" /><figcaption><strong><em>Left</em></strong><em>:</em> High model confidence (R² = 0.78) reflects a predictable, learnable cognitive trajectory. <strong><em>Right</em></strong><em>:</em> Low model confidence (R² = 0.21) captures erratic performance, identifying session-level instability in real-time.</figcaption></figure><p><em>Note: Data points and trend lines are for illustrative purposes and do not represent actual user data.</em></p><p>There is also an important signal in the model’s confidence score itself. High R-squared, above 70%, means your RT trajectory across the session formed a consistent, learnable pattern. This is itself a stability signal: a regulated cognitive system produces predictable responses. The model can learn it. Low R-squared means your trajectory was erratic. Prediction was difficult because the pattern was inconsistent. This erratic trajectory is itself an IIV signal: high session-level variability manifesting as an unpredictable within-session trajectory.</p><p>The real-time display of model predictions during the session is a deliberate pedagogical choice as well. Users can watch the model update with each trial. They can see when their actual RT diverges from the prediction. They can see their own trend forming. This is AI literacy through direct experience. The goal of PRECISE is ultimately to make cognitive science accessible and concrete. Watching a model update in real time is more educational than reading a description of what a model does.</p><p>The PRECISE Intelligence Engine, which I will cover in the next section, uses rule-based classification combined with a structured response generation system. That is also built without a neural network, and it is also the right choice for the same reasons: interpretability, data scale appropriateness, and browser-based performance.</p><p>I want to be direct about where more sophisticated models belong in the PRECISE roadmap. PRECISE Max, the third layer, will integrate longitudinal Mini and Pro data across many sessions from many users. At that scale, with potentially hundreds or thousands of sessions per user across years, and population-level data to train on, transformer-based temporal models become appropriate. The longitudinal structure of the data, time series with multiple modalities, is exactly the kind of structure that self-attention mechanisms are well-suited to. But that is the right tool for that problem at that scale. It is not the right tool for a 25-point within-session trajectory.</p><p>Using a complex model where a simple model is appropriate is mismatch, full stop. It is mismatch between tool and problem. PRECISE Pro uses the right tool for the problem it is solving.</p><h3>9. The PRECISE Intelligence Engine: Building an On-Device Inference System</h3><p>The PRECISE Intelligence Engine (PRECISE-IE) started as a much simpler idea.</p><p>Originally, I planned to just display a classification table after each session. Your metrics, their normative ranges, a green/yellow/red indicator. Clean, simple, easy to understand. I mocked it up in about a day and it looked fine. Similar to what Mini had.</p><p>Then I showed it to two people. A classmate who had no neuroscience background asked: “What does high SD mean? Is that bad?” A cousin who does have a neuroscience background asked: “How does it handle the interaction between the speech and visual signals? Are you weighting them together?”</p><p>Both questions pointed at the same gap. A table of metrics with color coding is useful for someone who already understands what the metrics mean and how they interact. For everyone else, it is <strong>noise dressed up as information.</strong> A color-coded table also fails to express nuance, context, or the relationship between signals across modalities.</p><h4><strong>I scrapped the table and started over. ONCE More.</strong></h4><p>The replacement became PRECISE-IE: a conversational interface where you upload your session CSV and ask questions in plain language, and receive structured answers grounded in the published literature. The intelligence engine analyzes your data, classifies every metric against its normative reference, determines what you are asking, assembles a response from structured blocks, and delivers it in language that a student, an athlete, or a parent can read without a background in neuropsychology.</p><p>Building this required solving a problem that most AI-powered analysis tools sidestep by using an external API: <strong>how do you deliver genuinely useful, scientifically grounded analysis of a user’s specific data without sending that data anywhere?</strong></p><p>PRECISE Pro has zero institutional funding. Zero API budget. Even if those resources existed, routing cognitive performance data to a commercial LLM API would be ethically problematic: the data is sensitive, the privacy implications are real, and the analysis produced by a general-purpose language model would likely lack grounding in the specific published norms PRECISE uses. The analysis needed to be entirely local and deterministic.</p><p>PRECISE-IE is a deterministic inference engine. Every response it generates is the product of explicit, auditable logic rather than probabilistic text generation. This is a design principle, not a workaround. Deterministic analysis means every response is traceable to a specific classification rule and a specific published source. There are no hallucinations. There is no stochastic variation between runs on the same data. If you upload the same CSV twice and ask the same question, you get the same answer.</p><h4>Here is the pipeline in detail.</h4><p><strong>Parsing.</strong> Your uploaded CSV is read entirely in the browser using the FileReader API. PRECISE-IE uses fuzzy header matching rather than exact string matching to locate the relevant metric fields. The fuzzy matcher checks whether a line contains any of a set of associated keywords, case-insensitively, using substring matching. “Mean Reaction Time (ms)”, “mean rt”, “average reaction time”, and “session mean RT” all map to the same field. This matters because the CSV format has evolved across PRECISE Pro versions and because users sometimes open CSVs in Excel, modify them slightly, and re-upload. The fuzzy parser handles minor variations gracefully. If a metric fails to parse or its value is outside the range of valid finite numbers, the metric is silently omitted. Signal blocks only render if the underlying value parsed successfully and passed a finite-number validation check. This is what eliminated the null rendering bugs that plagued earlier versions of the interface.</p><p><strong>Classification.</strong> Each parsed metric is classified against the tier thresholds derived from the published literature. Every threshold traces back to a specific source: RT norms from <a href="https://pubmed.ncbi.nlm.nih.gov/14976521/">Munoz and Everling (2004)</a>, IIV thresholds from <a href="https://pubmed.ncbi.nlm.nih.gov/11055261/">Hultsch et al. (2002)</a>, accuracy norms from <a href="https://pubmed.ncbi.nlm.nih.gov/21697711/">Crawford et al. (2011)</a>, speech latency from <a href="https://psycnet.apa.org/record/1989-97544-000">Levelt (1989)</a>. Nothing is made up. Zero estimation. If I cannot cite it, it does not go into the classification system.</p><p><strong>Intent routing.</strong> Your question is matched against a set of intent categories through keyword-based routing with priority ordering. Categories include: summary, signals, confidence, population comparison, next steps, worry assessment, speech detail, visual detail, baseline comparison, and model explanation. If your question matches multiple categories, the most specific match takes priority. “Should I be worried about my SD?” routes to worry assessment, and bypasses the general signals response. “What does the model confidence mean?” routes to model explanation. “Give me everything” routes to the full summary.</p><p><strong>Response assembly.</strong> Based on your classified data and detected intent, PRECISE-IE assembles a response from pre-built structural blocks. Text blocks carry narrative prose. Signal blocks display your metric with its tier classification, a plain-language interpretation, and a citation reference. Confidence bars visualize signal quality with animated fills. Section headers visually separate visual from speech metrics when both datasets are present. A disclaimer block appears on every response involving individual data interpretation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*w53mvUDnLMA78S7i897cqQ.png" /><figcaption>The interface translates raw reaction time trends into color-coded, plain-language alerts. Local browser processing identifies cognitive states like fatigue or instability without external compute lag.</figcaption></figure><p><em>Note: For privacy reasons, this AI-generated mockup (Gemini) replaces real user data.</em></p><p>The worry assessment response is the most carefully designed response in the system. It counts how many metrics are outside their normal tier, distinguishes between zero flags (clear, direct reassurance with data citation), one to two flags (contextualizes with session noise and biological variability, recommends retest rather than alarm), and multiple flags (acknowledges the pattern, recommends professional consultation if the pattern persists across sessions, provides specific guidance on what to bring to a clinician). Every worry response ends with a reminder that a single session is a snapshot and that longitudinal patterns are the meaningful signal.</p><p>The baseline comparison functionality is where PRECISE-IE becomes genuinely useful for ongoing monitoring. When your CSV includes a prior session comparison (generated when you upload an earlier session’s CSV at the start of a new assessment), PRECISE-IE applies the Reliable Change Index methodology from <a href="https://pubmed.ncbi.nlm.nih.gov/2002127/">Jacobson and Truax (1991)</a>. This methodology accounts for both the magnitude of change and the expected measurement error, producing a standardized score that indicates whether the change is beyond what would be expected from measurement noise alone.</p><p>A 15% or greater change in mean RT or speech latency crosses the meaningful change threshold under the Jacobson-Truax framework at this measurement noise level. A 10-point or greater change in composite fluency score likewise. Changes below these thresholds are attributed to normal session-to-session biological variability. Changes above these thresholds, when the direction is consistent with degradation, are flagged as worthy of attention and ideally, professional follow-up.</p><p>The development of PRECISE-IE took longer than any other part of PRECISE Pro. The first version, built in a day, was the color-coded table I scrapped. The second version was a chatbot with hardcoded responses that skipped actual data analysis entirely, it just responded to keywords with canned text. A teacher pointed out within the first ten minutes of a demo that I was unable to demonstrate it analyzing a real CSV. <strong>Back to the drawing board lol.</strong></p><p>The third version parsed the CSV correctly but had null rendering issues: metrics that failed to parse would still generate signal blocks, filled with “undefined” and “NaN” text. This is the bug that required the fuzzy parser, numeric guards, and safe label helpers that are now foundational to the system. The debugging process for those null rendering issues took a week, mostly because the failures were intermittent and depended on subtle variations in CSV structure.</p><p>The fourth version, the current one, is the first one that works cleanly across all tested CSV formats and all intent categories. It is also the first one I have been willing to describe publicly in technical detail, because the earlier versions would have collapsed under scrutiny. This one will hold in the long run.</p><h3><em>10. The Population Norms: Where the Numbers Come From</em></h3><p>One of the things that separates PRECISE Pro from most brain health apps on the market is that every threshold it uses traces back to a specific published source. I want to document that source library in full, because scientific credibility requires showing your work.</p><p>The norm library was assembled over several months of literature review. The process was iterative: I would find a relevant paper, read it in full, extract the normative data, check whether it had been subsequently replicated or challenged, and decide whether to incorporate it. Papers that were methodologically robust, widely cited, and whose findings had been confirmed by subsequent work are the ones that made it in. Papers with small samples, specific populations that may limit generalizability, or findings that conflicted with the broader literature were deprioritized.</p><h4><strong><em>Click </em></strong><a href="#1aa5"><strong><em>HERE</em></strong></a><strong><em> to skip this section.</em></strong></h4><p>Here is the complete norm library as implemented in PRECISE Pro.</p><h4><strong>Visual Processing Norms</strong></h4><p><strong>Anti-saccade reaction time:</strong></p><p><em>Primary source-</em> <a href="https://doi.org/10.1038/nrn1345">Munoz, D.P., and Everling, S. (2004). “Look away: the anti-saccade task and the voluntary control of eye movement.” Nature Reviews Neuroscience, 5(3), 218–228.</a> Healthy adults aged 18–35: mean RT of 280 to 450 milliseconds.</p><p><em>Secondary sources-</em></p><ul><li><a href="https://onlinelibrary.wiley.com/doi/abs/10.1111/j.1469-8986.2006.00403.x">Hutton, S.B., and Ettinger, U. (2006). “The antisaccade task as a research tool in psychopathology: a critical review.” Psychophysiology, 43(3), 302–313.</a> Confirmed range of 250 to 500 milliseconds with paradigm-specific variation.</li><li><a href="https://pubmed.ncbi.nlm.nih.gov/9740362/">Anti-saccade paradigm circuitry and inhibition requirements: Everling, S., and Fischer, B. (1998). “The antisaccade: A review of basic research and clinical studies.” Neuropsychologia, 36(9), 885–899.</a> This review clarifies the mechanism by which anti-saccades require suppression of a prepotent prosaccade and confirms that responses below physiological minimums indicate anticipatory rather than genuine reactive behavior.</li></ul><p><strong>Intra-individual RT variability:</strong></p><ul><li><a href="https://psycnet.apa.org/doiLanding?doi=10.1037%2F0882-7974.18.3.510">MacDonald, S.W.S., Hultsch, D.F., and Dixon, R.A. (2003). “Performance variability is related to change in cognition: Evidence from the Victoria Longitudinal Study.” Psychology and Aging, 18(3), 510–523.</a> Young adults aged 18–35: SD typically 40 to 80 milliseconds in sustained attention tasks.</li><li><a href="https://www.taylorfrancis.com/chapters/edit/10.4324/9780203837665-10/intraindividual-variability-cognition-aging-david-hultsch-esther-strauss-michael-hunter-stuart-macdonald">Hultsch, D.F., Strauss, E., Hunter, M.A., and MacDonald, S.W.S. (2002). “Intraindividual variability, cognition, and aging.” Neuropsychology, 16(2), 199–207.</a> Established that older adults with higher RT IIV at baseline showed accelerated cognitive decline over six years.</li><li><a href="https://psycnet.apa.org/doiLanding?doi=10.1037%2Fa0017798">MacDonald, S.W.S., Li, S-C., and Backman, L. (2009). “Neural underpinnings of within-person variability in cognitive functioning.” Psychology and Aging, 24(4), 792–808.</a> Documents that elevated IIV reflects CNS compromise and is an early sensitive marker across disorders.</li></ul><p><strong>Inhibitory control accuracy:</strong></p><ul><li><a href="https://psycnet.apa.org/doiLanding?doi=10.1037%2Fa0022032">Crawford, T.J., et al. (2011). “Antisaccade performance in normal and clinical populations: Normative data and diagnostic implications.” Neuropsychology, 25(1), 1–10.</a> Healthy adults: 75 to 92 percent accuracy.</li><li>Hutton and Ettinger (2006): below 70 percent is clinically notable.</li><li>Munoz and Everling (2004): error rates of 20 to 25 percent are within normal range for the standard paradigm.</li></ul><p><strong>Fatigue index:</strong></p><ul><li><a href="https://journals.sagepub.com/doi/10.1080/17470214808416738">Mackworth, N.H. (1948). “The breakdown of vigilance during prolonged visual search.” Quarterly Journal of Experimental Psychology, 1(1), 6–21.</a> Foundational vigilance decrement research.</li><li><a href="https://journals.sagepub.com/doi/10.1518/001872008X312152">Warm, J.S., Parasuraman, R., and Matthews, G. (2008). “Vigilance requires hard mental work and is stressful.” Human Factors, 50(3), 433–441.</a> Within-task decrement exceeding 20 percent classified as significant.</li></ul><h4><strong>Speech and Verbal Fluency Norms</strong></h4><p><strong>Response latency:</strong></p><ul><li><a href="https://books.google.ca/books?hl=en&amp;lr=&amp;id=LbVCdCE-NQAC&amp;oi=fnd&amp;pg=PR13&amp;dq=Speaking+From+Intention+to+Articulation+Levelt+1989&amp;ots=J0hDUi4dwO&amp;sig=i2D7t41XFA0XqC6old1wJ32MCLM&amp;redir_esc=y#v=onepage&amp;q=Speaking%20From%20Intention%20to%20Articulation%20Levelt%201989&amp;f=false">Levelt, W.J.M. (1989). Speaking: From Intention to Articulation. MIT Press.</a> Speech initiation for familiar words: 600 to 1,000 milliseconds from concept to articulatory onset.</li><li><a href="https://www.sciencedirect.com/science/article/abs/pii/S0010027703002294?via%3Dihub">Indefrey, P., and Levelt, W.J.M. (2004). “The spatial and temporal signatures of word production components.” Cognition, 92(1–2), 101–144.</a> Mapped neural timeline of lexical access and phonological encoding.</li><li>PRECISE threshold: above 1,500 milliseconds signals retrieval friction.</li></ul><p><strong>Pause frequency:</strong></p><ul><li><a href="https://books.google.ca/books/about/Psycholinguistics.html?id=BAsNAAAAIAAJ&amp;redir_esc=y">Goldman-Eisler, F. (1968). Psycholinguistics: Experiments in Spontaneous Speech. Academic Press. Foundational psycholinguistics.</a> Typical: 3 to 6 pauses per 60-second verbal fluency task.</li><li><a href="https://link.springer.com/article/10.1007/BF01067306">Beattie, G., and Bradbury, C. (1979). “The pause and speech.” Linguistics, 17(221), 147–160.</a> Linked elevated pause frequency to working memory load.</li></ul><p><strong>Words per minute:</strong></p><ul><li><a href="https://medium.com/emic.oup.com/applij/article-abstract/11/1/90/255991?redirectedFrom=fulltext&amp;login=false">Tauroza, S., and Allison, D. (1990). “Speech rates in British English.” Applied Linguistics, 11(1), 90–105.</a> Conversational speech: 120 to 180 words per minute.</li><li><a href="https://academic.oup.com/acn/article-abstract/11/4/329/1614">Ruff, R.M., Light, R.H., Parker, S.B., and Levin, H.S. (1996). “Benton Controlled Oral Word Association Test: Reliability and updated norms.” Archives of Clinical Neuropsychology, 11(4), 345–358.</a> Normative verbal fluency data stratified by age.</li><li><a href="https://www.tandfonline.com/doi/abs/10.1080/13854046.2020.1762931">Marquine, M.J., et al. (2016). “Demographically adjusted norms for selected tests of verbal fluency.” The Clinical Neuropsychologist, 30(7), 1–23.</a> Confirms that age and education significantly affect fluency scores and should be accounted for in norm application.</li></ul><p><strong>Composite fluency:</strong></p><ul><li>The PRECISE composite formula integrates latency, pause frequency, and output rate against the normative frameworks established in Ruff et al. (1996) and Marquine et al. (2016), weighting latency at 30%, pause frequency at 25%, words per minute at 30%, and a pause duration estimate at 15% to reflect the relative diagnostic sensitivity of each component.</li></ul><h4><strong>Reliable Change Methodology:</strong></h4><p><a href="https://psycnet.apa.org/doiLanding?doi=10.1037%2F0022-006X.59.1.12">Jacobson, N.S., and Truax, P. (1991). “Clinical significance: A statistical approach to defining meaningful change in psychotherapy research.” Journal of Consulting and Clinical Psychology, 59(1), 12–19.</a> Reliable Change Index methodology for determining whether observed change is beyond measurement noise. Applied to session-to-session baseline comparison in PRECISE-IE. Threshold: 15 percent change in RT or latency, 10 composite points in fluency.</p><p>I want to make one additional point about the norm library that I think is important. All of these norms come from lab-based research. They were established using controlled stimuli, calibrated hardware, trained administrators, and carefully selected participant samples. Moreover, I acknowledge how to the naked eye some of these sources may seem outdated, they are the most recent + reliable stats I could find given my status as a junior high student.</p><blockquote>Translating them to browser-based capture introduces known and unknown differences. The timing noise floor is different. The stimulus presentation conditions are different. The absence of an administrator introduces variability in how participants understand and follow task instructions.</blockquote><p>These differences are real and I acknowledge them in the platform’s methodology documentation. But the alternative, making up thresholds or using thresholds from a single small convenience sample, would be scientifically worse. Published lab norms are an imperfect reference for browser-based data, but they are the best available reference. Using them, while documenting the limitations of that use, is the intellectually honest choice.</p><p>As PRECISE Pro accumulates session data, with user consent and appropriate privacy protections, the norm library can be calibrated to browser-based capture specifically. That calibration is part of the Fusion/Max roadmap.</p><h3>11. Signal Architecture: How PRECISE Pro Processes What It Captures</h3><p>The signal processing pipeline in PRECISE Pro is the part of the platform that most directly reflects the tension between scientific rigor and practical constraints, and I want to walk through it in detail because the design decisions are non-obvious.</p><p>For the visual module, every trial generates three primitive measurements: the timestamp of stimulus onset (captured in requestAnimationFrame), the timestamp of response (captured in pointerdown), and a boolean indicating whether the response was on the correct side. From these three primitives, PRECISE Pro computes the trial-level RT as the difference between response and stimulus timestamps, and classifies the trial as correct, incorrect, false start, or anticipatory based on the response timing and direction.</p><p>Trials classified as false starts (response before stimulus) and highly anticipatory trials (RT below 100 milliseconds) are excluded from RT calculations. They are logged separately in the CSV export with their classification code, so the full trial-level record is preserved for anyone who wants to analyze the raw data. But they are excluded from the session metrics because including them would corrupt the central tendency and SD measures.</p><h4><strong>Across the 25 valid trials, the session-level computation is:</strong></h4><ul><li><strong>Mean RT: </strong>arithmetic mean of valid trial RTs. Standard deviation: sample SD (dividing by n-1) of valid trial RTs.</li><li><strong>Accuracy rate:</strong> proportion of valid trials answered on the correct side.</li><li><strong>Fatigue index: </strong>((mean RT of trials 18–25) — (mean RT of trials 1–8)) / (mean RT of trials 1–8) × 100, expressed as a percentage change.</li></ul><p>The live linear regression model adds two more computed values: <strong>slope</strong> (in milliseconds per trial, the rate of RT change across the session) and <strong>R-squared</strong> (the proportion of RT variance explained by the linear trial-number predictor). The confidence score reported to the user is R-squared × 100, rounded to the nearest integer.</p><p>For the <strong>CSV export</strong>, every session produces the following structure. A metadata section at the top lists the session date and time, the assessment module version, the user’s reported age and sex (used for norm lookup), and the session parameters. Below that, the session-level metrics are listed with their computed values. If a prior session CSV was uploaded at session start, a comparison section follows, showing the percentage change for each metric, the direction of change, and a flag indicating whether the change exceeds the Jacobson-Truax meaningful change threshold. After that comes the trial-level data table: trial number, stimulus side, response side, RT in milliseconds, and trial classification code.</p><p>Finally, a notes section at the bottom explains each metric, its normative range, and how to interpret the comparison flags. This section was added because early beta users consistently opened the CSV, looked at the numbers, and had no idea what they meant. The notes section is documentation embedded directly in the data file, ensuring that anyone who opens the CSV six months later has the context they need to interpret it without searching for a separate guide.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*lWeBHZvGG018dMIhf-DOKQ.png" /><figcaption><em>The PRECISE data output includes embedded metadata and longitudinal comparisons for seamless tracking. This structured format ensures full data ownership and portability for the user. For privacy reasons, this AI-generated mockup replaces real personal data.</em></figcaption></figure><p>The speech module’s pipeline is structurally similar but more complex at the raw data level.</p><p>The Web Speech API returns results as SpeechRecognitionResult objects with two properties: a transcript string and a confidence value. PRECISE Pro listens to interim results (partial transcripts as speech is ongoing) and final results (committed transcripts after a pause). The response latency is measured from task prompt appearance to the timestamp of the first final result object that contains at least one word, minus the calibrated API delay correction.</p><p>Pause frequency is estimated from the temporal gaps between consecutive final result objects. When the API commits a final result, it indicates that a pause longer than the API’s internal threshold has occurred. PRECISE Pro counts these commits per 60-second task, minus one (since the final commit at task end is not a mid-task pause), as the pause frequency estimate. This is an approximation. The API’s internal pause threshold is not precisely documented and varies between browsers. The CSV methodology note acknowledges this limitation.</p><p>Words per minute is computed from the full session transcript length divided by 60. The full transcript is assembled from all final result strings concatenated in chronological order. Composite fluency applies the weighted formula described in section 5.</p><p>The speech CSV structure mirrors the visual one: metadata, session metrics, task-level breakdown for each of the five tasks, and the baseline comparison section if prior speech data was uploaded. The task-level breakdown shows latency, pause frequency, and words per minute for each individual task, allowing users to see variation across task types (category naming versus sentence completion) in addition to session-level averages.</p><p>All of this, both the visual and speech pipelines, runs entirely in JavaScript in the browser. No server. No API calls. No external dependencies except the Web Speech API itself, which is native to the browser. The platform functions entirely offline after the initial page load, which is a specific design choice reflecting the privacy architecture: data that never leaves the device cannot be breached.</p><h3>12. Anti-Sandbagging: Why You Cannot Fake This</h3><p>When I first explained PRECISE Pro to a researcher I had connected with through LinkedIn, one of the first questions she asked was: “What stops an athlete from just performing better on purpose when they want to be cleared to return to play?”</p><p>For a single session, the answer is: nothing. If you know which direction you want your scores to go and you try hard in that session, you can probably move your mean RT down by 5 to 10 percent, your accuracy up by a few points, and your speech fluency score up by being more deliberate and focused than you otherwise would be.</p><h4><strong>A single session can absolutely be gamed by a motivated individual.</strong></h4><blockquote><strong>This is why a single session is almost never the relevant unit of analysis in PRECISE Pro.</strong></blockquote><p>The genuine anti-sandbagging mechanism is longitudinal, and it works through three interlocking properties.</p><p><strong>First</strong>: variability cannot be gamed over time. Your intra-individual variability pattern across dozens of sessions reflects the stability of your underlying attentional system. A regulated, healthy attentional system produces consistent RT patterns session after session. A compromised system, whether from concussion, burnout, chronic stress, or sleep deficit, produces inconsistent patterns. You might be able to produce a fast mean RT in a single session through deliberate effort. Producing a consistently low RT standard deviation across 20 sessions through deliberate effort is beyond any conscious strategy, because the SD reflects automatic regulatory processes that operate below the level of conscious control.</p><p>Think of it this way. You can hold your breath for a minute, but you cannot hold your breath consistently for 20 sessions without your body’s automatic respiratory regulation eventually asserting itself. RT variability works similarly. The automatic cognitive regulation that produces consistent reaction times is not something you can consciously sustain when the underlying system is genuinely stressed.</p><p><strong>Second</strong>: the anti-saccade task’s difficulty is structural. The core challenge is suppressing a reflexive prosaccade response and generating a voluntary countermove. This suppression depends on the functional capacity of your dlPFC in real time. You can try harder. That might improve your mean RT slightly in a session where you are otherwise coasting. Your accuracy and variability remain anchored to your actual inhibitory control capacity, regardless of effort. The task is probing a capacity, not a willingness.</p><p><strong>Third</strong>: the speech task introduces a different kind of resistance. Verbal fluency under time pressure is an output of multiple automatic processes happening simultaneously: lexical retrieval, phonological encoding, speech motor planning. You can try to speak faster, but speaking faster without increasing error and pause frequency requires your lexical retrieval system to actually produce words faster, which requires the underlying neural system to be functioning at that rate. Trying harder raises effort, but effort and capacity are different. A system with reduced lexical retrieval speed will produce more pauses under pressure, because the planning stage cannot keep up with the output demand, regardless of how hard the speaker is trying.</p><p>The combination of two modalities also matters in a specific way. Performing well on both tasks simultaneously requires your executive control system (visual task) and your language production system (speech task) to both be at near-baseline capacity in the same session. If one is compromised, it shows in that modality. And critically, the pattern of which modality shows degradation carries information: a concussion that primarily affects executive control will show in the visual task before the speech task. A stress response that primarily affects verbal working memory will show in speech latency before it shows in RT.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/788/1*aK9QSwEjLF6x6RBvV53_8A.png" /></figure><blockquote>There is a longer-term roadmap for anti-sandbagging that I am designing for PRECISE Max/Fusion. Session timing patterns carry information: an individual who always completes sessions very quickly, regardless of condition, may be rushing through trials rather than engaging genuinely. Passive signals like motor planning time (the delay between stimulus and response movement onset, different from RT) are largely outside conscious control and will be captured in Max’s fine motor coordination module. Anomaly detection across session histories, flagging sessions whose profile is statistically implausible relative to the individual’s history, will add a further layer.</blockquote><p>For now, PRECISE Pro’s anti-sandbagging properties are real but acknowledge their limits. A single session can be performed with deliberate effort. A longitudinal history cannot be faked. And the longitudinal history is the signal.</p><h3>13. Why I Chose These Specific Tasks</h3><p>The task selection process for PRECISE Pro was more iterative and more difficult than the engineering work. Technical problems have solution paths you can work through methodically. Choosing which cognitive signals to capture required weighing evidence, tradeoffs, and constraints simultaneously, and being willing to sit with uncertainty before committing.</p><p>I want to walk through every candidate paradigm I seriously considered, the reasons each one was evaluated, and the reasoning that led to the final selection. This is the part of the design process that is most invisible in the finished product but most important for understanding whether the product is defensible.</p><p>My evaluation criteria were four: established clinical validity (the paradigm has a published evidence base for the conditions PRECISE targets), sensitivity to the specific conditions PRECISE monitors (fatigue, stress, cognitive load, concussion), feasibility in a browser without hardware beyond a standard screen and microphone, and signal multiplicity (the paradigm generates multiple distinct signal types from a single session rather than a single scalar output).</p><h4><strong>Anti-saccade task (selected, visual module)</strong></h4><p>Meets all four criteria at the highest level of any visual task I evaluated. Clinical validity: four decades of published use, covering ADHD, TBI, concussion, aging, schizophrenia, and Parkinson’s. Sensitivity: documented sensitivity to sleep deprivation, concussion, and executive load at subclinical levels. Feasibility: requires only a screen and a cursor or touchscreen. Signal multiplicity: generates mean RT, RT SD, accuracy, and fatigue index from a single 25-trial session.</p><p>The one concern I had was the gap between eye-tracking based anti-saccade paradigms (the standard in research) and cursor-response based implementation (what PRECISE Pro does). Everling and Fischer (1998) review the task’s fundamental mechanism, which is the suppression of a reflexive orienting response and the generation of a voluntary countermove, and that mechanism is engaged regardless of whether the output is an eye movement or a cursor response. The cognitive processes are the same. The motor output is different. I apply conservative norm adjustments to account for the output-measurement gap, and the longitudinal structure of PRECISE means that within-person change remains the meaningful signal even where absolute calibration to lab norms is approximate.</p><h4><strong>Verbal fluency (selected, speech module)</strong></h4><p>Meets all four criteria from a completely different neural angle than the visual task. Clinical validity: decades of use in neuropsychological assessment, including in concussion recovery, aging, and dementia research. Sensitivity: documented sensitivity to cognitive load, fatigue, and language-related processing changes. Feasibility: runs through the browser’s built-in speech API. Signal multiplicity: generates response latency, pause frequency, words per minute, and composite fluency from a single 60-second task.</p><p>The concern was the Web Speech API’s noise floor, specifically the variable transcription delay corrupting latency measurements. I addressed this with the calibration approach described in section 5. The noise is real but manageable, and it is consistent within a session, which is what matters for session-level metric computation.</p><h4><strong>Simple reaction time (in PRECISE Mini, bypassed for Pro)</strong></h4><p>Excellent baseline formation tool. Well-validated. Fast. Easy to implement cleanly in a browser. But its sensitivity to the specific subclinical changes Pro targets is lower than the anti-saccade paradigm because it involves no executive control component. When the dlPFC is stressed, simple RT may show a signal, but it is less specific and less sensitive than anti-saccade RT. Simple RT’s strength is precisely its simplicity, which makes it ideal for Mini’s repeated engagement context and less appropriate for Pro’s signal depth goals.</p><h4><strong>Stroop task (considered, bypassed)</strong></h4><p>Strong candidate. Stroop interference (naming the ink color of a word that spells a different color) is one of the most replicated cognitive control measures in the literature. I was genuinely excited about including it. The problem is implementation: the standard version requires verbal color naming, which demands word-level accurate speech recognition. PRECISE Pro’s speech pipeline can handle spontaneous fluency reasonably well, but word-level forced-choice recognition is a harder problem. In informal testing, the Web Speech API misclassified spoken color names at a rate that would corrupt the interference metric. A button-press version of Stroop is possible, but the evidence base for its psychometric properties is considerably weaker than for the verbal version.</p><p>Stroop remains on the roadmap. It is on the Pro v2 roadmap pending improvements in the speech API’s word-level classification accuracy, or a button-press variant with sufficient published validity.</p><h4><strong>N-back task (considered, bypassed for current version)</strong></h4><p>N-back is an excellent working memory and sustained attention measure with a well-established literature. 2-back specifically has been used extensively in cognitive load research and shows documented sensitivity to fatigue and sleep deprivation. I intend to add it to Pro v2.</p><p>The reason it is absent from the current version is a prioritization decision rather than a conceptual one. I wanted to get the anti-saccade and verbal fluency modules to high signal integrity before adding a third paradigm. Each paradigm added to the session increases total session time, and there is a user engagement tradeoff: a session that takes eight minutes will have higher completion rates than one that takes twelve. Adding N-back before the first two modules were working cleanly would have added complexity without proportional benefit.</p><h4><strong>Trail Making Test (considered, not selected)</strong></h4><p>Trail Making Test Part B, where you alternate between connecting numbers and letters in sequence, is one of the most widely used measures of cognitive flexibility and executive function in neuropsychological assessment. Its browser implementation is feasible and has been done by several research groups. The reason for bypassing it: the signal it captures largely overlaps with the anti-saccade paradigm’s executive control signal, without adding the speech modality’s complementary perspective. Two tasks measuring similar constructs from similar modalities would add session time without proportionally increasing signal breadth.</p><h4><strong>Fine motor coordination (on the roadmap)</strong></h4><p>A stylus-based drawing consistency measure, where you trace a shape or follow a moving target, captures signals from the cerebellar and fine motor planning systems that neither the visual cognitive task nor the speech task accesses. This is on the PRECISE Pro v2 roadmap specifically because it adds a modality dimension that neither current task covers.</p><p>The final selection reflects a specific set of priorities. Two modalities that access completely different neural systems, so that degradation in one is interpretable in the context of the other. Sufficient session depth to generate multiple metrics per modality. Feasibility within the browser constraints. And a published evidence base that can be cited in response to scientific scrutiny.</p><p>I am confident in this selection. I am also honest that other defensible selections exist, and that other researchers building similar tools would make different choices with valid reasons. The task selection for a cognitive monitoring platform is a design decision with scientific constraints, and reasonable scientists can reach different conclusions.</p><h3>14. What PRECISE Pro Is Preparing For</h3><p>PRECISE Pro is Replicate 2 in a program I am in… known as TKS (the world’s best accelerator for teens). It is the proving ground for the assumptions that PRECISE Max requires.</p><h4>(those who know if not already, PRECISE Fusion is being rebranded as PRECISE Max)</h4><p>Understanding what Max will be requires understanding what it needs to be true before it can be built defensibly.</p><p>Max is a multimodal longitudinal prediction engine. It will integrate Mini and Pro signals across many sessions from the same individual, and potentially across many individuals, to build models of what “stable” looks like for each person, detect drift from that stability, and generate early-warning signals with quantified confidence. It will use transformer-based temporal architectures capable of modeling how signals co-vary across time and modalities. It will be, if built correctly, a genuinely novel contribution to the cognitive monitoring space.</p><p>But Max cannot be built on assumptions. It requires validated foundation.</p><p>PRECISE Pro is designed to validate four specific assumptions that Max requires.</p><p><strong>Multimodal signal capture is feasible without dedicated hardware.</strong> This is the most basic assumption and also the one that had to be proved first. Speech and visual signals can be captured with sufficient quality to be longitudinally meaningful using only a standard device with a microphone and screen. Pro is the empirical proof-of-concept for this. Every session completed in Pro that produces clean, plausible data is a data point in favor of this assumption. Every bug discovered and fixed in the signal pipeline is learning that feeds into Max’s architecture.</p><p><strong>AI modeling adds interpretive value beyond raw metrics.</strong> The live linear regression model and PRECISE-IE demonstrate that structured analysis of the captured data produces insights more useful than the raw numbers alone. Pro validates this with a transparent, interpretable approach. The lesson from Pro that will inform Max is not the specific model architecture, but the principle: AI in this context needs to be explainable, grounded, and appropriate to the data scale. Max’s models will be more sophisticated, but that sophistication will be earned by data scale, earned by data scale rather than assumed.</p><p><strong>Users will complete structured multi-component sessions.</strong> Mini validated repeated engagement with lightweight tasks. Pro tests whether users will complete a more demanding, multi-modality session: 25 visual trials plus five 60-second speech tasks plus the intelligence engine upload. Engagement patterns from Pro, completion rates, dropout points, session frequency across users, will directly inform Max’s session design. If a significant fraction of users drop out at the speech module, that is a signal that the session design needs recalibration before Max inherits it.</p><p><strong>Published lab norms apply at an approximate level to browser-captured signals.</strong> The norm library in PRECISE-IE is drawn from controlled lab research. Pro tests whether browser-captured signals cluster in the expected normative ranges for healthy young adults completing the tasks under natural conditions. If healthy young adults consistently score in the normal tiers on all metrics, the norm library’s approximate applicability is confirmed. If they consistently score outside normal tiers, the norms need browser-specific adjustment factors, and Pro is the dataset that enables those adjustments.</p><p>Every session completed in PRECISE Pro is, in a real sense, a contribution to the evidence base that Max’s architecture will be built on. The aggregate signal-to-noise ratios observed in Pro will determine the feature engineering in Max’s fusion model. The session-to-session variability patterns observed in Pro will calibrate Max’s personal baseline algorithms. The engagement patterns will inform Max’s session design.</p><p><strong><em>This is what the staged architecture means in practice. Each layer is not just a product. It is an experiment whose results shape the layer that follows.</em></strong></p><h3>15. The Bigger Vision: From Reactive to Dynamic</h3><p>I want to step back from the architecture entirely for a moment and sit with the question that motivated all of this.</p><p>I am fourteen years old. I am building a neurotechnology platform. I have no lab. I have zero institutional funding. I have no PhD, or a co-founder with a PhD, or a research institution whose credibility I can borrow. I have a laptop, a browser, several months of reading, and a community of people at TKS who push me to think more carefully than I would otherwise.</p><h4><strong>Why am I doing this?</strong></h4><p>The honest answer has <strong>two </strong>parts.</p><p>The <strong>first</strong> part is that the problem is real and unsolved at the scale I care about.</p><blockquote>For context, one of my closest friends recently had her third concussion. What made this one so upsetting was that it was different from the ones she had before. After that happened, our friendship was never quite the same. Another moment that showed me how serious this issue is was when a friend at TKS told me their coach had literally encouraged them to lie during verbal tests, like when asked, “Are you dizzy?”</blockquote><p>I care about youth. I care about students and athletes who push their brains hard and have no visibility into whether the pushing is sustainable. I care about the gap between “I feel fine” and “something is actually wrong” that clinical assessment systems are currently unable to close. I care about the fact that a hockey player can sustain a hit on Friday and lack visibility into whether their executive control is degraded until they get reassessed a week later, if they get reassessed at all. These are real gaps. The tools to address them, in accessible, frequent, affordable form, remain absent at scale.</p><p>The <strong>second </strong>part is that I believe I am actually in a position to contribute to closing them, by building defensible, rigorous pieces of the puzzle that can be handed off, extended, or collaborated on by people with more resources and authority than I have. PRECISE Pro operates upstream of clinical assessment. It is trying to make the space between clinical assessments, the weeks and months where currently nothing is measured, less dark.</p><p><strong>The current model of brain health is reactive. Something goes wrong. You notice. You respond. Assessment happens after the fact. Treatment begins after the fact. The entire clinical infrastructure is organized around identifying problems that have already manifested.</strong></p><p>This model has a fundamental limitation that cannot be resolved from within its own assumptions. By the time something is identifiable as a problem using reactive tools, the underlying process has often been underway for a long time. Neural compensation masks early decline. Broad population-based thresholds miss personal deviation. Single-session snapshots miss trends. The gap between “something is changing” and “something is clearly wrong” can be months or years.</p><h4>PRECISE asks a different question. The question is: is something changing?</h4><p>The old system asks “did something happen?” PRECISE asks “is this brain behaving differently today than it was behaving last week, and is that difference meaningful relative to this individual’s normal pattern?”</p><h4><strong>These are fundamentally different questions, and they require fundamentally different tools.</strong></h4><p>You answering “is something changing?” requires more than a single measurement. It requires more than comparing one person to a population average. It requires a longitudinal history and a personal baseline. PRECISE is built to generate that history, to build that baseline, and to surface meaningful deviation from it in a way that is accessible to a student, an athlete, a coach, or a parent without a clinic visit or a specialist referral.</p><p>The long-term vision, the one I think about when I am working on this at 11 PM after finishing school, is a world where cognitive monitoring is as routine and unremarkable as measuring your resting heart rate. Where a student notices their verbal fluency score drifting over three days before an exam and recognizes it as a sign that they need to sleep, rest instead. Where an athlete catches a subtle inhibitory control decline after a small hit and rests before a real concussion develops. Where a founder notices cognitive instability during a period of extreme work and makes a structural change before burning out.</p><p>These outcomes are achievable. The technology to build toward them exists. The scientific foundation exists. The gap is accessible, well-engineered implementations that are honest about what they are and what they are not.</p><p>PRECISE is building toward closing that gap, one carefully validated layer at a time.</p><h3>16. What Comes Next</h3><p>PRECISE Pro is live and open for use at this <a href="https://precisepro-alizasamirkhoja.netlify.app/">link</a>. Both the speech and visual assessments are fully functional. The PRECISE Intelligence Engine is embedded in the homepage and will analyze any CSV you upload from either module.</p><p>I want to say plainly what “live” means in the context of a research prototype built by a student. The platform works. The signal capture is clean. The intelligence engine functions correctly. But it lacks formal validation in a controlled study. It has been tested only outside clinical populations. It has bypassed peer review at this stage. It is a research prototype that is public because accessibility and feedback are valuable, and because building in isolation from users produces worse tools than building with them.</p><h4>The immediate roadmap for Pro has four priorities.</h4><p><strong>Microphone calibration improvements.</strong> The speech latency measurement is the weakest signal in the current implementation because the Web Speech API’s transcription delay is variable and only partially corrected by the current calibration approach. A more robust calibration protocol, using multiple fixed prompts at session start to estimate delay with higher precision, would reduce the latency noise floor and improve the reliability of latency-based classifications in the intelligence engine.</p><p><strong>Session-linking improvements.</strong> Baseline comparison currently requires manually uploading a prior session’s CSV at the start of each new session. A session code system, where users generate a persistent identifier that links their sessions in the CSV metadata, would make longitudinal trend analysis substantially more accessible. The code would be embedded in every exported CSV, making it trivial to load a prior session for comparison without navigating a file system.</p><p><strong>Research partnerships and pilot design.</strong> I am actively looking for conversations with concussion researchers, youth sports organizations, cognitive neuroscience labs, and academic institutions interested in piloting PRECISE Pro in structured environments. The platform’s CSV export is designed for compatibility with standard statistical software. If you are researching or deploying at the intersection of cognitive health and youth performance and want to explore what a collaboration might look like, find me on LinkedIn.</p><p><strong>PRECISE Max (rebranded from Fusion) </strong>is in early design. The architecture will integrate Mini and Pro signals into a multimodal temporal fusion model, most likely transformer-based given the temporal and cross-modal structure of the data. Early design conversations have focused on the feature representation question: how do you encode Mini’s simple RT time series alongside Pro’s anti-saccade and verbal fluency data into a format that a self-attention architecture can operate on meaningfully? That question remains genuinely open, and the answer matters enormously for whether Max produces defensible predictions or statistical noise.</p><p>The design phase for Max is using feedback from every Pro session to inform these decisions. The signal-to-noise ratios observed in Pro will determine the feature weighting. The session-to-session variability patterns will calibrate the personal baseline algorithms. The engagement patterns will shape the session design.</p><p>This is a long project. I am a fourteen-year-old with no institutional affiliation, no lab, and no co-founder. I am building it in a browser, one carefully chosen piece at a time, with published science as my primary reference and honesty about limitations as my governing design constraint.</p><p>I started this project because a problem had taken hold of me. I have stayed with it because every time I thought I understood it well enough to stop learning, something new emerged. An unread paper. An unaccounted limitation. A feedback conversation that reframed my assumptions. That cycle of not knowing, learning, and rebuilding is what PRECISE is made of.</p><p><strong><em>The goal is to be right. Depth is everything. Speed is essentially irrelevant.</em></strong></p><h3>About Me and About PRECISE</h3><p>I am Aliza Samir Khoja, a fourteen-year-old innovator at The Knowledge Society (TKS). My focus area is neurotechnology and brain-computer interfaces, with a specific interest in proactive cognitive health monitoring and the translation of laboratory neuroscience into accessible tools.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*FKgBhq13WM7XdKso4haU9g.png" /></figure><p>PRECISE (Predictive Real-time Evaluation of Cognitive Instability using Speech and Executive-functioning) is a staged neurotechnology platform I am building to shift brain health from reactive to dynamic. It is designed for youth, athletes, students, and anyone who wants to understand how their brain is doing across time, without requiring clinical infrastructure.</p><p><strong>PRECISE Pro is live at the link above. The PRECISE Intelligence Engine is embedded in the Pro homepage. All data processing is local. Nothing is transmitted. Nothing is stored on any server.</strong></p><p>Enormous thanks to my TKS director Azar, whose sustained pressure-testing of my assumptions and refusal to let me skip depth is the reason PRECISE is built the way it is and not the way I originally imagined it. Thanks to Hannah Conkey and to every practitioner, researcher, and clinician who has taken time for a conversation that shaped the methodology. This project has been built through feedback loops, and the quality of those feedback loops has directly determined the quality of what exists today.</p><p><em>The roadmap: PRECISE Mini (live) → PRECISE Pro (live) → PRECISE Max (in design)</em></p><p><em>If you are researching, building, or working at the intersection of cognitive health, neurotech, and youth performance, I want to hear from you. Find me on </em><a href="http://www.linkedin.com/in/aliza-samir-khoja"><em>LinkedIn</em></a><em> or </em><a href="https://alizasamirkhoja.substack.com/"><em>Substack</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fe6e3b1b7edf" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[PRECISE Mini: A Browser-Based System for Detecting Micro-Level Cognitive Drift]]></title>
            <link>https://medium.com/@khoja.aliza/precise-mini-a-browser-based-system-for-detecting-micro-level-cognitive-drift-046617a4da31?source=rss-5307378d07b3------2</link>
            <guid isPermaLink="false">https://medium.com/p/046617a4da31</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[precise]]></category>
            <category><![CDATA[brain-health]]></category>
            <category><![CDATA[neuroscience]]></category>
            <category><![CDATA[mental-performance]]></category>
            <dc:creator><![CDATA[Aliza Samir Khoja]]></dc:creator>
            <pubDate>Thu, 25 Dec 2025 01:05:41 GMT</pubDate>
            <atom:updated>2025-12-25T15:35:54.147Z</atom:updated>
            <content:encoded><![CDATA[<p><em>“Your brain tells the truth </em><strong><em>long </em></strong><em>before </em><strong><em>you </em></strong><em>realize it.”</em></p><p>That’s the core idea behind PRECISE Mini; a browser-based cognitive tracking system designed to capture the tiny day-to-day fluctuations in reaction time, memory, decision speed, and inhibition that signal how “stable” your brain is at any moment. It might sound dramatic, but neuroscience says otherwise. Before you feel tired, your simple reaction time drifts by a few milliseconds. Before stress becomes conscious, your decision speed slows. Before brain fog hits, your working memory contracts. Before anyone notices you’re off, your inhibitory control slips.</p><blockquote>Your brain whispers before it screams.</blockquote><p>Neuroscience <strong>consistently </strong>shows that millisecond-scale changes in reaction time and variability <strong>often </strong>conscious symptoms across <strong>stress</strong>, <strong>fatigue</strong>, and <strong>mild brain injury (mTBI)</strong>.</p><blockquote><strong>PRECISE Mini is built to listen to the whispers.</strong></blockquote><blockquote>PRECISE Mini is my proof-of-concept pilot within a larger system for tracking cognitive instability in real time.</blockquote><p>Adding to that,<strong> PRECISE </strong>stands for: <strong>P</strong>redictive, <strong>R</strong>eal-time <strong>E</strong>valuation <em>of </em><strong>C</strong>ognitive <strong>I</strong>nstability <em>using </em><strong>S</strong>peech <em>and </em><strong>E</strong>xecutive-functioning…</p><p>To be clear: PRECISE Mini is not a diagnostic tool, and it’s not trying to replace clinicians. It’s designed to sit upstream; before labels, before symptoms, before something “goes wrong.” Its job is simply to surface change early, when intervention is still easy.</p><p>This article breaks down the science, engineering, and reasoning behind PRECISE Mini; and how it fits into the larger PRECISE Fusion framework I’m building for proactive, accessible cognitive health.</p><h4><strong>By the end of this article, you’ll understand how cognitive instability shows up before symptoms… and how it can be measured outside a clinic </strong>↓↓↓</h4><p>(PS., if you want to skip ahead and check out the app go to: <a href="https://precisemini-alizasamirkhoja.netlify.app/"><strong>https://precisemini-alizasamirkhoja.netlify.app/</strong></a><strong>)</strong></p><h4>1. <a href="#cb6f">Introduction</a></h4><p>How the brain changes before you consciously notice; and why invisible cognitive drift is the earliest, most honest biomarker of stress, fatigue, overload, and injury.</p><h4>2. <a href="#5a7b">The PRECISE Framework</a></h4><p>Breakdown of the PRECISE Mini engine (SRT, CRT, Working Memory Span, and Go/No-Go) and the science behind why these micro-tasks detect drift earlier than symptoms.</p><h4>3. <a href="#6c28">Engineering PRECISE Mini: Browser Neuroscience Done Right</a></h4><p>How I built a stable, timing-accurate cognitive testing environment in the browser; 10,000+ lines of code, latency fixes, trial architecture, and precision engineering.</p><h4>4. <a href="#049c">The Philosophy</a></h4><p>Why PRECISE Mini isn’t a medical labeler, but a personal baseline system built to detect “Are you different from your normal?” instead of “Are you concussed?”</p><h4>5. <a href="#a761">The Future</a></h4><p>Where this is going and the long-term vision of cognitive baselines as universal as something like measuring your heart rate.</p><h4>6. <a href="#7f64">Why I Built This</a>…</h4><p>My motivation, the problem I’m solving, how this started, and why PRECISE Mini is only the beginning of a much larger cognitive-health movement.</p><h3>INTRODUCTION</h3><h4>Why Micro-Fluctuations Matter</h4><p>Most people imagine cognitive decline as some dramatic event: a concussion, a panic attack, a total burnout moment.</p><p><strong>But in reality, brains don’t “break” suddenly.</strong></p><p>They leak. A touch more fatigue here. A bit more trouble recalling numbers. A slightly slower reaction time. A few extra inhibition failures. Small enough that you dismiss them; but meaningful enough that, if you measured them, you’d see the early warning signs.</p><p>The truth is that your brain rarely alarms you when something is wrong; it compensates.</p><blockquote>Humans are experts at “pushing through it,” even when the neural load is already too high. That’s how students keep studying, athletes keep playing, and founders keep working … even as the underlying signals degrade and things are quietly compounding overtime.</blockquote><p>PRECISE Mini exists for the exact opposite reason: to capture the <em>micro-stuff</em>, the fluctuations no one feels, the invisible changes that matter more than we admit. These changes are often the earliest biomarkers of sleep debt, concussion, burnout, cognitive overload, emotional stress, neural fatigue, and high-performance breakdown.</p><p>Traditional neuropsych tests capture this too; but the experience is terrible… both<strong><em> metaphorically </em>and <em>literally</em>.</strong></p><p>They’re long, expensive, inaccessible outside clinics, dependent on equipment, and impossible to run regularly. PRECISE Mini flips the model. Instead of a heavy testing battery, it uses four tightly engineered micro-tasks that run in a browser.</p><h3>THE PRECISE FRAMEWORK; Four Tasks, Four Neural Pathways</h3><p>Cognition isn’t one system. It’s a coordination of dozens of neural circuits that shift at different speeds under stress or injury. PRECISE Mini isolates four circuits that research consistently shows are the earliest to fluctuate. Altogether, these tasks span speed, decision-making, memory, and inhibition… four systems that degrade independently, which is why fusion matters more than any single score.</p><p>A single metric can fluctuate for dozens of reasons. Reaction time alone might reflect sleep. Working memory alone might reflect stress. Inhibition alone might reflect emotion. But when multiple systems drift together (even subtly), that pattern is far harder to ignore. PRECISE Mini is built around that principle: signal emerges from convergence, not magnitude.</p><p><strong>Check out the full app at: </strong><a href="https://precisemini-alizasamirkhoja.netlify.app/"><strong>https://precisemini-alizasamirkhoja.netlify.app/</strong></a></p><h3>1. Simple Reaction Time (SRT)</h3><h4>The Nervous System’s “Heartbeat”</h4><p>If your brain had a resting heart rate, this would be it.</p><p><strong>Task:</strong><br>You wait. The screen turns green. You click instantly.<br>Thirty trials… just raw speed.</p><p><strong>What it measures:</strong><br>The pure conduction speed of your retina → occipital cortex → motor cortex → spinal cord → finger. It’s the cleanest human performance metric in existence.</p><p><strong>Why it’s powerful:</strong><br>You cannot strategize or “try harder” to react faster.<br>SRT is brutally honest.<br>It shifts early under:</p><ul><li>acute sleep deprivation</li><li>sedation</li><li>mild head impacts</li><li>metabolic exhaustion</li><li>slowed brainstem arousal</li></ul><p>A drifting SRT is your first instability flag.</p><blockquote>In PRECISE Mini, this task is tracked longitudinally, not judged in isolation.</blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FUQ-Seoe7qS8%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DUQ-Seoe7qS8&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FUQ-Seoe7qS8%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/71701942c3c8f275ad04970711ba4ba1/href">https://medium.com/media/71701942c3c8f275ad04970711ba4ba1/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*t_snpGu2PRXeWt8KDjOAYw.png" /><figcaption>Quick screenshot of how the Simple RT Data CSV would look like!</figcaption></figure><h3>2. Choice Reaction Time (CRT)</h3><h4>The Brain’s Decision Engine Under Load</h4><p>If SRT measures <em>notice speed</em>, CRT measures <em>decision speed</em>.</p><p><strong>Task:</strong><br>Left arrow = click left<br>Right arrow = click right<br>Twenty trials.</p><p>Simple on the surface; but your brain is doing heavy lifting.</p><p><strong>What it measures neurologically:</strong></p><ul><li>stimulus discrimination (parietal cortex)</li><li>response selection (dorsolateral prefrontal cortex)</li><li>motor planning (premotor cortex)</li><li>execution (motor cortex)</li></ul><p><strong>Why it matters:</strong><br>Concussions and stress hit CRT harder than SRT.<br>You can still “feel fine,” play well, or study well; while your CRT is noticeably slower.</p><p>Research shows CRT declines are early markers of:</p><ul><li>stress</li><li>cognitive overload</li><li>concussion</li><li>mental fatigue</li><li>emotional dysregulation</li></ul><p>This is why PRECISE Mini uses CRT as a core pillar.</p><blockquote>In PRECISE Mini, this task is tracked longitudinally, not judged in isolation.</blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FaaENN-UgLVw%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DaaENN-UgLVw&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FaaENN-UgLVw%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/aedd77a3b05863a8d31a25c5bc6311d3/href">https://medium.com/media/aedd77a3b05863a8d31a25c5bc6311d3/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*TYzkJoDZGsbfH0NU9sUoRA.png" /><figcaption>Quick screenshot of how the Choice RT Data CSV would look like!</figcaption></figure><h3>3. Working Memory Span</h3><h4>The Brain’s Mental Buffer</h4><p>Working memory is what lets you hold and manipulate information for a few seconds; the foundation of problem solving, reading, reasoning, planning.</p><p><strong>Task:</strong><br>A sequence of digits appears.<br>It disappears.<br>You type it back; forward OR backward.<br>Fifteen trials.</p><p><strong>Neural basis:</strong></p><ul><li>phonological loop (storing digits)</li><li>visuospatial sketchpad (tracking sequence)</li><li>central executive (prefrontal cortex)</li></ul><p><strong>Why it’s sensitive:</strong><br>Working memory collapses early under:</p><ul><li>concussion</li><li>overwhelm</li><li>burnout</li><li>sleep deprivation</li><li>emotional stress</li><li>anxiety</li><li>fatigue</li></ul><p>“Brain fog” is <em>essentially </em>working memory instability. PRECISE Mini catches this collapse long before you consciously feel it.</p><blockquote>In PRECISE Mini, this task is tracked longitudinally, not judged in isolation.</blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F--zBcQawvDM%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D--zBcQawvDM&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F--zBcQawvDM%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/ca816ffd13d2c39e7702c6ceac212a2f/href">https://medium.com/media/ca816ffd13d2c39e7702c6ceac212a2f/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6J1yC6zBJ6IogeYTZHvMIQ.png" /><figcaption>Quick screenshot of how the Memory Span Data CSV would look like!</figcaption></figure><h3>4. Go/No-Go</h3><h4>Your Brain’s Self-Control and Inhibition Circuit</h4><p>If SRT is speed and WM is clarity, Go/No-Go is control.</p><p><strong>Task:</strong><br>Green = tap<br>Red = don’t tap<br>Twenty trials.</p><p>Simple; but prefrontal inhibition is one of the first systems to destabilize.</p><p><strong>Neural basis:</strong></p><ul><li>orbitofrontal cortex (inhibition)</li><li>anterior cingulate (conflict monitoring)</li><li>basal ganglia (action gating)</li></ul><p><strong>Why this matters:</strong><br> High false-alarm rates often appear when someone is:</p><ul><li>stressed</li><li>anxious</li><li>overloaded</li><li>impulsive</li><li>sleep-deprived</li><li>recovering from mild brain injury</li></ul><p>This single task can reveal executive dysfunction before symptoms appear.</p><blockquote>In PRECISE Mini, this task is tracked longitudinally, not judged in isolation.</blockquote><blockquote>Up to this point, everything you’ve read describes the “why.”<br>What follows is the “how.”</blockquote><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FoxWoee355WA%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DoxWoee355WA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FoxWoee355WA%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/0eb7fd273460624ebf893cef6f472fd2/href">https://medium.com/media/0eb7fd273460624ebf893cef6f472fd2/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*SxKDgr8QtWqeBlzAb44xLQ.png" /><figcaption>Quick screenshot of how the GNG Data CSV would look like!</figcaption></figure><h3>ENGINEERING PRECISE MINI</h3><h4>Timing, Latency, and Browser Neuroscience</h4><p>Designing cognitive tests is easy.<br>Designing <em>accurate</em> cognitive tests in a browser is insanely hard.</p><p>Browsers have fluctuating event loops, inconsistent frame rendering, variable refresh rates, and input latency differences across devices. In early versions, a 16ms frame delay was enough to mask real reaction-time shifts… forcing a full rewrite of the timing engine.</p><blockquote><strong>PRECISE Mini required <em>10 000+ lines of code</em> across prototypes because every small inconsistency ruins the measurement.</strong></blockquote><p>I rebuilt entire subsystems multiple times after discovering timing drift, edge-case crashes, and input artifacts that made the data unreliable.</p><p>PRECISE Mini solves these issues with:</p><ul><li>high-resolution timing</li><li>strict stimulus preloading</li><li>locked, jitter-resistant response windows</li><li>no-layout-shift UI to avoid rendering delays</li><li>trial-averaging to smooth noise</li><li>variance metrics (not just mean RT)</li><li>consistent CSV export for analysis</li></ul><p>Most people think reaction time is “how fast you click.” It’s not. It’s 90% engineering, 10% neuroscience.</p><p>My goal was to create a <em>stable testing environment inside a browser</em>, something even many adult-engineered apps fail at.</p><p>The reason this matters is access (and me learning the skills needed to scale this to pro and fusion). A browser-based system removes clinics, equipment, scheduling, and cost from the equation. It makes cognitive baselining something you can do weekly, casually, and longitudinally… which is exactly what early detection requires. Precision only matters if people can actually use it.</p><blockquote>PRECISE Mini is not immune to noise. Browser hardware, input devices, and user context introduce variability. That’s why Mini focuses on longitudinal trends and variance, not single-session scores. These limitations are also what motivated the move toward multimodal fusion in later versions.</blockquote><h3>THE PHILOSOPHY</h3><h4>This Isn’t Diagnosis. It’s Detection.</h4><blockquote><em>Medical concussion tests are built to answer one question:</em><br><strong>“Are you concussed?”</strong></blockquote><blockquote>— — —</blockquote><blockquote><em>PRECISE Mini is built to answer a much more important question:</em><br><strong>“Are you different from your normal?”</strong></blockquote><p>All PRECISE systems are designed around local data ownership, user consent, and trend-based interpretation… not centralized scoring or opaque risk labels.</p><p>Brains vary wildly between people, but they are extremely consistent <em>within</em> a person. That’s why PRECISE Mini focuses on:</p><ul><li>baselines</li><li>trends</li><li>variability</li><li>deviation</li><li>fusion across tasks</li></ul><p>A CRT of 380 ms means nothing alone.<br>But if your baseline is 290 ms?<br>That’s a massive red flag.</p><blockquote><strong>Going more in-depth to that:</strong> a student whose CRT baseline sits at ~280 ms may show a gradual rise to 310 ms across three days. Individually, each score looks “normal.” Longitudinally, the trend signals instability.</blockquote><p>Working memory of 6 digits might seem “fine.”<br>But if you normally do 8?<br>Your brain is struggling.</p><blockquote><em>Most existing tools measure cognition at a single moment; PRECISE Mini is built to measure change over time.</em></blockquote><p>PRECISE Mini doesn’t label you.<br>It detects micro-instability; the earliest signal that something shifted.</p><h3>THE FUTURE</h3><h4>PRECISE Fusion and Multimodal Cognitive Health</h4><p>PRECISE Mini is the micro-task core of a much bigger system I’m now building: <strong>PRECISE Fusion</strong>, a multimodal cognitive-health engine that combines performance, vocal markers, reaction dynamics, micro-motor signals, and day-to-day fluctuations.</p><p>The vision is a <strong>Cognitive Vital Sign Dashboard; </strong>something students, athletes, founders, and high-performers can use to track how their brain is actually doing each day. Especially for young brains, where change happens fast and damage compounds quietly.</p><p>I’m already beginning conversations with:</p><ul><li>concussion researchers</li><li>youth neurotech labs</li><li>Alberta sports clinicians</li><li>founders building mental-performance tools</li></ul><p>👆If you’re building, researching, or deploying tools at the intersection of cognition, performance, or youth neurotech… let’s talk <a href="https://www.linkedin.com/in/aliza-samir-khoja/">here</a>!</p><p>The roadmap: <strong>Mini → Pro → Fusion → full predictive interface.</strong></p><p>The long-term vision is simple: <strong>In the future, everyone should know their cognitive baseline the same way they know their resting heart rate.</strong></p><p>Imagine a world where a student notices their CRT rising over three days and realizes they’re heading toward burnout <em>before</em> their exam week hits.</p><p>Or an athlete catches a subtle inhibition drop after a small impact and rests before a real concussion develops.</p><p>Or a founder catches cognitive instability during extreme workload weeks and adjusts before hitting a wall.</p><p>That’s not a decade away.<br>We can build that now; and <strong>PRECISE Mini is the first brick.</strong></p><h3>WHY I BUILT THIS</h3><h4>…And What Comes Next</h4><p>I built PRECISE Mini because cognitive decline often hides behind “I’m fine.” Because athletes keep playing. Students keep pushing. People keep ignoring the signal until it becomes a symptom.</p><p>I wanted to build something that listens to the small shifts, the quiet warnings, the early signs; something accessible enough that a 7-year-old student, a parent, or a coach could run it in minutes.</p><p>Cognition shouldn’t be a mystery.<br>It should be a vital sign.</p><p><strong>PRECISE Mini is the proof.<br>PRECISE Pro is the innovation.<br>PRECISE (Fusion) Max is the future.</strong><br>And I’m building it… one micro-signal at a time.</p><blockquote><em>Because the future of brain health isn’t reactive.<br>It’s proactive + aware.</em></blockquote><h4>Where else can you find me?</h4><ul><li>On <a href="http://www.linkedin.com/in/aliza-samir-khoja"><strong>LinkedIn</strong></a>, I share short, sharp reflections every <strong>1–2 weeks</strong></li><li>On <a href="https://alizasamirkhoja.substack.com/"><strong>Substack</strong></a>, you’ll get posts <strong>every 2–3 weeks</strong> filled with insight, action, and the kind of growth no classroom teaches</li><li>And right here on <a href="https://medium.com/@khoja.aliza"><strong>Medium</strong></a>, I post deeper thought and research pieces about every <strong>2 months</strong> (or more during school breaks!)</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=046617a4da31" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Detecting the Invisible: How AI Can Transform Youth Concussion Care]]></title>
            <link>https://medium.com/@khoja.aliza/detecting-the-invisible-how-ai-can-transform-youth-concussion-care-f76d11eac2f3?source=rss-5307378d07b3------2</link>
            <guid isPermaLink="false">https://medium.com/p/f76d11eac2f3</guid>
            <category><![CDATA[concussions-and-the-brain]]></category>
            <category><![CDATA[youth-athelete-safety]]></category>
            <category><![CDATA[early-detection]]></category>
            <category><![CDATA[ai-and-tech-in-medicine]]></category>
            <category><![CDATA[limits-of-current-tools]]></category>
            <dc:creator><![CDATA[Aliza Samir Khoja]]></dc:creator>
            <pubDate>Sun, 28 Sep 2025 20:07:44 GMT</pubDate>
            <atom:updated>2025-12-25T15:37:29.543Z</atom:updated>
            <content:encoded><![CDATA[<h3>AI-Driven Detection of Micro-Cognitive Fluctuations: Redefining Youth Concussion Care</h3><p>One in five youth athletes will get a concussion at least once in their lifetime (Gawrys et al., 2025).</p><p><strong>Surprisingly enough, most of them won’t even know it.</strong></p><p>Sideline checks look for headaches, dizziness, or nausea. They don’t notice the small things; the tiny glitches in memory or focus that flicker before anything obvious appears. Those glitches are invisible. But they exist. <strong>And they matter.</strong></p><p>Artificial intelligence can see patterns that are too subtle for humans. Small shifts in attention, tiny drops in <a href="#ad79">working memory</a>, micro-fluctuations that hint the brain isn’t ready. Suddenly, what was invisible becomes visible. Risks you didn’t know were there can be avoided. Recovery can start sooner.</p><p>Here’s what you can expect to learn about throughout this article;</p><ul><li><a href="#5198"><strong>Why Working Memory and Attention Matter</strong></a><strong>: </strong>About the early cognitive disruptions after mild injury</li><li><a href="#b838"><strong>Limitations of Current Concussion Tools</strong></a><strong>: </strong>Why existing methods miss subtle warning signs</li><li><a href="#7839"><strong>A New Framework: AI for Micro-Fluctuation Detection</strong></a><strong>: </strong>How AI can capture our brain’s “invisible” patterns, cognitive task data streams, machine learning models for detection, personalized baselines and youth-friendly formats</li><li><a href="#7f05"><strong>Industry Landscape and Gaps</strong></a>: Current players, progress, and shortcomings</li><li><a href="#815b"><strong>Challenges and Barriers</strong></a><strong>: </strong>Data, ethics, and adoption hurdles</li><li><a href="#a241"><strong>Future Directions</strong></a><strong>: </strong>Multimodal fusion and broader applications</li><li><a href="#9b69"><strong>The Road Forward</strong></a><strong>: </strong>Shifting concussion care from reactive to proactive</li><li><a href="#3e32"><strong>References and Glossary</strong></a><strong>: </strong>Supporting resources and definitions</li><li><a href="#b186"><strong>Author’s Note</strong></a><strong>: </strong>About me (the author) and how you can support me and my journey ;)</li></ul><h3>Why Working Memory and Attention Matter</h3><p><a href="#ad79">Working memory</a> and attention are often the first <a href="#bb90">executive functions</a> to show disruption after <a href="#6465">mild traumatic brain injury</a> (Mavroudis et al., 2024). Even relatively small impacts can impair the <a href="#61b7">prefrontal cortex</a> and <a href="#0ad1">parietal lobes</a>, which normally coordinate to sustain these neurological processes.</p><blockquote>Shah-Basak et al. (2017) explain this as, “Although injuries sustained after the concussion were considered “mild,” these data suggest that they can have prolonged effects on early attentional and VWM processes.” For context, VWM means Visual Working Memory.</blockquote><p>At first, a child with a concussion may perform normally on standard tests, masking the injury and making detection difficult. Over time, inconsistencies such as lapses in focus, slower reaction times, and forgotten details begin to appear. These micro-fluctuations are invisible to the naked eye, yet reveal critical information about the brain’s stability. Without tools capable of detecting such patterns, athletes are often cleared to play while their brains remain vulnerable to permanent damage.</p><h3>Limitations of Current Concussion Tools</h3><p>Despite widespread use, existing concussion assessments are too coarse to detect early cognitive disruption. Baseline tests like <a href="#73e0">ImPACT</a> are designed to flag only large drops in performance, missing subtle trial-by-trial shifts that signal early dysfunction.</p><p>Sideline tools such as the <a href="#8f2c">SCAT5</a> rely heavily on self-reported symptoms. Athletes are often asked questions such as “Do you have a headache?” or “Do you feel dizzy?” Many athletes underreport or minimize these symptoms to remain in play, which reduces reliability.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*31IGWElRSfF2u0Uq.jpg" /><figcaption>Components of the <a href="#8f2c">SCAT5</a> Test taken from; Iverson et al., (2021), demonstrate SCAT5&#39;s emphasis on subjective responses rather than continuous data.</figcaption></figure><p>Wearables such as helmet sensors or accelerometers provide useful data about impacts, but they only track external forces, not internal brain changes. These devices may show that a hard hit occurred but cannot reveal whether an athlete’s cognitive function has been compromised. Collectively, these approaches miss fine-grained variability in performance over time, which means that concussions without obvious symptoms often go unnoticed and untreated.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*yOGy2_5cW9LlItjU768UtQ.png" /><figcaption>This screenshot from a <a href="https://www.sportswearable.net/5-best-wearables-for-protection-against-concussion/">Sports Wearable</a> website showcases FITGuard, an example of a wearable device. It highlights how current wearables in the field are primarily designed to detect external forces, proving the limitations of existing technology.</figcaption></figure><h3>A New Framework: AI for Micro-Fluctuation Detection</h3><p>AI-driven approaches provide a pathway to detect invisible patterns that traditional tools overlook. Instead of focusing on average scores or obvious impairments, the emphasis shifts to trial-to-trial variability. Smartphone or tablet-based microtasks can serve as data collection tools. Tasks such as the <a href="#0cf0">n-back test</a>, the <a href="#70c0">Stroop task</a>, and sustained attention tests such as <a href="#7088">Go/No-Go</a> can be administered in youth-friendly formats. These digital tasks generate streams of data that reflect reaction times, <a href="#8d4b">recall accuracy</a>, and attention consistency.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Jz7jb_s5uW42nv_h3LzloA.png" /><figcaption>Visual summary comparing various tasks designed to test memory and attention, sourced from <a href="http://researchgate.net/figure/The-experiment-paradigms-Three-tasks-N-back-Stroop-and-verbal-fluency-having-three_fig3_335662322?__cf_chl_rt_tk=JF6VbxXuR8NVAZTv4epaTBLTPxSUtDwK7SeJU2QTfZ4-1759096712-1.0.1.1-z94wPVCZIXT5iQT9Sigb0j8q8Ih4AoQp7V4cbkJUe_8">ResearchGate</a>. It provides a clear overview of how different tasks assess these cognitive functions.</figcaption></figure><p>The key signals of interest include slight variability in <a href="#8d4b">recall accuracy</a>, small reaction time delays measured in milliseconds, <a href="#99d7">clusters of errors</a> appearing in bursts, and occasional lapses where the participant fails to respond.</p><p>Machine learning models can analyze these features by comparing an athlete’s performance to their <a href="#2542">personalized baseline</a> rather than using <a href="#bf5b">normative data</a>. To minimize <a href="#0abc">sandbagging</a>, the system would collect baseline data <strong>continuously over time</strong>, rather than relying on a single pre-season session. Because the AI learns each athlete’s natural performance variability, sudden underperformance during a concussion evaluation would stand out as an anomaly rather than resetting the baseline. Moreover, techniques such as <a href="#e48e">time-series anomaly detection</a> and <a href="#1ba5">variability trend analysis</a> shall allow the system to identify early drift away from healthy performance. When these subtle deviations appear, the system can flag the athlete as at risk, prompting further medical evaluation.</p><p>The envisioned output is a simple dashboard for coaches, parents, or doctors. The purpose is <strong>not medical diagnosis</strong>, it is early warning and giving adults <strong>a chance to act</strong> before problems escalate. AI systems do not simply register slower reaction times or missed answers in isolation; they quantify how these changes emerge relative to the participant’s historical baseline. Reaction time distributions, for example, can be modeled as probability curves rather than single averages. A concussion may shift the curve slightly to the right, indicating slower responses, but the more sensitive marker is often the increase in variance, which signals unstable neural processing rather than a uniform slowdown. Missed responses are not treated as singular mistakes but as part of an <a href="#99d7">error cluster analysis</a>, where the frequency and <a href="#398f">temporal spacing of omissions</a> reveal whether attentional control is degrading systematically or only under irregular/sporadic distraction.</p><p>AI approaches concussion detection by breaking down trial-by-trial task data into millisecond-scale features. Each attempt in tasks such as the <a href="#0cf0">n-back</a> or <a href="#70c0">Stroop</a> yields multiple measurable outputs, including response speed, accuracy, and attention consistency. Instead of averaging scores, these features are treated as time series so that algorithms can track fluctuations across trials. Within these streams, methods detect abrupt shifts in mean or variance, <a href="#99d7">bursts of consecutive errors</a>, and missed or abnormally slow responses (Rosenblatt et al., 2021). The most predictive features of concussion-related disruption include, but are not limited to; reduced average accuracy, increased variability within individuals, clusters of mistakes, and gradual drift away from an established <a href="#2542">personal baseline</a>.</p><h3>Machine Learning Models for Concussion Detection</h3><p>Certain machine learning models are particularly well suited to analyzing cognitive task data. <a href="#5229">Recurrent neural networks</a> (RNNs) and <a href="#a889">transformer</a>-based architectures capture sequential patterns, making them effective for identifying how small trial-level variations evolve into broader instability (McLeod et al., 2025). <a href="#e48e">Time-series anomaly detection</a> methods, such as <a href="#c89b">sliding-window Z-score tracking</a> and <a href="#1ead">autoencoder-based outlier detection</a>, enhance <a href="#3822">sensitivity</a> to subtle fluctuations by flagging deviations at the individual level.</p><p>The use of AI for continuous trial-level monitoring could shift concussion care from reactive to proactive. Real-time flagging gives coaches and medical staff immediate alerts when a child’s performance begins to deviate from baseline, enabling evaluation before symptoms escalate. This contrasts with current systems that often wait until multiple concussions or noticeable impairments surface. <a href="#e2bb">Longitudinal tracking</a> also supports individualized recovery plans, especially with <a href="#33ed">sub-concussive impacts</a> being common. Physicians can evaluate whether micro-fluctuations are stabilizing over time rather than relying on one-off post-injury tests. This approach reduces the risk of both premature return-to-play and unnecessary restriction, producing safer, evidence-based timelines.</p><p>Children’s cognitive capacities evolve rapidly with age, which complicates the interpretation of concussion data. AI systems address this by dynamically recalibrating baselines. Rolling averages and adaptive models incorporate the most recent non-injury data, capturing developmental improvements in <a href="#ad79">working memory</a> and attention. This baseline-centered approach provides greater accuracy than comparisons against static population norms, which fail to reflect individual growth or variability. By anchoring results to each child’s personal trajectory, AI avoids the pitfalls of a one-size-fits-all system and ensures that both injury detection and recovery monitoring remain developmentally sensitive.</p><p>For broad adoption, assessments must be engaging and easy to deploy. Gamified mobile tasks, designed around role-play or rapid-response challenges, increase compliance by making frequent testing appealing to youth. Smartphones and tablets are especially effective platforms because they do not require specialized hardware, can be used in real-world settings such as schools or sports leagues, and provide instant data transmission. Compared to EEGs or lab-based systems, which are costly, time-consuming, and impractical for large populations, mobile AI solutions make continuous monitoring both scalable and accessible.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*U5O-6sNMMaZ63gw5.png" /><figcaption>An example of a fun, gamified, self-care platform called Finch. Taken from <a href="https://au.lifehacker.com/apps/112975/opinion/finch-motivates-me-to-complete-my-to-do-list-without-shaming-me">LifeHacker Australia</a>.</figcaption></figure><p>After all, the framework for AI in concussion detection rests on three pillars: precise signal capture from mobile cognitive tasks, sophisticated <a href="#398f">temporal modeling</a> to identify subtle instability, and clinically meaningful outputs that prioritize early warning over formal diagnosis. Furthermore, by now we all understand that this integration has the potential to redefine concussion care by making invisible injuries visible, shifting intervention timelines forward, and tailoring recovery protocols to the <a href="#8ea5">developmental trajectories</a> of young athletes. The unanswered questions; which features are most predictive, which models are most robust across populations, and how compliance can be maximized, will guide the next phase of research, but the foundation is already quite clear: <strong>AI is uniquely capable of extracting diagnostic value from the </strong><a href="#0a75"><strong>noise</strong></a><strong> of everyday variability</strong>.</p><h3>Industry Landscape and Gaps</h3><p>The current industry landscape illustrates progress but also major gaps. Widely used tools such as <a href="#8f2c">SCAT5</a> and <a href="#73e0">ImPACT</a> (see section <em>Limitations of Current Concussion Tools</em>) remain the baseline industry standards. Both function as snapshots rather than continuous monitoring systems. Their reliance on self-reports, population norms, or large score drops leaves fine-grained variability undetected.</p><p>Companies like Cogstate provide computerized batteries to assess attention, memory, and decision-making. Cambridge Cognition’s CANTAB offers highly validated cognitive assessments often used in research. These platforms have strong scientific foundations but generally aggregate results into averages, which reduces <a href="#3822">sensitivity</a> to trial-level fluctuations.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Nlwv_lPxQUe3viJiN_pzrA.png" /><figcaption>A figure from <a href="https://www.researchgate.net/figure/Cogstate-test-battery-task-descriptions-1_tbl1_333193114">ResearchGate</a> shows the Cogstate Battery Task Descriptions, highlighting the approach of these tools.</figcaption></figure><p>Other firms target different signals. SyncThink uses eye-tracking to identify concussion-related impairments, while NeuroCatch (Canada) leverages EEG to track brain function. HitCheck delivers smartphone-based neurocognitive testing in field settings, but its emphasis remains on overall scores rather than micro-variability.</p><p>Even new AI-powered models, such as those developed at the University of Delaware and highlighted in sports media, focus on diagnosing concussions after they occur or predicting secondary injuries. These models rarely aim to detect micro-patterns of instability beforehand. In contrast, the framework outlined here emphasizes trial-to-trial fluctuations, integrates multiple signals from <a href="#ad79">working memory</a> and attention, and applies personalized <a href="#e2bb">longitudinal analysis</a>. The design is proactive rather than reactive, with accessibility supported through gamified smartphone or tablet tasks.</p><h3>Challenges and Barriers</h3><p>Although promising, this approach faces several challenges. <a href="#e520">Confounding</a> factors can make it difficult to determine whether variability is due to concussion or everyday influences. For example, a tired athlete may show slower reaction times even without a brain injury.</p><p>Reliable baseline data presents another obstacle. Many youth athletes do not undergo preseason testing, and natural development means that baselines shift with age and learning. This complicates the task of distinguishing injury-induced changes from normal growth.</p><p>Small sample sizes and inconsistent follow-up limit the scope of youth concussion studies. These gaps restrict the ability to train robust AI models that generalize across populations. Ethical concerns are equally pressing. Neurocognitive data from minors is highly sensitive, raising questions of privacy, consent, and fairness if datasets reflect socioeconomic or access-based bias.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/601/0*Id93hGtrHuGBrHZs.png" /><figcaption>An infographic explaining what to take in consideration to ensure ethical AI models. Taken from: <a href="https://www.probe.group/insights/why-ethics-matter-in-the-race-to-develop-ai">Probe Group</a></figcaption></figure><p>Finally, clinical adoption also remains a barrier. Even strong research results do not guarantee that parents, coaches, and clinicians will trust or use new tools. Without transparency and careful validation, new systems risk being seen as useless apps.</p><h3>Future Directions</h3><p>Several promising directions could expand this approach. One is multimodal fusion, where cognitive micro-fluctuation data is combined with other signals such as <a href="#25f9">biomarkers</a>, wearable motion data, or eye-tracking streams. Integrating these sources may help distinguish concussion-related variability from <a href="#e520">confounds</a> like fatigue or stress.</p><p>Another direction involves creating youth-friendly dashboards for parents and coaches. Complex AI outputs could be translated into simple, actionable visuals such as trend lines that highlight growing inconsistency in memory or attention compared to a child’s baseline. These tools would make data usable for non-specialists while still supporting clinical oversight.</p><p>The same methods may also extend beyond sports. In neurodegenerative conditions such as Alzheimer’s or Parkinson’s, <a href="#cd1b">intra-individual variability</a> already serves as an early <a href="#25f9">biomarker</a>. Applying micro-fluctuation analysis to concussion care could also contribute to broader neurological health, bridging pediatric injury prevention with long-term monitoring of brain function.</p><h3>The Road Forward</h3><p>Youth concussions remain a hidden crisis because the earliest signs emerge not in visible symptoms but in millisecond-level disruptions to <a href="#ad79">working memory</a> and attention. Traditional tools such as <a href="#8f2c">SCAT5</a> and <a href="#73e0">ImPACT</a> can rule out major injuries, yet they fail to capture these subtle fluctuations. AI-driven analysis of trial-to-trial variability offers a clear path forward: continuous, individualized monitoring that transforms concussion care from reactive diagnosis to proactive prevention.</p><p>The obstacles ahead; limited datasets, risks of algorithmic bias, and barriers to adoption in schools and clinics, are significant but solvable. What matters is that the technical foundation already exists. By converting <a href="#0a75">noise</a> in everyday variability into meaningful signals, AI can give athletes, coaches, and physicians a reliable early warning system. Current tools rely on explicit testing. The next frontier is passive neural fingerprinting: using natural speech and reflexive cough patterns, fused with task variability, to detect concussions earlier and more reliably than ever before. This shift has the potential to shorten recovery timelines, reduce repeat injuries, and protect the developing brains of young athletes. If implemented at scale, it could mark the end of undiagnosed concussions in youth sports and lay the groundwork for broader neurological health monitoring in the future.</p><h3>Author’s Note</h3><p><strong>About Aliza Samir Khoja: </strong>Hi, I’m Aliza Samir Khoja. I’m a student researcher at The Knowledge Society (TKS), where I explore the intersection of neuroscience, AI, and youth brain health. Right now, I’m focused on developing AI-driven approaches to detect subtle changes in working memory and attention after concussions. I write to share what I’m learning, raise awareness, and spark conversations about how emerging technologies can protect kids and teens in sports and everyday life. One day, when concussion care is proactive instead of reactive, I want people to look back a and say this began with a 13-year-old who refused to let invisible injuries keep stealing futures, and chose to build a world where no potential is overlooked.</p><h4>Where else can you find me?</h4><ul><li>On <a href="http://www.linkedin.com/in/aliza-samir-khoja"><strong>LinkedIn</strong></a>, I share short, sharp reflections every <strong>1–2 weeks</strong></li><li>On <a href="https://alizasamirkhoja.substack.com/"><strong>Substack</strong></a>, you’ll get posts <strong>every 2–3 weeks</strong> filled with insight, action, and the kind of growth no classroom teaches</li><li>And right here on <a href="https://medium.com/@khoja.aliza"><strong>Medium</strong></a>, I post deeper thought and research pieces about every <strong>2 months</strong> (or more during school breaks!)</li></ul><h3>References and Glossary</h3><p>Here is a pdf link to my references document: <a href="https://drive.google.com/file/d/1fYZqUidj1_QWu2HsUWx5v8lpS5iCMsaJ/view?usp=sharing"><em>https://drive.google.com/drive/u/0/folders/1sbJWkSOJ5YDGwhC37Z6sgTk8DXJ6qm7</em></a></p><p><strong>Definitions Glossary</strong></p><p>All the terms you need to understand to get through the technical jargon!</p><ul><li><strong>Mild traumatic brain injury (mTBI)</strong>: A brain injury caused by a bump, blow, or jolt to the head that disrupts normal brain function, often temporarily.</li><li><strong>Recall Lapses</strong>: Moments when an individual cannot retrieve needed information from their memory.</li><li><strong>Recurrent neural networks</strong>: a type of deep learning model designed to process sequential data, like text or time-series, by using feedback loops to remember past information.</li><li><strong>Feedback Loops</strong>: processes where the output or result of a system is fed back as an input, influencing future actions and leading to a continuous cycle of self-regulation or change</li><li><strong>Neural maturation</strong>: The gradual process where the developing brain strengthens important pathways, eliminates inefficient ones, and stabilizes into adult-like function.</li><li><strong>Second-impact syndrome</strong>: A rare but life-threatening brain swelling that happens if a person gets a second concussion before the first has healed.</li><li><strong>Executive functions</strong>: High-level brain processes that help with attention, planning, decision-making, and controlling impulses.</li><li><strong>Working memory</strong>: The brain’s “scratchpad;” the short-term system that lets you hold and use small amounts of information (like remembering a phone number while dialing).</li><li><strong>Prefrontal cortex</strong>: Brain region behind the forehead, crucial for decision-making, planning, and self-control.</li><li><strong>Parietal lobes</strong>: Regions near the top/back of the brain, important for spatial awareness and integrating sensory information.</li><li><strong>Transformer model:</strong> a type of neural network that excels at processing sequential data, like text, by using attention mechanisms to understand relationships between different parts of the sequence</li><li><strong>SCAT5 (Sport Concussion Assessment Tool 5th edition)</strong>: A sideline checklist used by coaches/medical staff to quickly screen athletes for concussion symptoms.</li><li><strong>ImPACT (Immediate Post-Concussion Assessment and Cognitive Testing)</strong>: A computerized test athletes take before and after injuries to compare memory, reaction time, and attention.</li><li><strong>Sandbagging</strong>: When athletes purposely do poorly on a baseline test so that later concussion test scores look “normal.”</li><li><strong>n-back test</strong>: A memory test where participants must remember a sequence of items and identify when the current item matches one seen “n” steps earlier.</li><li><strong>Stroop task</strong>: A test of attention where you name the color of a word’s font (e.g., the word “RED” printed in blue ink). Concussion slows responses here.</li><li><strong>Go/No-Go test</strong>: A task where participants must respond quickly to certain signals (“Go”) and withhold responses to others (“No-Go”), measuring impulse control and sustained attention.</li><li><strong>Time-series anomaly detection</strong>: An AI technique that looks for unusual patterns over time compared to a person’s normal baseline.</li><li><strong>Sliding-window Z-score tracking</strong>: Tracks how much a new data point deviates/changes from the recent average, not the entire history.</li><li><strong>Autoencoder-based outlier detection: </strong>A technique that uses a type of neural network to find unusual data points in a dataset.</li><li><strong>Error clustering</strong>: When mistakes happen in bursts instead of randomly, signaling instability in brain function.</li><li><strong>Developmental trajectories</strong>: the path or progression of an individual’s development over time, describing the sequence and pattern of changes in a specific behavior, skill, or domain (like cognitive or social) across the lifespan.</li><li><strong>Temporal spacing of omissions</strong>: refers to the timing, or the intervals between, when an expected event fails to occur.</li><li><strong>Cognitive biomarker</strong>: A measurable signal (like reaction time variability) that reflects underlying brain health.</li><li><strong>Intra-individual variability</strong>: The moment-to-moment fluctuations in one person’s performance (rather than average group scores).</li><li><strong>Confounds</strong>: Factors that interfere with results (like fatigue, stress, or lack of sleep), making it harder to know if changes are due to concussion.</li><li><strong>Post-concussion syndrome (PCS)</strong>: A condition where concussion symptoms (headaches, difficulty concentrating, dizziness) persist for weeks or months after the injury.</li><li><strong>Sub-concussive impacts</strong>: Head impacts that don’t cause noticeable concussion symptoms but may accumulate damage over time.</li><li><strong>Longitudinal monitoring</strong>: Following the same athlete’s cognitive performance repeatedly over time to detect gradual changes.</li><li><strong>Normative data</strong>: Standardized scores from a large population used as a comparison point for individual performance.</li><li><strong>Personalized baseline comparison</strong>: Using an athlete’s own pre-injury test results as the reference point, rather than comparing them to population averages.</li><li><strong>Variability trend analysis</strong>: Tracking whether fluctuations in cognitive performance are increasing over time; a possible sign of early brain dysfunction.</li><li><strong>Noise</strong>: Random fluctuations in data caused by distractions, fatigue, or measurement error, which can hide real concussion effects.</li><li><strong>Sensitivity</strong>: A test’s ability to correctly identify athletes who really have a concussion (true positives).</li><li><strong>Specificity</strong>: A test’s ability to correctly identify athletes who are healthy (true negatives).</li></ul><p>Here is a pdf link to the glossary above: <a href="https://drive.google.com/file/d/1ra8iJ6rLY7fnybfRKHOzLa42DDJgHZ86/view?usp=sharing"><em>https://drive.google.com/file/d/1ra8iJ6rLY7fnybfRKHOzLa42DDJgHZ86/view?usp=sharing</em></a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f76d11eac2f3" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>