<?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 Anitha Kannan on Medium]]></title>
        <description><![CDATA[Stories by Anitha Kannan on Medium]]></description>
        <link>https://medium.com/@anithakan?source=rss-891e9f27a4cf------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*r9e5hqYpwlZmLqmB.jpg</url>
            <title>Stories by Anitha Kannan on Medium</title>
            <link>https://medium.com/@anithakan?source=rss-891e9f27a4cf------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 21:39:01 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@anithakan/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[Beyond Vignettes: Why Primary Care AI Needs Conversational Benchmarks]]></title>
            <link>https://medium.com/curai-tech/beyond-vignettes-why-primary-care-ai-needs-conversational-benchmarks-014680d90ff6?source=rss-891e9f27a4cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/014680d90ff6</guid>
            <category><![CDATA[benchmarks-in-medicine]]></category>
            <category><![CDATA[healthai]]></category>
            <category><![CDATA[ai-safety]]></category>
            <dc:creator><![CDATA[Anitha Kannan]]></dc:creator>
            <pubDate>Thu, 07 May 2026 16:20:46 GMT</pubDate>
            <atom:updated>2026-05-07T23:05:00.461Z</atom:updated>
            <content:encoded><![CDATA[<p>Agustina Saenz, Danny Rosenthal, Anitha Kannan</p><p>Consider a conversational AI that takes a history from a patient presenting with chest tightness. It reasons fluently from the information it has, arrives at the correct diagnosis, and recommends appropriate management. It also never asks about exertional onset, never elicits radiation to the jaw, and never establishes whether the tightness resolves with rest. On every current benchmark, it passes. In a real encounter, it could miss a STEMI.</p><blockquote><strong><em>The failure is not one of reasoning, it is one of not asking.</em></strong></blockquote><p>This gap is the central problem in evaluating conversational primary care AI.​​</p><p>Clinical medicine has never relied on a single method of evaluation. Randomized trials prespecify populations, endpoints, and exclusions and demand transparent accounting of who was studied and who was not.[1] Objective structured clinical examinations (OSCEs) assess interactive clinical performance with standardized patients and structured rubrics rather than with written answers alone.[2] Clinical decision support systems (CDSS) are judged not only by correctness, but by timing and context.[3] And clinical quality measurement has long distinguished structure, process, and outcomes.[4] Primary care conversational AI should be evaluated with the same discipline.</p><p>That need is becoming urgent. Conversational systems that can ask follow-up questions, refine hypotheses, and recommend diagnoses and management are no longer speculative; early prospective pilots suggest that, in constrained primary care settings, they can approach clinician-level performance.[5,6] For these systems, prospective studies with real patients remain the gold standard, but pilots are expensive, slow, and feasible only for a narrow range of questions. The field therefore depends on benchmarks. The problem is that most current benchmarks are insufficient to evaluate readiness for a pilot.[7]</p><p>Recent proposals have argued, persuasively, that clinical LLM evaluation must move beyond static snapshots by embedding models in dynamic clinical environments where actions alter downstream patient and system states.[8,9] <strong><em>However, virtual primary care raises a complementary upstream evaluation problem: before a model can act safely, it must acquire the information required to justify action. In symptom-driven outpatient care, safety failures often arise not from resource allocation but from omission — failing to elicit key red flags or discriminating history — and from miscalibrated certainty when critical details are unknown. Benchmarks that do not measure information elicitation and “ask vs act” behavior therefore evaluate a different capability than the one deployed conversational systems require.</em></strong></p><h3>When the Benchmark Freezes the Information State</h3><p>Today’s dominant paradigms are the diagnostic vignette and the partial clinical conversation. Both are useful, but neither directly measures the capability deployed conversational systems require: safely acquiring the information needed to justify action. In both formats, models are largely evaluated on reasoning conditional on a benchmark-constructed information state, rather than on whether they can construct the state through questioning. Vignettes and partial transcripts can therefore reward correct endpoint answers while failing to reveal whether the model recognizes what is missing, asks the discriminating follow-ups and red flag questions in time, and avoids premature closure.</p><p>Neither paradigm reliably measures epistemic uncertainty about what has not yet been asked. In primary care, a safe conversational policy must be able to choose between asking a clarifying question and acting on current information, including appropriate deferral and safety-netting when critical details are missing. Benchmarks that force an immediate disposition from a fixed transcript cannot evaluate this choice and may overestimate real-world safety. In other words they under measure “ask vs act” behavior, central to outpatient safety when critical information is absent.</p><p>It is also a problem of external validity. Medicine has long recognized the danger of transporting evidence from population or purpose X to population or purpose Y. The AI analogue is benchmark-deployment mismatch: strong performance on a static benchmark does not establish fitness for use in a live, multi-turn primary care encounter. Emerging evaluations make this gap increasingly visible. Models that perform well when all clinically relevant facts are supplied upfront can degrade substantially when they must elicit those same facts through conversation.[10] For conversational primary care AI, that is not a peripheral issue; it is the core capability.</p><h3>Structure, Process, Outcomes — Not Outcomes Alone</h3><p>Donabedian’s framework helps clarify why.[4] Current AI benchmarks tend to overweight outcomes, under-specify structure, and under-measure process. <strong>Structure</strong> includes the benchmark cases, inclusion and exclusion criteria, prior chart context, simulator design, turn limits, and judging method. <strong>Process</strong> includes whether the model stays on the clinically relevant line of inquiry, elicits the necessary positives and negatives, recognizes what it does not yet know, escalates red flags within a reasonable number of turns, and calibrates its uncertainty appropriately. <strong>Outcomes</strong> include the final diagnosis, triage recommendation, and safety. In conversational AI, process is not ancillary; it is the mechanism by which safe outcomes are achieved.</p><p>This is also why static benchmarks can mislead. A model may arrive at the correct diagnosis while failing to gather the information that would justify confidence in that diagnosis. Another may ask the right questions, identify uncertainty, and escalate appropriately, even if the final label is imperfect. Those systems are not equivalent.</p><h3>Why OSCE-Style Evaluation Still Falls Short</h3><p>The closest clinical analogue is the OSCE.[2] It gets one thing right: we should evaluate what the system asks, not just what it answers.<br>But OSCE-style evaluations are still not enough. They depend on variable actors, are hard to reproduce, and miss the complexity of real primary care. What’s missing is infrastructure: standardized simulators, fixed cases, and scoring that captures the interaction itself.</p><p><strong><em>As these systems move toward greater autonomy, evaluation will need to more closely resemble how clinicians are assessed during residency, where performance is judged not only on knowledge, but on the ability to gather the right information, recognize uncertainty, and make safe decisions under incomplete data. Conversational benchmarks therefore function less like written examinations and more like supervised clinical encounters.</em></strong></p><p><strong><em>This implies a concrete benchmark requirement: evaluations should score the interaction trajectory, not just the endpoint correctness.</em></strong> That means attending to what the system chose to ask and what it failed to ask, whether it escalated red flags in time, whether it provided appropriate safety-netting, and whether its confidence was calibrated to the information actually elicited.</p><h3>Primary Care Conversational AI Needs its MIMIC</h3><p><strong><em>In inpatient AI, resources such as MIMIC transformed the field by making rich, temporally structured clinical data widely available.</em></strong>[11] MIMIC’s impact was not merely that it provided more examples; it changed what could be evaluated. By exposing time-stamped trajectories, notes, labs, vitals, treatments, it enabled questions about evolution, timing, and downstream consequences that snapshot datasets could not support.</p><p><strong><em>Primary care now needs an analogous, longitudinal resource for conversational systems: multi-turn encounters linked to downstream workup and outcomes, with explicit recording of what information (including chart context) was available at the time decisions were made.</em></strong></p><p>The rapid adoption of ambient documentation and other health AI tools makes this newly plausible.[12] For the first time, health systems are beginning to capture something closer to the encounter itself: what was asked, what was answered, what was volunteered, and what remained unresolved.</p><p><strong><em>Such datasets could support a more appropriate evaluation paradigm: standardized simulated conversations grounded in real encounter data and paired with judges that assess the trajectory of the interaction, not only the final answer.</em></strong> This would also address a major weakness of current conversational evaluation: simulator inconsistency. Different actors and patient simulators vary in how much information they volunteer, how literally they answer, and how difficult they are to interview, making benchmark results partly a function of the simulator rather than the model.[2,10] Grounding simulation in real transcript distributions offers a path toward standardization without stripping away realism.</p><p>The stakes will rise quickly. In the near future, these systems will be asked not only to triage acute symptoms but also to support longitudinal chronic disease management across interacting comorbidities. Managing diabetes, hypertension, chronic kidney disease, heart failure, depression, and polypharmacy over time is fundamentally different from answering a single vignette correctly. It is also precisely the setting in which regulators will demand stronger evidence that benchmark performance reflects intended use, risk, and safety.[13] As in randomized trials, benchmark reports should therefore clearly specify case selection, exclusions, available context, simulator behavior, and scoring design.[1]</p><h3>A Proposal: A Versioned Conversational Benchmark for Virtual Primary Care</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*zzcCDq-6fMlcKls_.png" /></figure><p><strong><em>We propose a versioned benchmark stack with three components: a fixed case set derived from real primary care encounter transcripts, a standardized patient simulator that produces consistent responses across evaluations, and a standardized judge that scores a taxonomy of safety-relevant conversational failures.</em></strong> Given the public health and regulatory implications, this benchmark should be developed and maintained through a federally guided public–private consortium to ensure broad representativeness, transparency, and long-term stewardship.</p><p><strong>Case set (real encounter transcripts). </strong>Ambient documentation workflows increasingly retain an encounter transcript (or near-verbatim conversational record) as an intermediate artifact used to generate the clinical note. These transcripts preserve what static benchmarks remove: the sequence of questions and answers, what patients volunteered, and what was never elicited. Each case can be packaged as an initial patient presentation plus explicit metadata on what clinical context was available at the time (for example, whether a medication list or problem list was accessible), linked to downstream signals such as diagnostic testing, follow-up diagnoses, reattendance, and escalation to urgent care or emergency services.</p><p><strong>Patient simulator (shared and versioned). </strong>To avoid variability introduced by human actors or unconstrained generative patient agents, the benchmark should ship with a “replay-first” simulator whose behavior is fixed and versioned. When a model asks a question that matches (semantically) one asked in the source encounter, the simulator returns the recorded answer. When a question cannot be answered from the transcript or from explicitly available chart context, it returns “unknown/not discussed/can’t recall,” rather than fabricating detail. Distributing the simulator with pinned versions (case IDs, turn limits, matching rules, and response policy) ensures that all evaluated models face the same patients under the same conditions.</p><p><strong>Judge (taxonomy-based, shared and versioned). </strong>Scoring should move beyond whether a final diagnosis label matches a reference. The judge should output a structured scorecard that includes: unsafe omissions (failure to elicit red-flag or discriminator information that could change triage or management), premature closure, under-triage and over-triage (graded by harm), hallucinated history or implied access to unavailable records, misinformation or unsafe advice, unnecessary or repetitive questioning, and excessive turns to escalate when high-risk features are present. Efficiency penalties should be applied only after safety-critical information needs are met, so benchmarks do not reward fast but unsafe behavior. As with any measurement instrument, the judge should be versioned and periodically audited on a fixed anchor set to support comparability over time.</p><p>A versioned, transcript-grounded benchmark also addresses a structural limitation of the current evaluation landscape: heterogeneity and fragmentation. Even within nominally single-task benchmarks, meaningfully different use cases are often pooled, for example, triage scenarios written for bystanders, patients and clinicians, and for adults and children — despite substantial differences in information needs, baseline risk, and appropriate thresholds for escalation. Across tasks, fragmentation is greater: triage, diagnosis and management are typically assessed on separate datasets with different inclusion criteria and labeling conventions, limiting the interpretability of cross-model comparisons. By contrast, an encounter-level primary care dataset that captures the full clinical episode, from initial presentation through workup, management and follow-up, can support multiple task formulations on the same underlying cases, while versioned simulators and scoring procedures preserve comparability across studies.</p><h3>Better Models Need Better Measurement</h3><p><strong><em>Primary care AI does not just need better models. It needs better measurement. </em></strong>The central risk in primary care is not misreasoning from available information, it is failing to obtain the information that reasoning requires. Current benchmarks cannot detect that failure, because they supply what the deployed system must elicit. Evaluating structure, process, and outcomes together is not a methodological preference; it is the minimum condition for a benchmark to be valid for the claim it is being asked to support. Until that standard is met, the field will know a great deal about how well these systems answer questions, and very little about whether they ask the right ones.</p><h3>References</h3><ol><li>Schulz KF, Altman DG, Moher D; CONSORT Group. CONSORT 2010 statement: updated guidelines for reporting parallel group randomised trials. BMJ. 2010;340:c332.</li><li>Harden RM, Gleeson FA. Assessment of clinical competence using an objective structured clinical examination (OSCE). Med Educ. 1979;13(1):41–54.</li><li>Kawamoto K, Houlihan CA, Balas EA, Lobach DF. Improving clinical practice using clinical decision support systems: a systematic review of trials to identify features critical to success. BMJ. 2005;330(7494):765.</li><li>Donabedian A. The quality of care. How can it be assessed? JAMA. 1988;260(12):1743–1748.</li><li>Saenz A, Schumacher E, Naik D, Khosla N, Kannan A. From concept to clinic: real-world evidence for autonomous AI deployment in primary care telemedicine. medRxiv [Preprint]. 2026 [cited 2026 May 1]. doi:10.64898/2026.03.18.26348749</li><li>Brodeur PG, Koshy T, Palepu A, et al. A prospective clinical feasibility study of a conversational diagnostic AI. arXiv [Preprint]. 2025 [cited 2026 May 1]. arXiv:2603.08448.</li><li>Rao AS, Esmail K, Lee RS, et al. LLM performance on clinical reasoning tasks. JAMA Netw Open. 2026;9(4):e264003.</li><li>Luo L, Kim SE, Zhang X, et al. A clinical environment simulator for dynamic AI evaluation. Nat Med. 2026;32:820–827.</li><li>Jiang Y, Black KC, Geng G, Park D, Zou J, Ng AY, et al. MedAgentBench: a virtual EHR environment to benchmark medical LLM agents. NEJM AI. 2025;2(9):AIdbp2500144.</li><li>Tu T, Azizi S, Dhingra N, et al. Towards conversational diagnostic AI. Nature. 2024;629:353–363.</li><li>Johnson AEW, Pollard TJ, Shen L, et al. MIMIC-III, a freely accessible critical care database. Sci Data. 2016;3:160035.</li><li>American Medical Association. 2 in 3 physicians are using health AI — up 78% from 2023 [Internet]. Chicago (IL): AMA; 2025 Feb 26 [cited 2026 Apr 14]. Available from:<a href="https://www.ama-assn.org/"> </a><a href="https://www.ama-assn.org/practice-management/digital-health/2-3-physicians-are-using-health-ai-78-2023">https://www.ama-assn.org/practice-management/digital-health/2-3-physicians-are-using-health-ai-78-2023</a></li><li>U.S. Food and Drug Administration. Clinical Decision Support Software: Guidance for Industry and Food and Drug Administration Staff [Internet]. Silver Spring (MD): FDA; 2022 [cited 2026 May 1]. Available from: <a href="https://www.fda.gov/">https://www.fda.gov/</a></li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=014680d90ff6" width="1" height="1" alt=""><hr><p><a href="https://medium.com/curai-tech/beyond-vignettes-why-primary-care-ai-needs-conversational-benchmarks-014680d90ff6">Beyond Vignettes: Why Primary Care AI Needs Conversational Benchmarks</a> was originally published in <a href="https://medium.com/curai-tech">Curai Health Tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2020 ML Research Internships at Curai]]></title>
            <link>https://medium.com/curai-tech/2020-ml-research-internships-at-curai-25d368dfcbb7?source=rss-891e9f27a4cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/25d368dfcbb7</guid>
            <category><![CDATA[startup]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[internships]]></category>
            <dc:creator><![CDATA[Anitha Kannan]]></dc:creator>
            <pubDate>Wed, 11 Aug 2021 16:28:04 GMT</pubDate>
            <atom:updated>2021-08-11T16:28:04.660Z</atom:updated>
            <content:encoded><![CDATA[<p>At Curai, <a href="https://medium.com/curai-tech/research-at-curai-fa09cae6727a">research</a> has been a top priority from the very beginning: we value longer term innovation, pushing the state of the art, and sharing advances with the community to validate our findings and to receive feedback. Our research internship program provides the opportunity for students to explore some of the research questions we are tackling towards scaling access to healthcare while gaining valuable research experience and, most of the time, coming out with a publication worthy piece of work. Our 2020 research internship program was special and likely to be the norm going forward — We hosted three research interns <em>remotely</em> <em>through the yea</em>r.</p><p>We had three creative research interns: <a href="https://scholar.google.com/citations?user=opUjZ0wAAAAJ&amp;hl=en">Anirudh Joshi</a> and <a href="https://www.linkedin.com/in/ali-mottaghi">Ali Mottaghi</a> from Stanford, and <a href="https://www.linkedin.com/in/bhachintagunta">Jai Chintagunta</a> from Virginia Tech. They made substantial research progress in areas of medical conversation summarization and active learning for long-tailed multilabel distributions:</p><ul><li>Anirudh introduced a <a href="https://arxiv.org/abs/2009.08666">medical conversation summarization model that encodes medical knowledge</a> presented at <a href="https://2020.emnlp.org/">EMNLP-2020</a>.</li><li>Jai studied the problem of <a href="https://aclanthology.org/2021.nlpmc-1.9/">imbibing medical knowledge into GPT-3</a> for use as a data simulator for medical dialog summarization. Jai’s work won the best paper award at the <a href="https://sites.google.com/view/nlp4mc2021/home">2021 NAACL Workshop on medical conversations</a> and will also be presented at <a href="https://www.mlforhc.org/">MLHC-2021</a>.</li><li>Ali studied the problem of <a href="https://arxiv.org/abs/2011.06874">long-tailed multilabel classification of patient text using active learning</a> approach/ He presented his work at <a href="https://ml4health.github.io/2020/">NeurIPS 2020 ML4H</a>.</li></ul><p>While you can read their work in the above links, what follows in this blog post is an edited short version of the same, in their own words. It has taken us a bit longer to put the blog together precisely because we have prioritized getting the work published in peer-reviewed publications.</p><p>If you are interested in NLP and clinical decision making research and want to positively impact quality healthcare access for all, please apply for a <a href="https://jobs.lever.co/curai/b15be159-f16a-49f9-bd05-95759dc2daf2">research internship</a> or any of our <a href="https://www.curai.com/work">other open positions</a>. You can also read about the amazing work of our engineering interns in this other <a href="https://medium.com/curai-tech/2020-21-internships-at-curai-69920f9ff686">blog post</a>.</p><p><strong>Confidentiality note</strong><em>: In accordance with Curai’s privacy policies, all the illustrative examples included in this post do NOT correspond to real patients. They are either synthetic or fully anonymized.</em></p><h3>Automated Summarization of Medical Dialogue</h3><p>Anirudh Joshi, Stanford University | Mentors: Namit Katariya, Anitha Kannan</p><p><em>You can find out more in our </em><a href="https://arxiv.org/abs/2009.08666"><em>paper</em></a><em> presented at </em><a href="https://2020.emnlp.org/"><em>EMNLP-2020</em></a><em>.</em></p><p>At Curai our mission is to provide affordable quality healthcare to everyone. To realize this mission, it is important to scale key aspects of the healthcare workflow to lower costs and increase access. Last Spring I had the opportunity to work with Anitha and Namit to leverage advances in natural language processing to help scale dialogue summarization.</p><p>Summarizing patient history is a core component of Curai doctors’ responsibility. This summary can be useful both to the patient to ensure accurate capture of all the information they conveyed, as well as to other Curai doctors to whom this conversation might be handed off to. This process takes 5–10 minutes per patient and reduces the amount of time health coaches can spend with patients. As Curai’s patient population grows, it is necessary for this task to also scale with the company in order to provide the best patient experience.</p><p>Our key insights while working on this problem were that when working with sparse medical data, it is critical to 1) design the data in a form that is conducive to machine learning and 2) design model architectures that possess the right inductive biases for the task.</p><p>Before we get into the science, let’s enjoy some of Dr. Summarize creations!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*juZ70rU62fC2iBgn" /><figcaption>Figure 1. Example of Dr. Summarize output on a conversation related to COVID-19.</figcaption></figure><p>Designing the data</p><p>From dialogue summarization literature, it is evident that some of the biggest challenges to accurate summarization is the ability to remember context over long dialogues. We tackled this problem by taking advantage of the local structure of our patient history taking data. Each “snippet” is the set of dialogue turns between a doctor and patient between subsequent questions from the doctor. We use snippets as a section of the dialogue to summarize and obtain ground truth summaries for these snippets from human doctors. Each snippet is independent and generating summaries for a snippet does not require other snippets.</p><p>While generating the summaries, we asked the doctors to include a special token [NO] if the medical condition mentioned was negated by the patient.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-vnDBXw_20-vUmWn" /><figcaption>Figure 2. Three training examples and their ground truth labels.</figcaption></figure><p>Designing the model</p><p>Dialogue summarization can be formulated as an abstractive task (model <em>generates</em> summaries), extractive task (model extracts exact phrases from the text) or a hybrid task (mix of abstractive and extractive). We formulate medical dialogue summarization as a hybrid task because it is important to retain the integrity of patient conversations through exact phrases, however for fluency it is important to allow the model to generate text. Given this formulation, we chose to extend the Pointer Generator model architecture that has proven to be state of the art for hybrid summarization tasks.</p><p>The Pointer Generator model produces probability distributions over the original text (copy-mode) and over the generative decoder (generate-mode). This allows it to determine whether to copy from the input or generate a word token for a given location in the summary. While this inductive bias is perfect for our task, there are important limitations of the vanilla model (2M-BASE) that are not well suited for medical dialogue. 1) The pointer generator network biases to generation during inference. 2) Since it is pre-trained on news corpora, it biases towards sequential text copying as opposed to alternating between generation and copying. For our task, we want the model to bias to copying to maintain the integrity of the patient response and we want the model to alternate between copying and generation.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*VU0OkRKcEME80_k3" /></figure><p>In order to build those properties into the pointer generator network we penalize the probability of generation in our loss function so the model learns to rely more on copying. We also propose to encourage the model to focus on medical concepts and negations more than ordinary words. We do this through a combination of influencing the attention distribution during copying and also through the loss function.</p><p>We find that our extension to the pointer generator network improves summaries both on automated metrics like ROGUE-F1 but also through evaluations from human doctors. Two doctors evaluated the outputs of the vanilla baseline model compared to extension (Dr. Summarize) and the results showed that they preferred Dr. Summarize on twice the number of examples compared to the baseline. For a more detailed analysis of the results, check out our publication in EMNLP Findings linked at the top!</p><p>Dr. Summarize illustrates that automated medical dialogue summarization can be achieved with current advances in NLP and can be translated to help scale healthcare.</p><p>I had an incredible time as an ML researcher at Curai as it has the perfect balance between research focus and tying the research back to improving patient experience in the product. The promise of machine learning in healthcare has always been about scaling and industrialising parts of the workflow and Curai’s product and research are directly aligned with that. A huge thank you to the health coaches who enable high quality AI work to happen through data creation and evaluation. An extra special thanks to Anitha who has been the best machine learning research mentor I’ve had!</p><h3>Medical symptom recognition from patient text: An active learning approach for long-tailed multilabel distributions</h3><p>Ali Mottaghi, Stanford | Mentor: Anitha Kannan</p><p><em>You can find out more in our </em><a href="https://arxiv.org/abs/2011.06874"><em>paper</em></a><em> presented at </em><a href="https://ml4health.github.io/2020/"><em>NeurIPS 2020 ML4H</em></a><em>.</em></p><p>In alignment with the standards for medical history taking, Curai asks its patients for a concise statement describing the reason for their visit: the reason for encounter (RFE), sometimes also called <a href="https://en.wikipedia.org/wiki/Chief_complaint">chief complaint (CC)</a>. Extracting medical findings (symptoms) from the RFE is an important step in history taking and asking follow-up questions. In my internship at Curai, I worked on developing a more data-efficient machine learning model for this task. The following example shows an example of the extracted medical symptoms from RFEs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1012/0*yT9x6NuZDoiTFGtU" /><figcaption>Figure 3: An example RFE and corresponding set of medical symptoms</figcaption></figure><p>There is no large and publicly available dataset for the task of medical symptom recognition from <em>patient text.</em> Therefore, at Curai we curate our own dataset. However, labeling RFEs is time-consuming and expensive. In this internship, we developed a new method that achieves the highest performance and coverage in medical symptom recognition while using the fewest number of labeled examples.</p><p>As shown in Figure 4, RFEs and their extracted symptoms have a long-tailed distribution. Some symptoms are common while multiple symptoms tend to co-occur. In order to model this kind of data, we developed a new active learning method that iteratively selects a set of unlabeled RFEs that covers a wide range of medical symptoms. Therefore, by labeling them, we can achieve the highest accuracy using the least number of labels.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/804/0*EvutATGM4SDClmqd" /><figcaption>Figure 4: Long-tailed data distribution. Also, shown example RFEs from dataset: Note the patient language and co-occurrence of symptoms.</figcaption></figure><p>Our active learning procedure starts with a large pool of unlabeled RFEs, available in encounters on Curai Health Platform. At the beginning we only had a small fraction of RFEs labeled. Our goal is to iteratively select and label the most informative samples. As shown in the Figure 5 below, in each iteration, we train our neural network model on the current labeled set and extract the features (latent representations) from labeled and unlabeled RFEs.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/0*_RPrcOAgq0VvZlCH" /><figcaption>Figure 5: Overview of our active learning approach</figcaption></figure><p>Then the Affinity propagation algorithm [ref: Frey et al. 2007] is utilized to reveal the underlying structure of the latent space. This algorithm is an exemplar clustering method based on probability propagation and it does not require the number of clusters to be specified beforehand. Our algorithm, Active Long-Tailed Learning (ALTL), selects a diverse set of representations based on their distance to the cluster centroids and already labeled data points. We want to select data points that explore both new regions of the latent space (and hence introduces previously unseen labels), as well as those that are close to the cluster centroids (captures data density in the latent space). Selected data points are labeled by an expert and are added to the labeled set.</p><p>To evaluate our model, we gathered 1232 RFEs from Curai Health Platform and had medical experts label them with corresponding symptoms. We consider a universe of 20 most frequently occurring medical symptoms in patient RFEs. Figure 4 shows the distribution of the data. Due to the size of available datasets, we use a pretrained InferSent [ref: Conneau et al. 2017] for encoding patient RFE text. Text embeddings are fed into a MLP and the outputs before the last fully connected layer are used as the main features in our method.</p><p>We set up our experiments to first start with a small set of 10 labeled RFEs, and then at each iteration, we select a new batch (again of size 10) of unlabeled RFEs, with ALTL acquisition function. The new data points are then labeled by the oracle and added to the labeled training dataset. The model is then retrained on this new training dataset. This process is continued for multiple iterations until the labeling budget is exhausted. We compare the performance of our method with the following baselines:</p><ol><li>Random baseline, as the name suggests, randomly selects new data points from the unlabeled data pool following a uniform distribution</li><li>Fully supervised baseline is a fully supervised model trained on a completely labeled training data set and serves as a reference</li><li>Core-set [ref: Sener et. al. 2017] is a representation- based method that solves the K-Center problem to select the points in the latent space</li><li>Max-entropy [ref: Settles et. al. 2009] is the best-known uncertainty-based method in Active Learning that uses the entropy of the model’s output as a measure for uncertainty.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/936/0*zoTFeQX69fYeTvU2" /><figcaption>Figure 6: Performance (LRAP and F1-Score) of our method compared to Core-set, Maxentropy, Random, and fully-supervised baselines.</figcaption></figure><p>Figure 6 shows the performance of our model compared to the baselines. We can see that our approach outperforms the baselines. Since we start with only 10 labeled RFEs, many labels in the universe of symptoms do not have associated data samples. Our method is able to identify the previously unseen and unlabeled data points for exploration. Simultaneously, since the data has a long-tailed distribution, our methods are not susceptible to oversampling outliers along the boundaries of the data.</p><p>We also qualitatively analyzed the performance of our model based on the selected points in the latent space. <a href="https://arxiv.org/pdf/2011.06874.pdf">Figure 4 in the paper</a> shows that at first, our model explores new clusters (labels) and later it balances the exploration and exploitation to avoid outliers in the data. As we showed in our experiments, this can address long-tailed distribution in the dataset and outperform compared baselines in the symptom recognition task.</p><h3>Medically aware GPT-3 as data generator for medical conversation summarization</h3><p>Jai Chintagunta, Virginia Tech | Mentor: Namit Katariya</p><p><em>You can find out more in our </em><a href="https://aclanthology.org/2021.nlpmc-1.9/"><em>paper</em></a><em> that won the best paper award in </em><a href="https://www.aclweb.org/portal/content/second-workshop-nlp-medical-conversations"><em>2020 ACL Workshop on medical conversations</em></a><em> and also got accepted as a full paper at </em><a href="https://www.mlforhc.org/"><em>MLHC-2021</em></a><em>.</em></p><p>Hi I’m Jai. Prior to Curai, I studied mathematics at Virginia Tech. I joined Curai’s machine learning team specifically working on the problem of medical dialogue summarization.</p><p>As Anirudh described previously, summarizing patient history is a crucial task in the workflow of Curai providers. With Dr. Summarize work from Anirudh’s internship, we had a foundation for approaching medical dialogue summarization. However, we lacked a large labeled dataset, which is crucial for summarization tasks. In the medical world, obtaining labeled examples is both costly and challenging. At Curai, we’ve started on the journey of generating high-quality labeled summarization examples, obtaining a set of 6,000 labeled examples vetted by clinical professionals. Although this is a good start, it is nowhere near enough labeled data to train a summarization model, especially in the medical domain where errors are deleterious. To put this into perspective, the main summarization datasets (CNN/Daily Mail, XSum) used in research have upward of 200k labeled examples. With such few labels we hypothesized that a large scale language model which had been trained on a vast amount of data could offer us good few shot performance. GPT-3 certainly fits the bill and we tried it out on our task, priming it with 21 labeled examples. GPT-3 off-the-shelf (referred to as Vanilla GPT-3 from now on) produces coherent summaries; however, it does not do so well at retaining relevant medical information from the conversation. For example in Conversation 1 in Figure 7, Vanilla GPT-3 fails to capture the name of the medication and dosage information.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*CP7HYIbJyaHiJO4D" /><figcaption>Figure 7: Qualitative comparison of summaries outputted by Vanilla GPT-3 and GPT-3-ENS</figcaption></figure><p>We realized that we need to incorporate medical domain knowledge into Vanilla GPT-3. Our approach, GPT-3-ENS, illustrated in the first box of Figure 8 relies on ensembling. We have 10 different instances of Vanilla GPT-3 for which we provide 21 unique training examples each (total of 210 human annotated examples) and have each produce a summary for a total of 10 summaries. Out of these 10 summaries we pick the summary which has the most medical concepts in common with the original dialogue using the help of a medical entity recognizer. This approach yields more medically correct summaries as opposed to Vanilla GPT-3. We can see from Conversation 1 in Figure 7 that GPT-3-ENS now captures the relevant dosage and medication information that was missed by Vanilla GPT-3.</p><p>Now we have a way of producing medically correct summaries using relatively few manual annotations. However we can’t send every doctor-patient chat to OpenAI servers due to privacy concerns and furthermore, the approach is not very amenable to corrections/supplementations from doctors. Thus, we propose to use GPT-3-ENS as a way to synthesize labels upon which we train a downstream summarization model (Figure 8).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/799/0*D8q0-d0QB2fAk3tR" /><figcaption>Figure 8: First box introduces GPT-3-ENS: infusing medical knowledge to identify best medically correct summary from GPT-3. GPT-3-ENS synthesized and human labeled data are combined to learn summarization model that performs better than models trained on either of the sources</figcaption></figure><p>Using this approach, we are able to train a model from 210 human labeled examples to be on par in performance with a model trained from 6400 human labeled examples. Furthermore, in our experiments we discovered that the best model comes from training on a mix of GPT-3-ENS synthesized examples and human-labeled examples.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/800/0*Ng1EA-hMJdoBeSY7" /><figcaption>Figure 9: Doctor evaluations in regards to coherency and coverage of generated summaries</figcaption></figure><p>In Figure 9:</p><ul><li>Human is a model trained from 6400 human labeled examples</li><li>GCF… is a model trained from 6400 GPT-3-ENS synthesized examples (which it itself needed 210 human labeled examples)</li><li>Human + GCF … is a model trained from 3200 GPT-3-ENS synthesized examples and 3200 human labeled examples</li></ul><p>You can find out more in our <a href="https://aclanthology.org/2021.nlpmc-1.9/">paper</a> that won the best paper award in <a href="https://www.aclweb.org/portal/content/second-workshop-nlp-medical-conversations">2020 ACL Workshop on medical conversations</a> and also got accepted as a full paper at <a href="https://www.mlforhc.org/">MLHC-2021</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=25d368dfcbb7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/curai-tech/2020-ml-research-internships-at-curai-25d368dfcbb7">2020 ML Research Internships at Curai</a> was originally published in <a href="https://medium.com/curai-tech">Curai Health Tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Research at Curai]]></title>
            <link>https://medium.com/curai-tech/research-at-curai-fa09cae6727a?source=rss-891e9f27a4cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/fa09cae6727a</guid>
            <category><![CDATA[medical-diagnosis]]></category>
            <category><![CDATA[medical-summarization]]></category>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[research-at-curai]]></category>
            <dc:creator><![CDATA[Anitha Kannan]]></dc:creator>
            <pubDate>Thu, 01 Oct 2020 18:04:31 GMT</pubDate>
            <atom:updated>2020-10-02T03:50:41.754Z</atom:updated>
            <content:encoded><![CDATA[<p><em>A quick overview of our research publications is available </em><a href="https://medium.com/curai-tech/research-publications-at-curai-ee22a820b807?sk=3deb6e26508df699e63c72c0298a069e"><em>here</em></a><em>.</em></p><p>We started our company to fundamentally reimagine what it means to truly provide the best healthcare to every human being. This requires transformative solutions that scale medical providers’ abilities as well as lower users’ barrier to entry to care. A natural component to the solution is that of tight seamless integration of AI/ML in the doctor-patient interaction flow that augments and scales medical providers, and not merely as a plug-and-play in an existing workflow. To truly achieve this, there are several open research problems at the intersection of machine learning and healthcare that need to be addressed. That’s why research has been one of the top priorities in the company, from the very beginning.</p><p>We also acknowledge that not all research problems can be solved by ourselves: We seek and actively collaborate with the research community through external collaborations and through our research internship program. Our recent external collaborations include partnership with Stanford, MIT, Georgia Tech, Massachusetts General Hospital and Vanderbilt University. We actively publish in top tier conferences to give back to the community and also to have our research vetted.</p><p>In this blog post, we describe some of our research work specifically in the areas of clinical decision making and natural language processing, which have been our two main areas of focus because of their connection to healthcare in general and to <a href="https://curaihealth.com/">Curai Health</a>, in particular.</p><h3>Clinical decision making</h3><p>In a perfect world, clinical decision making is a complete information game in which the medical provider has all the information about the patient and the entirety of this information is used to draw conclusions regarding diagnosis and care plan. As we described in an earlier <a href="https://medium.com/curai-tech/the-science-of-assisting-medical-diagnosis-from-expert-systems-to-machine-learned-models-cc2ef0b03098">blog post</a>, many factors prevent this from happening in the real world. Our hypothesis is that AI diagnostic systems will complement medical providers to solve the complete information game by gathering all the pertinent information from the patients, while at the same time suggesting diagnostic possibilities and treatment plans [<a href="https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2726709">1</a>].</p><p>For clinical AI/ML solutions to be truly effective, these systems need to be deployable “in-the-wild”. By this we mean that AI/ML systems need to be correct in making the decisions in environments and medical situations that they may have never encountered. If the system is extremely good at thousand conditions; when it’s deployed in the wild, it may need to attend to a patient with that thousand-and-one~th condition that it has never seen before.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ekdrABNxtxj5UqBD" /></figure><p>While it’s a tall goal to strive for AI systems to perform in ‘in-the-wild’, we can make progress towards this goal by having medical provider-in-the-loop. For this to happen, we need these systems to have two properties. First, AI/ML systems need to be <em>extensible</em>: which means that they automate parts of the workflow in the interaction between the medical provider and patient, and incrementally update itself as more data becomes available or by learning from the provider-in-the loop. This would mean the models become more knowledgeable over time as they interface more. Because of the need to selectively automate, these AI models also <em>need to know what it knows and what it doesn’t</em> so that the system can switch between either automating the task or graciously falling back to the providers. This also builds trust for the providers.</p><h3>Medical diagnosis</h3><p>Diagnosis is <em>“a mapping from a patient’s data (normal and abnormal history, physical examination, and laboratory data) to a nosology of disease states”, [</em><a href="https://link.springer.com/chapter/10.1007/978-3-319-31913-1_11"><em>3</em></a><em>]</em></p><h4>Extensible models for diagnosis</h4><p>Since the 1970s, there has been significant efforts in building AI driven clinical decision support. The early approaches depended extensively on expert systems where medical experts curated knowledge bases and inference algorithms, based on medical studies. These medical studies are costly to run and may be limited in their size and their ability to generalize to the population as a whole, rapidly. Machine learned models for diagnosis hold a lot of promise in terms of scalability and extensibility by leveraging data stored in clinical electronic health records, but do not offer an immediate way to encode prior medical knowledge.</p><p><em>Our earliest research asked the following question: “How do we design extensible machine learned models, from scratch? How can we incorporate medical knowledge encoded in over fifty years of manually curated AI expert systems for medical diagnosis?”</em></p><p>We tackled these questions by treating expert systems as data prior: We used expert systems as a clinical case simulator, and these cases are used as training data for learning a model for diagnosis.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*nVlmjfzctSPGneH5" /><figcaption>Machine learned (ML) model for differential diagnosis (DDx) by combining simulated data from expert systems and COVID-19 assessments data. The resulting model is COVID-aware and can reason about multiple hypothesis based on the patient symptoms.</figcaption></figure><blockquote><a href="https://arxiv.org/abs/1804.08033"><strong>Learning from the experts: From expert systems to machine learned diagnosis models</strong></a><br>M. Ravuri, A. Kannan, G. Tso, X. Amatriain<br>Machine Learning for Healthcare (MLHC), 2018</blockquote><p>With the COVID-19 pandemic, we showed how we can quickly update these models to take into account our evolving knowledge about COVID-19 by leveraging data from COVID-19 symptom checker updated frequently by practitioners. We show that our approach is able to accurately model new incoming data about COVID-19 while still preserving accuracy on conditions that have been modeled in the past.</p><blockquote><a href="https://arxiv.org/abs/2008.03323"><strong>COVID-19 in differential diagnosis of online symptom assessments</strong><br></a>A Kannan, R Chen, V Venkataraman, G Tso, X Amatriain</blockquote><p>One of the biggest challenges in learning a comprehensive diagnosis model for all diseases is lack of access to sufficient data because of HIPAA regulations or because of the rarity (which is good!) of diseases or their phenotypes. According to International classification of diseases (ICD), there are thousands of diseases, while expert systems cover a few hundred diseases.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*-IrtVUePsmia8Ltm" /><figcaption>Long-tailed class distribution Also shown are nearest neighbors to four of the many prototypes learned for select classes using the proposed Prototypical Clustering Network approach. This is illustrative of the huge intra-class variability in the data. For a novel test image, shown at the upper right corner, the model predicts the correct class by measuring weighted similarity to per-class clusters in the embedding space.</figcaption></figure><p>We tackled this problem by explicitly posing medical diagnosis as a low-shot learning problem that can handle the long-tail in the data and perform well on classes in both the head and the tail of distribution. We extended prototypical networks to learn local clusters for each disease to handle multiple phenotypes. Once deployed, the model is also easily extensible to new disease classes by leveraging a few clinical cases, labeled by a practitioner on the platform.</p><blockquote><a href="https://arxiv.org/abs/1811.03066"><strong>Prototypical Clustering Networks for Dermatological Disease Diagnosis</strong></a><br>V Prabhu, A Kannan, M Ravuri, M Chablani, D Sontag, X Amatriain<br>NeurIPS Machine Learning for Health (ML4H), 2018<br>Machine Learning for Healthcare (MLHC), 2019</blockquote><h4>Knowing when to fall-back to providers</h4><p>In a recent work published at AMIA, we introduced the problem of trade-off between the number of diseases and the accuracy of diagnosis models in user-facing settings. We show that modeling the diseases of interest (eg. commonly seen on a platform), provides much better accuracy than modeling all diseases.</p><blockquote><a href="https://arxiv.org/pdf/1912.08041.pdf"><strong>The accuracy vs. coverage trade-off in patient-facing diagnosis models</strong><br></a>A Kannan, J Fries, E Kramer, J Chen, N Shah, X Amatriain<br>AMIA Informatics Summit, 2020</blockquote><p>In the subsequent work, we posed the problem of open set medical diagnosis in which the proposed model explicitly models out-of-distribution. When the model predicts a clinical case as belonging to out-of-distribution, it’s indicative of the model to completely hand-off to the provider without making any suggestion. We also show how to model open-set diagnostic ensembles when training data is distributed across several healthcare sites that do not allow data pooling.</p><blockquote><a href="https://arxiv.org/abs/1910.02830"><strong>Open Set Medical Diagnosis</strong><br></a>V Prabhu, A Kannan, G Tso, N Katariya, M Chablani, D Sontag, X Amatriain<br>NeurIPS Machine Learning for Health (ML4H), 2019<br>ACM Conference on Health, Inference, and Learning (ACM- CHIL), 2020</blockquote><h3>Natural language processing</h3><p>At <a href="https://curaihealth.com/">Curai Health</a>, the primary mode of interaction with the patients/users is text based chat. For patients, this fits into their busy lives as it is affordable and accessible at the moment. For the practitioners on the platform, chat interface improves ability to scale operationally since it can combine synchronous and asynchronous modalities, at the appropriate level of prioritization. Therefore, as we describe in our previous<a href="https://medium.com/curai-tech/nlp-healthcare-understanding-the-language-of-medicine-e9917bbf49e7"> blog post</a>, advancements in natural language understanding is important to our work.</p><p>There has been tremendous advancements in NLP (eg, GPT-3, BERT) in recent years. However, there are number of important challenges to be tackled, when we apply NLP to healthcare. This includes imparting medical knowledge into language and overcoming lack of task-specific datasets. We will now describe how we are tackling these challenges.</p><h4>Medical question answering</h4><p>The rate at which medical questions are asked online far exceeds the capacity of practitioners to answer them. One way to address this problem is to build a retrieval system that can automatically match unanswered questions to answered questions with the same meaning, or mark them as priority for a doctor if no similar answered questions exist. Building such a retrieval system has two main challenges: First is the dataset that encodes nuances in medical knowledge (eg. Can I take aleve during pregnancy? Vs. Can I take aleve after pregnancy? ). Second, we recognize that in healthcare, a simple answer might not always solve the concern (such as the example above) and therefore, the machine learned models for question answering should be able to direct the patient to the providers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*hb-ytv1aE6q9X71P" /><figcaption>Examples of questions from the users that are matched to questions in our FAQ. Notice how we also suggest discussing their health issue with the providers for certain user questions.</figcaption></figure><p>To solve the first problem, in a recent work presented at NeurIPS <a href="https://ml4health.github.io/">Machine Learning for Health</a>, 2019, we augment the features of language models using the depth of information that is stored within a medical question-answer pair dataset to provide the needed medical knowledge. We show how pre-training a transformer network on a medical question-answer matching task embeds relevant medical knowledge into the model that it otherwise does not have. We then fine-tune on an in-house (<a href="https://github.com/curai/medical-question-pair-dataset">made publicly available</a>) dataset of question-question similarity.</p><blockquote><a href="https://arxiv.org/abs/2008.13546"><strong>Effective Transfer Learning for Identifying Similar Questions: Matching User Questions to COVID-19 FAQs</strong></a><br>C McCreery, N Katariya, A Kannan, M Chablani, X Amatriain<br>KDD, 2020</blockquote><p>During this COVID, there is a huge surge in the number of questions from users. We extended the above work to be used in the context of COVID related curated FAQ dataset. We enabled an option on our platform to directly connect the patients with our providers, when needed.</p><h4>Learning medical conversational agent</h4><p>Another area of research we actively investigate is in learning conversational agents for interaction with patients. The task of learning medical dialog agents is hard for many reasons. First, the definition of “task” is subtle: the task of medical diagnosis encompasses the task of identifying the correct diagnosis (over a few thousands ) for a patient. Second, the models need to be extremely precise to uncover all the pertinent information from the patient, by jointly doing medical reasoning while gathering information with the patients.</p><p>We took an initial step in this direction by formulating the problem as a conversational assistant to a practitioner wherein the model would suggest responses to the practitioner based on the context of the conversation so far. We formulated this response generation as a discrimination task where the goal is to rank similar responses from response clusters. We show that our approach of clustering contexts and identifying responses describing the context had much better accuracy than learning a full generative model for response generation.</p><blockquote><a href="https://arxiv.org/abs/1910.03476"><strong>Classification As Decoder: Trading Flexibility For Control In Neural Dialogue</strong><br></a>S Shleifer, M Chablani, N Katariya, A Kannan, X Amatriain<br>NeurIPS Machine Learning for Health (ML4H), 2019</blockquote><h4>Medical Dialogue summarization</h4><p>Another important research area is around medical dialogue summarization. Medical dialogue summaries enable effective hand-off of the patient between the practitioners so that they can be used for medical decision making and subsequent follow-ups. For the patient, it also serves as a detailed summary of their conversation with their provider. Understanding a medical conversation between a patient and a provider poses unique natural language understanding challenges since it combines elements of standard open-ended conversation with very domain specific elements that require expertise and medical knowledge.</p><p>In a recent work, we extend the pointer generator network by incorporating important properties of medical conversations such as medical knowledge coming from standardized medical ontologies , and also taking into account the affirmations of the patient symptoms</p><blockquote><a href="https://arxiv.org/abs/2009.08666"><strong>Dr. Summarize: Global Summarization of Medical Dialogue by Exploiting Local Structures</strong></a><br>A Joshi, N Katariya, X Amatrian, A Kannan<br>EMNLP — Findings, 2020</blockquote><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SoqOAieoaoxWLago" /><figcaption>Example of a medical dialogue and summary generated by our model. Note that the summary captures affirmatives, negatives and medical concepts.</figcaption></figure><p>We are working in a time with tremendous opportunity as well as obligation to scale access to care. We envision that technology and AI will play a huge role in this. At the same time, AI systems can not be simply dropped in the existing workflow — they need to be integrated into end-to-end medical care benefiting both patients and providers. At Curai, we are focused on exactly this. Please reach out if this area of research interests you, and interested in being part of our mission.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=fa09cae6727a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/curai-tech/research-at-curai-fa09cae6727a">Research at Curai</a> was originally published in <a href="https://medium.com/curai-tech">Curai Health Tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Research Publications at Curai]]></title>
            <link>https://medium.com/curai-tech/research-publications-at-curai-ee22a820b807?source=rss-891e9f27a4cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/ee22a820b807</guid>
            <category><![CDATA[nlp]]></category>
            <category><![CDATA[ai-expert-systems]]></category>
            <category><![CDATA[machine-learning-research]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[ehr]]></category>
            <dc:creator><![CDATA[Anitha Kannan]]></dc:creator>
            <pubDate>Thu, 01 Oct 2020 18:03:03 GMT</pubDate>
            <atom:updated>2026-05-08T21:34:36.844Z</atom:updated>
            <content:encoded><![CDATA[<p><em>Last updated: May 8, 2026</em></p><h3><strong><em>Curai’s commitment to AI is to impact primary healthcare and advance research/science meaningfully</em></strong></h3><p>Since Curai was founded in 2017, we have published 25 peer-reviewed papers and received over 740 citations, including citations from other leading healthcare organizations. This commitment to transparency in AI research and the pursuit of academic, peer-reviewed publications is central to our mission.</p><p><a href="https://www.medrxiv.org/content/10.64898/2026.03.18.26348749v1"><strong>From Concept to Clinic: Real World Evidence for Autonomous AI Deployment in Primary Care Telemedicine</strong><br></a>A Saenz, E Schumacher, D Naik, N Khosla, A Kannan<br>MedRxiv, 2026</p><p><a href="https://medium.com/curai-tech/beyond-vignettes-why-primary-care-ai-needs-conversational-benchmarks-014680d90ff6"><strong>Beyond Vignettes: Why Primary Care AI Needs Conversational Benchmarks</strong></a><br>A Saenz, D Rosenthal, A Kannan<br>Medium Blog Post, 2026 [unpublished work]</p><p><a href="https://arxiv.org/abs/2502.15069"><strong>Rare Disease Differential Diagnosis with Large Language Models at Scale</strong><br></a>E Schumacher, D Naik, A Kannan<br>Machine Learning for Healthcare, 2025</p><p><a href="https://arxiv.org/abs/2311.08303"><strong>Extrinsically-Focused Evaluation of Omissions in Medical Summarization</strong></a><strong> </strong><br>E Schumacher, D Rosenthal, V Nair, L Price, G Tso, A Kannan<br>Machine Learning for Health Symposium, 2024, GenAI for Health: Potential, Trust and Policy Compliance, 2024<br>Arxiv, 2023</p><p><a href="https://arxiv.org/abs/2304.14364"><strong>CONSCENDI: A Contrastive and Scenario-Guided Distillation Approach to Guardrail Models for Virtual Assistants</strong></a><br>A Sun, V Nair, E Schumacher, A Kannan<br>NAACL, 2024</p><p><a href="https://arxiv.org/abs/2303.17071"><strong>DERA: Enhancing Large Language Model Completions with Dialog-Enabled Resolving Agents</strong></a><br>V Nair, E Schumacher, G Tso, A Kannan<br>NAACL, Clinical NLP Workshop 2024</p><p><a href="https://arxiv.org/abs/2306.03652"><strong>Injecting knowledge into language generation — A case study in auto-charting after-visit care instructions from medical dialogue </strong><br></a>M Ereemev, I Valmianski, X Amatrian, A Kannan<br>ACL, 2023</p><p><a href="https://arxiv.org/abs/2305.05982"><strong>Generating medically-accurate summaries of patient-provider dialogue: A multi-stage approach using large language models</strong></a><br>V Nair, E Schumacher, A Kannan<br>ACL Clinical NLP Workshop, 2023</p><p><a href="https://arxiv.org/abs/2304.01974"><strong>Dialog-Contextualized Re-ranking for Medical History Taking </strong><br></a>J Zhu, I Valmianski, A Kannan<br>Machine Learning for Healthcare, 2023</p><p><a href="https://arxiv.org/abs/2210.02658"><strong>Learning functional sections in medical conversations: iterative pseudo-labeling and human-in-the-loop approach</strong></a><br>M Wang, I Valmianski, X Amatrian, A Kannan<br>Machine Learning for Healthcare, 2023</p><p><a href="https://arxiv.org/abs/2207.05817"><strong>OSLAT: Open Set Label Attention Transformer for medical entity span extraction</strong></a><br>R Li, I Valmianski, L Deng, X Amatrian, A Kannan<br>Machine Learning for Health (ML4H), 2022</p><p><a href="https://arxiv.org/abs/2111.09381"><strong>MEDCOD: A medically-accurate, Emotive, Diverse, and Controllable Dialog System</strong></a><strong> </strong><br>R Compton, I Valmianski, L Deng, C Huang, N Katariya, X Amatrian, A Kannan<br>Machine Learning for Health (ML4H), 2021</p><p><a href="https://arxiv.org/abs/2111.07564"><strong>Adding more data does not always help: A study in medical conversation summarization with PEGASUS</strong></a><br>V Nair, N Katariya, I Valmianski, X Amatrian, A Kannan<br>Machine Learning for Health (ML4H), 2021</p><p><a href="https://arxiv.org/abs/2110.07356"><strong>Medically-aware GPT-3 as a data generator for medical dialog summarization</strong></a><strong><br></strong>B Chintagunta, N Katariya, X Amatrian, A Kannan<br>(Best Paper award at) NAACL-Workshop on NLP for Medical Conversations, 2021 and Machine Learning for Healthcare (MLHC) 2021</p><p><a href="https://arxiv.org/abs/2011.06874"><strong>Medical symptom recognition from patient text: An active learning approach for long-tailed multilabel distributions</strong></a><br>A. Mottaghi, P. Sarma, X. Amatriain, S. Yeung, A. Kannan<br>NeurIPS — Machine Learning for HealthCare, 2020</p><p><a href="https://arxiv.org/abs/2009.08666"><strong>Dr. Summarize: Global Summarization of Medical Dialogue by Exploiting Local Structures</strong></a><br>A Joshi, N Katariya, X Amatrian, A Kannan<br>EMNLP — Findings, 2020</p><p><a href="https://arxiv.org/abs/2008.03323"><strong>COVID-19 in differential diagnosis of online symptom assessments</strong></a><br>A Kannan, R Chen, V Venkataraman, G Tso, X Amatriain<br>NeurIPS —ML4H, 2020</p><p><a href="https://arxiv.org/abs/2008.13546"><strong>Effective Transfer Learning for Identifying Similar Questions: Matching User Questions to COVID-19 FAQs</strong></a><br>C McCreery, N Katariya, A Kannan, M Chablani, X Amatriain<br>KDD, 2020</p><p><a href="https://arxiv.org/abs/1910.02830"><strong>Open Set Medical Diagnosis</strong><br></a>V Prabhu, A Kannan, G Tso, N Katariya, M Chablani, D Sontag, X Amatriain<br>NeurIPS Machine Learning for Health (ML4H), 2019<br>ACM Conference on Health, Inference, and Learning (ACM- CHIL), 2020</p><p><a href="https://arxiv.org/pdf/1912.08041.pdf"><strong>The accuracy vs. coverage trade-off in patient-facing diagnosis models</strong><br></a>A Kannan, J Fries, E Kramer, J Chen, N Shah, X Amatriain<br>AMIA Informatics Summit, 2020</p><p><a href="https://arxiv.org/abs/1910.03476"><strong>Classification As Decoder: Trading Flexibility For Control In Neural Dialogue</strong><br></a>S Shleifer, M Chablani, N Katariya, A Kannan, X Amatriain<br>NeurIPS Machine Learning for Health (ML4H), 2019</p><p><a href="https://arxiv.org/abs/1910.04192"><strong>Domain-Relevant Embeddings for Medical Question Similarity</strong><br></a>C McCreery, N Katariya, A Kannan, M Chablani, X Amatriain<br>NeurIPS Machine Learning for Health (ML4H), 2019<br><br><a href="https://arxiv.org/abs/1811.03066"><strong>Prototypical Clustering Networks for Dermatological Disease Diagnosis</strong></a><br>V Prabhu, A Kannan, M Ravuri, M Chablani, D Sontag, X Amatriain<br>NeurIPS Machine Learning for Health (ML4H), 2018<br>Machine Learning for Healthcare (MLHC), 2019</p><p><a href="https://arxiv.org/abs/1804.08033"><strong>Learning from the experts: From expert systems to machine learned diagnosis models</strong></a><br>M. Ravuri, A. Kannan, G. Tso, X. Amatriain<br>Machine Learning for Healthcare (MLHC), 2018</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ee22a820b807" width="1" height="1" alt=""><hr><p><a href="https://medium.com/curai-tech/research-publications-at-curai-ee22a820b807">Research Publications at Curai</a> was originally published in <a href="https://medium.com/curai-tech">Curai Health Tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[2019 Machine Learning Research Internships at Curai]]></title>
            <link>https://medium.com/curai-tech/2019-machine-learning-research-internships-at-curai-8fc15313b7a8?source=rss-891e9f27a4cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/8fc15313b7a8</guid>
            <category><![CDATA[neurips-2019]]></category>
            <category><![CDATA[research-internship]]></category>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[health-technology]]></category>
            <dc:creator><![CDATA[Anitha Kannan]]></dc:creator>
            <pubDate>Fri, 06 Dec 2019 04:35:08 GMT</pubDate>
            <atom:updated>2020-02-18T22:37:17.627Z</atom:updated>
            <content:encoded><![CDATA[<p>Curai’s mission is to provide the world’s best healthcare for every human being. We are building an augmented intelligence system to help scale physicians’ abilities as well as to lower users’ barrier to entry to care.</p><p>Our internship program brings together students from diverse and strong technical backgrounds. Teamed with mentors, interns work on problems that are important to make deep inroads into our mission. Our internship program is open throughout the year, and there are three different <a href="https://www.curai.com/work/internships">internship openings</a> for students: engineering, research and medical. This post covers the research of our research interns this summer. You can also read about the work of our engineering interns in this other <a href="https://medium.com/curai-tech/2019-summer-internships-recap-cd4cd44ed0ea">blog post</a>.</p><p>We had three awesome research interns this summer: <a href="https://chmccreery.github.io/">Clara McCreery</a> and <a href="https://www.linkedin.com/in/sam-shleifer-a5307a77/">Sam Shleifer</a> from Stanford, and <a href="https://virajprabhu.github.io/">Viraj Prabhu</a> from Georgia Tech. It’s worth noting that Viraj Prabhu was a returning research intern! They made tremendous research progress in three main areas: <a href="https://arxiv.org/abs/1910.04192">domain relevant embeddings for question similarity</a>, formulation of <a href="https://arxiv.org/abs/1911.08554">decoder for medical dialog as a classification task for trading flexibility for control</a> and <a href="https://arxiv.org/abs/1910.02830">open set medical diagnosis</a>. We are quite excited to share that the research from their internships have been peer-reviewed and presented at <a href="https://ml4health.github.io/2019/">NeurIPS 2019 Machine Learning for Healthcare</a> (ML4H) workshop. While you can read their work in the above links, what follows in this blog post is an edited short version of their work, in their own words.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*dWHyZHB_5uM9fM3j2TbWzw.png" /></figure><p>If you are interested in helping us with our mission, please apply for an <a href="https://www.curai.com/work/internships">internship</a> or on any of our <a href="https://www.curai.com/work">other positions</a> by emailing your resume to jobs@curai.com.</p><h3><strong>1. For Finding Similar Medical Questions, Domain Matters</strong></h3><p><em>Clara McCreery, Stanford</em></p><p><strong>TL;DR: The task of determining whether a given pair of medical questions are “similar” or not is closely related to the task of medical question answering. In fact, these tasks may be more closely linked (in terms of low-level features) than question-similarity tasks in different domains are. We show how the task of classifying whether a statement adequately answers a medical question (medical question answering) works well for transfer learning to the medical question-similarity task. Our work highlights the importance of choosing both, the right pre-training task as well as the domain from which that pre-training data comes. </strong><a href="https://arxiv.org/abs/1910.04192"><strong>Full paper</strong></a></p><p>This Summer I completed a 10-week research internship at <a href="https://www.curai.com/">Curai</a>, a start-up in Palo Alto whose goal is to make primary care affordable and scalable. Curai’s technology drives their chat-based platform, <a href="https://curaihealth.com/">Curai Health</a>, where doctors interact with patients. This internship gave me an opportunity to take the lead on an open-ended research question and address the challenges of working with real-world data. Whereas many academic research questions are formulated around datasets that already exist, companies often care about problems for which labeled and vetted data do not yet exist. Not a lot of start-ups prioritize contributing to the research community the way Curai does, so this internship was a rare chance to see how fast-paced start-up culture can coexist with and even benefit from forward-thinking research.</p><h4>The Problem</h4><p>During my internship, I tackled the following natural language processing (NLP) problem: How can we accurately detect pairs of <em>semantically similar</em> medical questions? For example, the pairs of questions below should be identified as similar and dissimilar respectively, even though the first pair have few words in common, and the second pair differ by only one word.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/497/1*0hR7eq2-hmZBSS-LXsgy0w.png" /><figcaption>Examples of pairs of similar and different questions</figcaption></figure><p>There are several reasons why we care about medical question similarity:</p><ul><li>Some patient-asked questions are factual and will always have the same answer. If these questions have been answered before, it is faster to have an algorithm surface the answer than it is for a doctor to answer the question</li><li>It is expensive and not a good use of doctors’ time to answer the same questions over and over if the questions do not require follow-up questions or in-person examination</li><li>Different doctors have different experiences and expertise, and such a system can aggregate knowledge, following the principle that <a href="https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2726709">many heads are better than one</a></li><li>Collecting information about common primary complaints over time can give scientists insights into the spread of disease, onset of epidemics, etc.</li></ul><p>The problem of question similarity is of interest to many beyond the medical field. The question-answering forum Quora is one of many businesses that are always trying to improve their algorithms for detecting and combining duplicate questions. In fact, a common general language understanding benchmark, the <a href="https://gluebenchmark.com/">GLUE benchmark</a>, includes a task called Quora Question Pairs (QQP), in which a model is supposed to label question pairs as either similar or different. What is unique about this problem in the medical domain is that <strong>domain expertise</strong> is required to adequately label medical questions as similar or not, and most <a href="https://lilianweng.github.io/lil-log/2019/01/31/generalized-language-models.html">general-language models</a> have no such expertise. Furthermore, as no large labeled datasets of similar medical questions exists, it is unclear how to best embed this knowledge into a machine learning model.</p><h4>Transfer Learning</h4><p>One common technique for overcoming a dearth of training data is transfer learning, in which a neural network is trained first on a large dataset that is related to but different from the task of interest. Then, the weights from that first network are <em>transferred to</em> or used to initialize the weights of a network being trained on the actual task of interest. This often works well because related tasks leverage similar low-level features/weights, and learning these low-level features requires a lot of training data.</p><p>Although it is expensive and infeasible to generate a medical-question-pairs dataset large enough to train a neural network from scratch (QQP uses 363,000 question pairs), it is more manageable to generate a small dataset (3,000 question pairs) for fine-tuning and testing our models. We started with <a href="https://github.com/durakkerem/Medical-Question-Answer-Datasets">1,500 freely available patient-asked questions from HealthTap</a>, to ensure that questions were representative of what real patients ask online. For each question, we then had healthcare professionals generate two question pairs: one positive and one negative. The positive pair used different language to ask a similar question, whereas the negative pair used similar language to ask a different question.</p><p>We then set out to find a pre-training dataset and task for transfer learning that would produce useful low-level features for our task of medical question similarity. One obvious dataset to choose was the previously-mentioned QQP. Although these questions are not medical, they do represent many examples of semantically similar and different phrases. This is the same task in a different domain.</p><p>Additionally, although there are no large datasets of medical <em>question pairs</em> online, there are many medical <em>question-answer pairs</em> from websites such as HealthTap <a href="https://github.com/durakkerem/Medical-Question-Answer-Datasets">[1]</a> <a href="https://github.com/LasseRegin/medical-question-answer-data">[2]</a>. Some of this data is labeled with the category to which the question belongs. We leveraged this data to generate three more pre-training tasks in the medical domain:</p><ol><li>Question-Answer Pairs (QA): Here, positives were questions and the true answer that a doctor provided online. Negatives were questions and a doctor-written answer to a different question</li><li>Answer-Answer Pairs (AA): Positives were complete medical answers. Negatives were the first 2 sentences from one medical answer followed by the last several sentences of a different answer</li><li>Question-Category Pairs (QC): Positives were medical questions and their correct categorical label. Negatives were medical questions and a random incorrect categorical label</li></ol><p>For each pre-training dataset we trained a model first on the pre-training task, and then fine-tuned it on our medical question pairs dataset. We performed these experiments using the <a href="https://arxiv.org/abs/1810.04805">BERT</a> model (results below) and replicated them with <a href="https://arxiv.org/abs/1906.08237">XLNet</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/589/1*mn3-Xsbw3_NPT36-eifevw.png" /></figure><p>As shown in the plot (<em>X-axis: number of training examples used in the final task viz. medical question similarity</em>), we found that the best pre-training task was that of Question-Answer (QA) matching. This highlights the importance of both the pre-training task as well as the domain from which that pre-training data comes. For instance, despite AA and QC data being medically relevant, pre-training on QQP data (which is generic) outperforms both of them. However, pre-training on QA data (also medically relevant) outperforms QQP. Furthermore, we also observe that this outperformance is amplified for smaller final-task training sets. To dive further into the cause for this improvement, we looked at the errors made by each model. Our qualitative error analysis shows how some medical synonyms that prompted false negatives in the Quora model (<em>eg.</em> hypertension/high blood pressure and menstrual cycle/period) were learned by the models pre-trained on the medical datasets, which got these questions correct. We present many such examples in <a href="https://arxiv.org/abs/1910.04192">our paper.</a> Through our error analysis, we gained insights into the types of mistakes that our best models make.</p><h3>2. Classification as Decoder: Trading Flexibility for Control in Medical Dialogue</h3><p><em>Sam Shleifer, Stanford</em></p><p>Over the summer, I worked with Manish Chablani on AI assisted medical conversation automation. A step in that direction is suggesting responses to doctors they can use or make edits to in medical conversations with users of tele-medicine. Suggested Response is a difficult task because the model must be able to handle a wide range of primary care dialogue scenarios, including courtesy responses (<em>eg.</em> “Feel Better soon”), asking patients relevant questions about their symptoms to narrow down disease diagnosis space, and offering advice on managing the symptoms. Here are some examples:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/644/1*qx_mKTKUl81z68UBenvTaw.png" /><figcaption>Some examples of suggested responses</figcaption></figure><p>We experimented with generative models based on ULMFit as well as other transformer based architectures including GPT2. Generative models, are trained to predict the next word, and at inference time use beam search to generate a likely set of responses. Generative approaches can generate a likely response in any conversational context, making them incredibly flexible. This flexibility comes at the cost of control. Undesirable responses in the training data will be reproduced by the model at inference time, and longer generations often don’t make sense. For example, in one conversation about drug interactions generative system proposed “It is not advisable to take Sudafed with Sudafed” to be the most likely response.</p><p><strong>Our Approach: Classification as Decoder</strong></p><p>To avoid ineffective and inaccurate generations, I trained a classifier to choose from a predefined list of full responses. The classifier is trained on (conversation context, response class) pairs, where each response class is a noisily labeled group of interchangeable responses and is associated with an “Exemplar response”. At inference time, the classifier just looks up the exemplar response associated with the predicted class ID in a dictionary.</p><p>The discriminative approach has a few key advantages: First, doctors can edit and improve these exemplar responses over time, without retraining the model. For example, if we wanted to switch from recommending `users sleep 6–8 hours per night’ to `recommending 7–9 hours’, we could simply update the message associated with the output class and the discriminative model would immediately generate new advice in the same conversational context, without retraining.</p><p>The main drawback, and the reason this approach is not widely used, is that it requires a high-quality (conversational context, response class) training dataset, which requires clustering the responses observed in the data into interchangeable response “classes” that are numerous enough to contain reasonable responses in a wide range of conversational topics, while also having enough examples for each class that the model can learn it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VMLuugTVDrmp_XZP_47SIw.png" /></figure><p>I spent the first few weeks of the summer on this system, which includes a weakly supervised pipeline that relies on techniques like LSH with pretrained sentence encoders, a supervised BERT-based sentence similarity model, Agglomerative Clustering and three hours of manual labeling. The process is detailed at length in our <a href="https://arxiv.org/pdf/1911.08554.pdf">paper</a>, but eventually generates 187 distinct response classes that appear frequently in the conversational logs. Each response class has a “exemplar” response for that class which is used at inference time as suggested response. We ensure that exemplars are all factual, sensical and grammatical by allowing experts to edit them before or after training.</p><p><strong>Experiments</strong></p><p>The response clustering problem is unsupervised, and can only be evaluated as part of the end to end response suggestion problem. More specifically, by asking medical experts whether the responses generated by one system are better than responses generated by another system. Luckily, the results were very encouraging: For a random sample of 775 conversations-response pairs, only 12% of our discriminative approach’s responses are worse than the doctor’s response in the same conversational context, compared to 18% for the generative model.</p><p>The rest of the summer was spent implementing and comparing different architectures, including triplet loss, Huggingface’s GPT based model, <a href="https://arxiv.org/abs/1507.04808">HRED</a>, and <a href="https://arxiv.org/abs/1801.06146">ULMfit</a>. At the beginning of the summer we suspected that we would need a more traditional ranker approach, but with experimentation, we discovered that models that used triplet loss or a multiple choice loss took longer to train, spent more time at inference, and did not significantly improve classification accuracy, as shown below. For the final version, we settled on ULMFit, since it performed as well as the larger transformer based approach, but offered much faster training and inference.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*kM6N8F_Is6jxI51l" /><figcaption><em>Classification Accuracy for 187 classes using either 4 or 8 turns of conversational history</em></figcaption></figure><p>Our <a href="https://arxiv.org/abs/1911.08554">paper</a> discusses more experiments on different response class generation procedures, model architectures, and tradeoffs.</p><h3>3. Open set medical diagnosis</h3><p><em>Viraj Uday Prabhu, Georgia Tech</em></p><p>In the US, an increasing number of adults (about a third!)[<a href="http://www.nbcnews.com/id/3077086/t/more-people-search-health-online/#.Xein1OdKglJ">3</a>][<a href="https://www.beckershospitalreview.com/healthcare-information-technology/google-receives-more-than-1-billion-health-questions-every-day.html">4</a>] use the Internet to diagnose medical concerns and online symptom checkers are increasingly part of this process. These tools are powered by <a href="https://medium.com/curai-tech/the-science-of-assisting-medical-diagnosis-from-expert-systems-to-machine-learned-models-cc2ef0b03098">diagnosis models</a> similar to clinical decision support systems, and they walk patients through a series of questions about their symptoms, and finally provide a diagnosis. Such services are poised to revolutionize patient-facing telehealth services that could move from current rule-based protocols for nurse hotlines to more accurate and scalable AI systems.</p><p>However, existing models for diagnosis make a <em>closed-set </em>assumption, i.e. the universe of diseases is limited to those that the diagnosis model can understand. In practice, it is infeasible to obtain sufficient training data for every human condition (over 14,025 diagnosis codes and 2000 disease families exist in <a href="https://www.icd10data.com/ICD10CM/Codes">ICD-10</a>), and so it is highly likely that a diagnosis model, once deployed, will encounter cases corresponding to previously unseen conditions. To address this, telehealth providers often constrain their coverage to a specific area of care. However, determining whether or not a patient falls within diagnostic scope based on symptoms often necessitates additional models or human expertise, which introduces higher costs and is itself error-prone. Further, each misdiagnosis is a missed opportunity for better care, and may even be safety-critical.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*AZNOMYVH9PSVrhGT" /></figure><p>Consider the example above, that illustrates a typical symptom checking workflow. For the given list of user-reported findings, that include muscle rigidity, jaw pain, and muscle spasms, a closed-set diagnostic model predicts <em>whiplash injury</em> to be the most likely condition, followed by <em>scarlet fever</em> and <em>diverticulitis</em>. In this case, the underlying condition is actually <em>tetanus</em>, which the diagnosis model has never been trained to diagnose. Such misdiagnosis is an artifact of the ‘forced-choice’ nature of such closed-set models. Further, in this case the model is a deep neural network — while such models tend to achieve state of the art performance on several tasks, they often tend to be highly overconfident about their predictions, which is evidenced in the high confidence (95.3%) assigned to the incorrect prediction in this case. Both of these limitations pose significant challenges to the deployability of such diagnostic models.</p><p>Prior work in machine learning has studied the <em>open-set learning </em>problem (and the related problem of learning with reject option), where the goal is to design approaches that are aware of and can avoid misclassifying previously unseen classes. Inspired by this, we frame machine-learned diagnosis as an open-set classification problem, that we call <strong>Open Set Diagnosis</strong>, and study how well proposed approaches apply to diagnosis. In the example above, our goal is to design an open-set diagnosis model that predicts “Don’t know” for the given set of findings, and recommend additional diagnostic evaluation (such as physical exams, lab tests, and imaging studies), rather than misdiagnose.</p><p>Another critical challenge in building diagnosis models is <em>access to data. </em>Health data usually lives in hospital repositories and for privacy reasons can often not be taken outside its source site to be pooled with other sources. This makes training diagnosis models (particularly data-inefficient neural networks) difficult. Further, healthcare sites often have data that is complementary — for example, hospitals on the US east coast are likely to have far more patient encounters for hypothermia than on the west. To develop models with high accuracy and coverage, we need mechanisms to bridge models trained on separate sites. To this end, we introduce the task of <strong>Ensembled Open Set Diagnosis</strong>, where we ensemble models trained on data sources that cannot be shared, and evaluate their open-set diagnosis performance.</p><p>Concretely, we study two tasks that are illustrated below. In Task 1 (Open Set Diagnosis), our goal is to learn a model to diagnose a select set of diseases (call this Lselect), and reject unseen conditions (call this Lunknown), i.e declare “Don’t Know”. We are optionally given access to case data for a subset of conditions Lextra that is disjoint from Lselect, and that we may use for modeling unseen conditions. In task 2 (Ensembled Open Set Diagnosis), the goal and evaluation setting is identical to Task 1; however, training data is now <em>distributed</em> across multiple healthcare sites.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*8CBEGLERMPyYL47Z" /></figure><p>For our dataset, we simulate clinical cases using the simulation algorithm proposed in this <a href="https://arxiv.org/abs/1804.08033">prior work from Curai</a>, and create splits for Lselect, Lextra, and Lunknown. We measure performance as the diagnostic accuracy corresponding a given false positive rate (<em>i.e.</em> rate of misdiagnosing an unknown condition as a known one). Our goal is to achieve a high accuracy at a low false positive rate. We evaluate over a large and challenging testset of over 200k cases corresponding to diseases in Lselect and Lunknown.</p><p>We experiment with three different training approaches: i) a simple cross-entropy (CE) baseline where we threshold on the confidence of the final softmax layer to predict “Don’t Know” ii) training with an additional background (BG) class as a catch-all for unseen conditions, the extra class being trained with examples from Lextra, and iii) a state of the art open-set learning method called the Entropic Open-Set (EOS) loss, which trains with regular cross-entropy for seen (Lselect) datapoints, and encourages high entropy otherwise (for Lextra). For training ensembles in Task 2, we experiment with both a <em>naive</em> ensembling strategy (predicting the class with the highest confidence) and a <em>learned</em> strategy where we train an additional network to combine the outputs from individual expert models that are trained at different sites.<em> </em>All our models are parameterized as 2-layer neural networks, and we represent cases with one-hot encodings using a global symptom vocabulary.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/808/1*07Rcqsvrd89KF3Kac2L_JQ.png" /></figure><p>Our results are summarized above. On Task 1, we find the BG approach to clearly outperform EOS. In Task 2, with naive ensembling we find the same trend to hold. However with learned ensembling, EOS outperforms BG (not shown here). In Task 2, we also report the corresponding performance of the method on Task 1 as an “oracle” upper bound, with the gap representing the error introduced by distributed training. Across all settings, approaches that model unseen conditions (BG and EOS) <em>consistently</em> outperform a baseline that does not (CE). Further, we find that open-set approaches do not perform any worse at closed-set diagnosis (i.e. when only seen conditions occur at test time). For additional experiments, results, and analysis, please read our paper <a href="https://arxiv.org/abs/1910.02830">here</a>.</p><p>This work only scratches the surface and several exciting directions of follow-up work remain. These include varying model families across different sites, disease distribution shift across sites and deployments. All of these are essential challenges to overcome on the road to building reliable models for diagnosis.</p><p>Finally, I’d like to thank all my collaborators on this work (Anitha, Geoff, Namit, Manish, David, and Xavier) for their guidance through this project, and to everyone at Curai for the awesome internship experience!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8fc15313b7a8" width="1" height="1" alt=""><hr><p><a href="https://medium.com/curai-tech/2019-machine-learning-research-internships-at-curai-8fc15313b7a8">2019 Machine Learning Research Internships at Curai</a> was originally published in <a href="https://medium.com/curai-tech">Curai Health Tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The science of assisting medical diagnosis: From Expert systems to Machine-learned models]]></title>
            <link>https://medium.com/curai-tech/the-science-of-assisting-medical-diagnosis-from-expert-systems-to-machine-learned-models-cc2ef0b03098?source=rss-891e9f27a4cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/cc2ef0b03098</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-expert-systems]]></category>
            <category><![CDATA[medical-diagnosis]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[ehr]]></category>
            <dc:creator><![CDATA[Anitha Kannan]]></dc:creator>
            <pubDate>Mon, 15 Apr 2019 13:01:01 GMT</pubDate>
            <atom:updated>2019-04-15T13:01:01.371Z</atom:updated>
            <content:encoded><![CDATA[<p>Curai’s mission is to scale the world’s best healthcare for every human being. We are building an augmented intelligence system to help scale physicians’ abilities as well as to lower users’ barrier to entry to care. There are many components to such a system, but disease prevention, diagnosis, and treatment are central to providing best clinical service.</p><p>The medical diagnostic process, defined as <em>“a mapping from a patient’s data (normal and abnormal history, physical examination, and laboratory data) to a nosology of disease states”, </em>[<a href="https://link.springer.com/chapter/10.1007/978-3-319-31913-1_11">1</a>]<em> </em>starts with a <em>differential diagnosis</em> (DDx), consisting of a ranked list of possible diagnoses, that is used to guide further assessments and possible treatments. In 2015, the Institute of Medicine (IOM) called for a “significant re-envisioning of the diagnostic process” and reported that “nearly every person will experience a diagnostic error in their lifetime” [<a href="https://www.nap.edu/catalog/21794/improving-diagnosis-in-health-care">2</a>]. Just in the US alone, outpatient diagnostic errors account for 5.08% of medical errors, affecting 12 million US adults every year, with more than half of these errors being potentially harmful [<a href="https://qualitysafety.bmj.com/content/23/9/727">3</a>]. A recent <a href="https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2726709">study</a> demonstrated that diagnosis accuracy is around just 60% when only one medical provider is involved in the decision-making process.</p><p>The need for diagnostic decision support systems for assisting medical providers has been well understood for a long time. Medical experts, including Bruce G. Buchanan, Edward H. Shortliffe, and Jack Myers pioneered some AI efforts (<a href="https://dl.acm.org/citation.cfm?id=1096485">Mycin</a> and <a href="https://www.ncbi.nlm.nih.gov/pubmed/7048091">Internist-1</a>) directed towards this goal. At Curai, we are researching how to use these curated expert systems as a prior for modern machine-learning-based approaches. This represents a real step forward in AI-based diagnostic systems that combine the best of both old and new-age approaches: latent causal relationships that are encoded in such expert systems but are further probabilistically refined and continually self-improving by learning from new data.</p><h3>Why is medical diagnosis hard?</h3><p>Before discussing the role of AI systems in medical diagnosis, we need to understand why diagnosis is such a hard problem. In a perfect world, medical diagnosis can be viewed as a complete information game in which the medical provider has <em>all</em> the information needed for diagnosis, and the entirety of this information is used to draw conclusions. However, there are many factors that prevent this from happening in the real world:</p><ul><li>Patient holding back on pertinent information that they perceive as unimportant or irrelevant.</li><li>Cost of acquiring information, whether that be the cost of a test or the time it takes to further talk to the patient, given that a typical medical provider’s visit lasts only 15 minutes.</li><li>Reliance on recency and availability bias due to sheer number of diseases that are possible, and the need to quickly arrive at a diagnosis [<a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5093937/">4</a>].</li><li>Lack of appropriate diagnostic tools to gather complete information about the patient for certain diseases such as mental health, chronic fatigue and lyme disease.</li><li>Patient accessing different disconnected areas of healthcare resulting in incomplete view of their medical record.</li></ul><p>In addition to all these factors, we also want to draw attention to the fact that diagnosis is an evolving task that is temporal in nature — as a medical provider acquires more information, not only does the context in which they are diagnosing changes, but the actual diagnosis for a patient may change. For example, a simple cough could often be thought as the result of a cold, bronchitis, or acid reflux when seen in isolation. When that cough has been worsening slowly over a longer period of time, other diagnoses must be considered.</p><p>At Curai, we believe that AI diagnostic systems help medical providers solve the complete information game; they can play an integral role in gathering all the required information from the patients. With the task being hard enough as is, we believe software can assist medical providers not only in solving patients’ problems but also by enumerating possibilities and edge cases they may have otherwise missed.</p><h3>AI for medical diagnosis</h3><p>The figure below provides an overview of the evolution of AI models that assist doctors with differential disease diagnosis.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/960/0*YWGR2pfsmo9WKZGj" /><figcaption>Overview of evolution of AI models for differential disease diagnosis</figcaption></figure><h3>Medical Expert Systems</h3><p>Diagnostic expert-based systems are computer systems that seek to emulate the diagnostic decision-making ability of human experts. Some notable systems include <a href="https://dl.acm.org/citation.cfm?id=1096485">Mycin</a> for infectious diseases, and <a href="https://www.ncbi.nlm.nih.gov/pubmed/7048091">Internist-1</a>, <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC2233214/">QMR</a> and <a href="https://jamanetwork.com/journals/jama/fullarticle/366898">DXplain</a> for general internal medicine.</p><p>Medical expert systems generally include two components: (1) a knowledge base (KB), which encapsulates the evidence-based medical knowledge that is curated by experts, and (2) a rule-based inference engine devised by the expert, which operates on the knowledge base to generate a differential diagnosis.</p><p>Diagnostic knowledge bases generally consist of diseases, findings (i.e. symptoms, signs, history, or lab results), and their relationships. In many cases, they explicitly lay out the relationships between a set of findings and the things that cause them (diseases). For example, a KB might include <em>influenza</em> and show its relationships with <em>fever, coughing, </em>and <em>congestion.</em> A common approach to modeling relationships between diseases and findings is by using variables that encode positive predictive value and the sensitivity of a finding to a disease. In particular, the <em>evoking strength</em> (positive predictive value) captures how strongly one should consider a disease if the finding was observed, while the <em>frequency</em> (sensitivity) models how likely it is that a patient with a disease manifests a particular finding.</p><p>The rule-based inference engine outputs a ranked differential diagnosis by scoring the diseases in the knowledge base as a function of their relationship strengths over all of the input findings. In other words, given a set of findings from a patient, inference engine examines the strength of the relationships those findings have with each disease in the KB and sort based on some defined scoring function. See [<a href="https://link.springer.com/chapter/10.1007/978-3-319-31913-1_11">1</a>] and [<a href="https://arxiv.org/abs/1804.08033">4</a>] for additional details.</p><p>Unfortunately, the practical use of these systems has been constrained by several factors [<a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC116181/">5</a>]. First, maintenance of expert systems requires a dedicated team of experts who continually monitor and reconcile new research. As a practical example, most studies regarding <a href="https://www.ncbi.nlm.nih.gov/pmc/articles/PMC3715105/">diabetes in the South Asian population</a> have occurred fairly recently and the knowledge from these new studies can be incorporated into the knowledge base only with laborious and time consuming manual intervention. Second, it may take several years for a newly identified disease to be part of the expert system, as it may be costly to run medical studies. Consider, for example, the newly identified Zika virus, which is not yet well understood and thus cannot be incorporated into a medical knowledge base with any confidence. Third, an inference engine based on an underlying knowledge base is devised with the assumption that it will be used in a noise-free environment with all the information available. For example, a patient might answer no to a question from the system about having a hematoma simply because they don’t understand that the bruise on their leg is called a hematoma.</p><h3>Learning to Diagnose using Expert Systems as a Prior</h3><p>There has been prior research on <a href="http://www.jstor.org/stable/2345762">interpreting expert systems as a probabilistic graphical model</a>. These models typically have a bipartite causal graph structure modeling dependencies between diseases and findings (see figure below) . The prior probabilities of the diseases and conditional probability distribution of the findings are <em>fixed</em> and derived from the variables in the expert system. When some findings are observed, probabilistic inference can be used to obtain the posterior distribution (differential diagnosis) over the diseases. Due to explaining away and uncertainty modeling, the model can be resilient to noise. However, exact inference is difficult in this model family and several approximate inference methods such as <a href="https://arxiv.org/abs/1304.1141">Markov Chain Monte Carlo</a>, <a href="https://arxiv.org/abs/1105.5462">variational inference</a> and <a href="https://arxiv.org/abs/1301.2295">recognition networks</a> are computationally expensive or not accurate enough to be deployed in practical settings.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/285/1*dpir_NEIwA95yDsT1fx9SQ.png" /><figcaption>Probabilistic graphical model for differential disease diagnosis</figcaption></figure><p>Last year, we set out to tackle some of the discussed challenges, and asked the following questions: How could we improve the resiliency of expert systems in the presence of partial or noisy inputs? How could we leverage the curated medical knowledge encoded in the expert systems, as we learn models from other data sources such as electronic health records? We <a href="https://arxiv.org/abs/1804.08033">answered</a> these questions by letting the knowledge encoded in expert systems serve as the <em>prior</em> for learning a new diagnosis model from scratch. The central idea behind our work is that you can utilize expert systems as a <em>data generator</em> and the generated synthetic medical cases can be used as labeled data for training a model. The figure below gives a quick overview of our approach. Note that the models provide the optional ability to incorporate clinical cases from other data sources. We presented the detailed results in this <a href="https://arxiv.org/abs/1804.08033">paper</a> that showcased the resiliency of the approach to noisy inputs, and the efficacy of the approach on additional cases from electronic health records.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*NWgNXCKoO_dByNXQsthkxQ.png" /><figcaption>Medical Knowledge as prior for learning model for differential diagnosis</figcaption></figure><p><strong>Clinical case simulation: </strong>We used a simulation algorithm described in [<a href="https://www.ncbi.nlm.nih.gov/pubmed/2695784">6</a>] to generate clinical vignettes conditioned on different diseases. The simulator iteratively sampled findings based on their relationship to the pertinent disease for which the clinical vignette was generated. The figure below shows simulated examples corresponding to two diseases: acute viral hepatitis and acute septic arthritis.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*E5v9g8RWAd_0AijF" /><figcaption>Examples of clinical cases simulated from the expert system. For each disease, the figure shows two possible clinical presentations.</figcaption></figure><p><strong>Modeling choices and results</strong>: We posed diagnosis as a classification task and experimented with different modeling approaches. In the figure below, you can see that training a model drastically improved accuracy over other approaches including probabilistic inference on a graph derived from expert system and invoking the inference engine of the expert system.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/464/1*Gr9zb43IXMbn0xUc-B-GGw.png" /><figcaption>Comparison of performance of algorithms for differential diagnosis</figcaption></figure><p>Similarly, our deep neural network was much better than linear models such as logistic regression. This is not altogether surprising, since a deeper network allows for learning interdependencies between findings to explain away diseases. You can also see from the figure that our final learned model is much more resilient to noise at the time of diagnosis. This is an important property, especially when the model is deployed to assist real medical providers; in such a setting, inputs to the model can (and likely will) be noisy. As further evidence towards how this method can generalize and expand the capabilities of expert systems, we demonstrated that by adding additional data sources, we can expand disease coverage and more precisely model existing diseases.</p><h3>Conclusions</h3><p>So what does this all mean? The work described here represents, to us, an exciting and discrete step forward in self-improving diagnosis systems from the static and somewhat brittle world they have lived in for a long time. It is worth adding that we feel these types of methods represent an exciting area of AI research as they present opportunities for mixing causal forms of AI research with the massive step functions deep learning has introduced in the past decade to achieve better accuracy and resiliency to noise. We also found that this approach can be a valuable tool in settings where you can cobble together a number of diverse but related data sets.</p><p>There is very exciting research ahead of us that should vastly improve the quality of medical diagnosis and patients’ outcomes, eventually making this a world in which <em>“</em><strong><em>no</em></strong><em> person needs to experience a diagnostic error in their lifetime.</em>”</p><p>If the mission of scaling the world’s best healthcare for every human being excites you and you want the opportunity to contribute to these exciting advances in AI research, please check out our <a href="https://www.curai.com/careers">career page</a> or reach out directly.</p><h3>Acknowledgements</h3><p>I would like to thank my collaborators, Murali Ravuri, Geoffrey Tso, and Xavier Amatriain, who were integral to the <a href="https://arxiv.org/abs/1804.08033">research</a> described in this post. Thanks to Jack Craddock, Neal Khosla and Vignesh Venkataraman for their detailed inputs to this blog post.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cc2ef0b03098" width="1" height="1" alt=""><hr><p><a href="https://medium.com/curai-tech/the-science-of-assisting-medical-diagnosis-from-expert-systems-to-machine-learned-models-cc2ef0b03098">The science of assisting medical diagnosis: From Expert systems to Machine-learned models</a> was originally published in <a href="https://medium.com/curai-tech">Curai Health Tech</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>