<?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 Albert Bacelar on Medium]]></title>
        <description><![CDATA[Stories by Albert Bacelar on Medium]]></description>
        <link>https://medium.com/@abacelar?source=rss-7a8f1e8140cf------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/2*Iq_ejJC-zXAx6FYSG7xGPA.jpeg</url>
            <title>Stories by Albert Bacelar on Medium</title>
            <link>https://medium.com/@abacelar?source=rss-7a8f1e8140cf------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sun, 22 Nov 2026 12:55:16 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@abacelar/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[Agentic AI in Healthcare: From Pilot to Production]]></title>
            <link>https://abacelar.medium.com/agentic-ai-in-healthcare-from-pilot-to-production-de83c2f838f9?source=rss-7a8f1e8140cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/de83c2f838f9</guid>
            <category><![CDATA[healthcare-technology]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Albert Bacelar]]></dc:creator>
            <pubDate>Sun, 30 Aug 2026 18:35:32 GMT</pubDate>
            <atom:updated>2026-08-30T18:35:32.622Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ecb1epHcmiy8J26aPyas6g.png" /></figure><h3>When I use the term agentic AI in healthcare, I am describing an architecture that executes an entire clinical or administrative workflow through specialized functions coordinated by a supervisor, with controlled access to systems, formal validation, reconciliation, exception handling, auditing, and human intervention at the points where risk requires professional judgment. The unit of work becomes the entire process: receiving a data element, identifying its source, transforming its structure, normalizing its meaning, checking conformance, reconciling identity and context, deciding where the information should go, routing exceptions to the appropriate professional, and recording each step. The reference article distributes this work across a connection agent, a mapping and transformation agent, a terminology agent, a conformance validation agent, a reconciliation agent, and an exception triage agent, all coordinated by a supervisory layer. This decomposition makes execution traceable because every result is associated with a specific input, rule, version, and component.</h3><p>Moving from pilot to production starts before choosing the model. It starts by selecting the clinical workflow, defining exactly what the system will do within that workflow, and delimiting which actions will affect the patient record, a prescription, a device, a clinical queue, or another hospital system. The pilots described in the material stall for three recurring reasons: the absence of a confidence policy separating automation from human review, the absence of a durable audit trail, and treating the language model as if it were the entire system rather than one component inside a broader layer of information retrieval, validation, and rollback. The comments add another important point: the architecture needs to be designed from the outset as a governed workflow, with escalation rules, understandable communication between humans and agents, and a precise definition of what the human reviewer is actually reviewing.</p><p>I would start by describing the use case in operational terms. Let us use laboratory-result integration with the electronic health record as an example. The input is a message originating from the laboratory system. The output is a clinical observation written to the correct record, linked to the correct patient and encounter, containing the test, value, unit, reference range, collection time, release time, critical-result flag, and source identification. Between those two points lie several distinct risks: linking the result to the wrong patient, duplication, delay, incorrect unit conversion, loss of the reference range, incorrect semantic mapping, truncation of the result, confusion between preliminary and final results, and reprocessing an old message as if it were new. The first project meeting should produce this map, including the current workflow, responsible parties, systems involved, failure points, volume, urgency, operating hours, known exceptions, and the clinical consequence of each error.</p><p>This scoping prevents a common mistake: creating several agents first and then looking for a task for them. The number of agents should follow from the actual work. Stable transformations such as validating a required field, checking the data type, applying a known mathematical conversion, verifying a signature, controlling authorization, or preventing duplication belong in deterministic code. The language model enters where there is semantic ambiguity, document variability, a need to consult controlled sources, or coordination across steps. Even then, it operates within explicit boundaries, using authorized tools, defined input and output schemas, and stopping rules. A production design keeps the model as one component of the chain, surrounded by validation, auditing, and rollback.</p><p>A technical aside is useful here. An LLM is a large language model trained to process and generate text and other symbolic structures. An agent is a software unit that receives a constrained objective, consults tools, executes steps, checks results, and returns structured output. The supervisor, also called a planner or orchestrator, controls the permitted sequence, decides which agent acts at each point, aggregates validation signals, and closes, retries, or routes the case. In clinical production, I would restrict this supervisor to a predefined state graph with permitted transitions, entry conditions, retry limits, and exception destinations, preventing every message from generating a new and unpredictable path. This reduces unpredictable behavior and makes auditing easier.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*emHpdkvofhaTfWI2g8M5_A.png" /></figure><p>The second step is to build the interoperability foundation. The diagram on page 2 brings together the electronic health record, health information exchange networks, clinical and financial management systems, and medical devices connected through HL7 v2, FHIR R4, and C-CDA. These systems feed a connection agent; a supervisor then coordinates the specialized agents; every action is sent to an audit and observability layer; and changes affecting clinical meaning stop at a human approval stage before reaching the canonical model and downstream systems.</p><p>For a physician, these terms need to be translated into practice. HL7 v2 is an older standard that remains widely used in hospitals, based on messages composed of segments and fields and used for admissions, discharges, transfers, laboratory results, orders, and other events. FHIR R4 organizes clinical data into resources such as Patient, Encounter, Observation, Medication, and AllergyIntolerance, often transported as JSON through service interfaces. C-CDA structures clinical documents, such as summaries and reports, in a standardized format. EHR or EMR refers to the electronic health record. HIE and QHIN refer to structures for information exchange across organizations. PMS is the practice or service management system. RCM refers to revenue cycle management. A physician does not need to program these standards, but does need to understand that the same clinical concept may arrive with different structures, codes, units, and levels of completeness.</p><p>This is why the architecture needs a canonical clinical data model. The canonical model is a common internal representation used to receive data from different sources before sending it to downstream systems. One laboratory may call a test “TNIHS,” another “High-Sensitivity Troponin I,” and another may send a proprietary code. The canonical model stores a single representation containing terminology code, name, method, value, unit, reference interval, specimen, timestamp, and provenance. Provenance identifies where the data came from, when it was captured, which message carried it, and which transformations were performed. Without this layer, every new integration becomes an isolated project with duplicated rules and a risk of divergence.</p><p>The work begins with a data contract for each interface. This contract defines required fields, accepted formats, local codes, units, handling of missing values, behavior when duplicate messages are received, standard version, authentication rules, expected response times, and rejection criteria. The article mentions vendor-specific certification tiers and contract testing to connect different EHRs without modifying the platform core. A contract test is an automated verification that the interface continues to send and receive data according to the technical agreement. If a vendor changes a field, unit, or response, the test detects the break before it reaches the medical record.</p><p>Once the data foundation is defined, I would distribute responsibilities. The connection agent receives the message, authenticates its origin, records the timestamp, preserves the raw content, assigns a correlation identifier, and responds to the sending system according to the protocol. That identifier follows the message through every stage, making it possible to reconstruct its trajectory. The mapping and transformation agent converts the source structure into the canonical model. It knows that one HL7 field corresponds to the test value, another contains the unit, and another carries the encounter identifier. The terminology agent converts local codes into standardized vocabularies and checks whether the selected concept preserves the clinical meaning. The conformance agent verifies whether the output follows the required schema, whether mandatory fields are present, whether data types are correct, and whether destination constraints have been met.</p><p>The reconciliation agent handles conflicts involving identity, context, and repetition. It verifies whether the patient, encounter, order, and test belong to the same clinical chain; distinguishes a correction from a duplicate; identifies when two systems contain conflicting versions; and decides whether manual linkage is required. The exception triage agent classifies the failure, identifies the appropriate owner, creates the remediation task, and controls the retry. The supervisor collects these outputs, applies the risk policy, and decides among automatic write, human review, temporary hold, or rejection. The material presents this architecture as six cooperating functions, with every decision linked to an agent, an input, and a rule, providing the basis for technical and clinical auditing.</p><p>Now let me walk through one message from beginning to end. The laboratory releases a high-sensitivity troponin I result. The message contains the local test identifier, a value of 52, the unit ng/L, a reference interval below 19, a flag indicating an elevated result, collection time, release time, patient identifier, and encounter number. The connection agent receives the message, confirms the source, stores an immutable copy, and creates the correlation identifier. The mapping agent then transforms the fields into the canonical model. The terminology agent consults the institution’s approved dictionary and maps the local code to the correct standardized concept while preserving method and unit. The conformance agent checks whether the result is numeric, whether the unit is compatible, whether the timestamps are valid, whether the patient and encounter have been provided, and whether the final resource complies with the destination system schema.</p><p>Next, the reconciliation agent compares the patient, encounter, order, and result. If the same message has already been processed, it prevents another write. If a preliminary result exists and a final result arrives, it classifies the relationship between the two versions. If the patient identifier conflicts with the order, the workflow triggers a mandatory stop. If the laboratory changed the unit without updating the contract, the case is routed for review. The supervisor receives all of these signals and applies the policy defined for that interface. When the mapping has already been approved, required fields are complete, identity has been reconciled, the unit is known, no clinical meaning has changed, and no high-risk rule has been triggered, the message proceeds to automatic write in the destination system. The technical term auto-commit means exactly this: the final write occurs without another human action, within a predefined set of approved conditions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*8IQrBimE-tjYFqdgHKEsVw.png" /></figure><p>When there is a mapping change, a new code, an unknown unit, an identity conflict, or another predefined condition, the message goes to the human review stage. If the reviewer approves it, the system writes the observation and records who approved it, which evidence was reviewed, and which version of the rule was active. If the reviewer rejects it, the triage agent classifies the cause, routes the task to the appropriate team, and keeps the message on hold. Once the dictionary, contract, or registration problem has been corrected, the message is placed back into the queue. Requeue means putting an event back into the processing queue. Replay means re-executing a preserved event with version control while retaining the history of what occurred before. The flow on page 3 follows this logic: message arrival, connection, transformation, terminology normalization, validation, confidence-based decision, write or review, triage after rejection, retry, and logging of every branch in the audit trail.</p><p>The most important objection raised in the comments concerns confidence policy. A single number produced by the model does not represent clinical risk. A mapping with 98% confidence still requires interruption if it involves a potentially wrong patient, an incompatible unit, a drug allergy, a dose, a critical result, a diagnosis change, a change in clinical meaning, or an irreversible action. Risk depends on context, documentation quality, consequence of error, urgency, scope of the downstream effect, and the ability to reverse the action. The comments also note that calibration requires real workflow data and that a single global threshold creates two possible extremes: excessive human review or unsafe automation.</p><p>I would structure this policy in two layers. The first contains mandatory blocking rules independent of model confidence. Identity mismatch, unknown units, missing provenance, unapproved terminology changes, actions involving high-risk medications, critical results without destination confirmation, modifications to clinically sensitive fields, and authentication failures all belong in this layer. The second contains graded signals such as completeness, mapping stability, historical frequency of the pattern, agreement across agents, documentation quality, distance from calibration cases, reversibility, available time, and scope of the action. The supervisor combines these signals according to an approved service policy. The output should be decomposed into verifiable reasons, with every risk factor and triggered rule visible to the reviewer.</p><p>Thresholds also vary by interface and action type. A structured laboratory message with a stable contract and near-real-time response requirements uses strict rules, deterministic validation, and specific interruption conditions. Extracting problems from a free-text clinical note tolerates more ambiguity and tends to run in the background, with review before anything is written to the chart. A billing task has different consequences from a change to an allergy, medication order, or device setting. A real-time workflow requires latency handling and contingency planning; a batch workflow allows accumulated review. The comments in the second PDF specifically call attention to interface-specific thresholds, placement of human approval, and the risk of placing the reviewer in the middle of the process where they receive fragments that do not yet form an intelligible decision.</p><p>The position of the human review stage depends on the use case, but there is a practical rule. When the agents are able to produce a complete proposal, the reviewer should receive the package after transformation, normalization, validation, and reconciliation, but before the final write. The reviewer can then examine the original message, transformed version, modified fields, selected terminology concept, unit conversion, validation results, identified conflicts, destination, and consequence of approval. In workflows with progressive decisions, an intermediate authorization stage can also make sense, provided the reviewer receives enough context and understands exactly what decision is being made. The diagram on page 2 places the human stage after the specialized agents have finished their work and before the downstream system, an ordering that supports automation of routine cases while preserving supervision of sensitive ones.</p><p>The expression human in the loop remains incomplete unless we define who reviews and what that person reviews. An HL7 syntax error belongs to the integration analyst. A local laboratory code without a valid mapping belongs to the terminology professional, with participation from the laboratory. An identity conflict belongs to registration and patient-safety teams. A discrepancy involving dose or allergy requires an appropriately qualified clinician. A billing rule goes to the revenue-cycle team. Sending every case to a physician creates delay, cost, and fatigue; sending a clinical case to an interface technician creates risk. The queue needs to classify the exception, assign responsibility, define a deadline, present the relevant evidence, and record the decision.</p><p>Communication between humans and the system deserves the same rigor as the integration itself. The review interface should display verifiable data: source message, linked patient and encounter, proposed transformation, modified fields, unit, reference range, terminology, triggered rule, mapping version, degree of uncertainty, destination system, and the effect of approval. The reviewer needs to understand both the operational decision and the clinical requirement involved. Operational explainability is based on evidence, rules, data, and versions rather than on an internally generated narrative from the model. The comments emphasize that if users cannot interpret or trust the output, the system fails as part of the workflow even when the technical component is functioning. Trust in this architecture comes from visibility, clear escalation, the ability to challenge a decision, and alignment between the task given to the human and that professional’s competence.</p><p>There is another important adjustment. The 80% to 90% automation range cited in the article describes the specific case presented. An institutional target should be derived from the risk profile and observed performance of each interface. A team that measures success only by the proportion of messages automatically written tends to raise risk tolerance inappropriately or conceal exceptions. I would measure automation rate together with incorrect-write rate, sensitivity of escalation for high-risk cases, reversal frequency, human-queue time, rejection rate, age of unresolved exceptions, and distribution of errors by clinical consequence. In a sensitive workflow, lower automation with reliable risk capture is better than higher automation with silent failures.</p><p>The deployment and security layer appears on page 4 and tends to receive little attention in demonstrations. The material describes agents running as independent services without permanent local state, with automatic scaling, separate identity and access accounts, an ordered messaging layer, encrypted storage, authentication, encryption in transit and at rest, immutable auditing, OpenTelemetry tracing, and anomaly detection. This layer supports system behavior when volume rises, a component becomes unavailable, an interface slows down, messages are repeated, a vendor changes its implementation, or a partial failure occurs.</p><p>A second technical aside is useful here. A stateless service processes a task without relying on local memory from a previous execution; relevant state is stored in controlled storage. This makes restart, load distribution, and instance replacement easier. Autoscaling means automatically adjusting the number of running instances according to demand. IAM is identity and access management; each agent receives its own service identity and minimum necessary permissions. An event bus is infrastructure that transports messages across components while preserving ordering, acknowledgment, and controlled replay. OpenTelemetry is a tracing standard that links logs, metrics, and execution paths. With it, the team can follow a message from source to destination and identify where delay, failure, or degradation occurred.</p><p>Separating agent identities prevents every component from having access to everything. The terminology agent can consult the clinical dictionary but does not need permission to write to the medical record. The conformance agent reads the structure and returns errors without authorization to alter registration data. The component that writes to the destination receives access limited to the authorized resource, field, and operation type. Database credentials remain outside the language model, and available commands remain restricted. The model calls tools through defined contracts, validated parameters, and an allowlist of actions. The same logic applies to information retrieval: dictionaries, vendor specifications, institutional rules, and terminology versions are provided as controlled sources with origin and date recorded.</p><p>The audit trail needs to be append-only, meaning that new records are added without overwriting history. Each stage should record the message identifier, source, timestamp, agent, tool, relevant input, structured output, rule, model version, prompt version, mapping version, decision, human reviewer, and result in the destination system. The record should support replay, reconstructing the sequence and, when authorized, reprocessing the event with duplication and version controls. Without this capability, a clinical or technical investigation ends in assumptions.</p><p>Rollback is the ability to undo or compensate for an automated action. In the medical record, that mechanism should preserve history through a new version, addendum, replacement marker, or compensating transaction while retaining what was written, who wrote it, when it happened, and why it was corrected. The architecture needs to define this mechanism before production begins. The article places rollback among the four minimum governance requirements, alongside a documented confidence threshold, a human stage for sensitive clinical actions, and an immutable and reproducible audit trail.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*2x4ODGwQKQV717tpijaz2g.png" /></figure><p>Once the architecture is designed, I would move to validation. The first test set should use representative historical messages, de-identified according to institutional policy, including normal, incomplete, duplicate, corrected, delayed, and incompatible examples. The team creates an expert-reviewed reference set, often called a golden dataset. For each message, that set specifies the correct mapping, expected decision, destination, exception type, and whether human review is required. Evaluation should be performed by field and by risk. High average accuracy can hide concentrated failure in identity, unit, allergy, or dose, precisely where the consequence may be greatest.</p><p>Next, I would run deliberate failure tests. The team changes a unit, removes a required field, alters a code, duplicates a message, shifts a timestamp, mixes a patient with the wrong encounter, stops an agent, resubmits an old event, degrades response time, and simulates downstream unavailability. The purpose is to determine whether the system blocks, holds, reroutes, logs, alerts, and recovers according to design. The supervisor should also be tested when an agent returns an incomplete result, when two agents disagree, and when a new dictionary version becomes active. A production system needs to fail in a controlled way, with a known destination for the message and without silently writing questionable content.</p><p>The next stage is shadow mode. In this mode, the system receives copies of the real workflow, executes every step, and produces decisions, but does not write to the medical record. The team compares the architecture’s output with the current process, measures discrepancies, reviews false blocks, identifies cases that escaped escalation, and calibrates thresholds per interface. The comments emphasize that this calibration depends on real workflow data and that human oversight should select the decisions that genuinely require judgment rather than reviewing every message.</p><p>During shadow mode, I would evaluate mapping accuracy by field, terminology compatibility, correct reconciliation rate, prevented duplication, proportion of cases routed to the correct professional, sensitivity for high-risk situations, human approval and rejection rates, reviewer disagreement, processing time, queue time, tool failures, lost messages, repeated messages, audit completeness, and replay success. The central safety metric is incorrect automated write rate stratified by severity. The central efficiency metric is total event-to-resolution time, including the human queue. The central governance metric is the ability to reconstruct every decision from data, rules, and versions.</p><p>When shadow mode meets the defined criteria, production entry occurs in stages. I would begin with a stable interface, a small set of message types, and lower-risk actions while maintaining expanded review. I would then release automatic writes for previously validated classes, with continuous monitoring and immediate suspension capability. Changes to mapping, terminology, vendor, model, prompt, rule, or destination system enter version control and repeat the relevant tests. Expansion follows risk class rather than enthusiasm for the technology.</p><p>Daily operation requires both clinical and technical observability. The team monitors volume, latency, failure rate, queue age, exception distribution, vendor-specific behavior, model version, mapping version, and review frequency. It also monitors changes in the incoming-data distribution, known as drift. A laboratory changes equipment, an EHR updates an interface, a department changes documentation practices, a new abbreviation appears, or the clinical case mix changes. Performance measured during the pilot is no longer valid when the input changes. Drift detection triggers contract review, recalibration, and, when necessary, temporary return to human review.</p><p>The incident plan should define who can stop automation, how to isolate an interface, how to preserve messages, how to notify clinical teams, how to identify affected records, how to execute corrections, how to reconcile downstream systems, and how to document the event. If an incorrect unit version reaches the medical record, the response includes locating all messages processed under that version, identifying which were written, issuing traceable corrections, reviewing the highest-risk cases, and blocking further writes until validation is complete. Audit, replay, and rollback turn a diffuse failure into a bounded set of traceable events.</p><p>Governance also requires clear ownership. The clinical owner defines consequences, escalation criteria, and acceptability. The integration owner maintains contracts and connectors. The terminology owner manages dictionaries and versions. Security manages identity, access, and incidents. Operations manages queues, availability, and response times. The human reviewer operates within a defined scope of competence. A change-control group approves modifications that affect clinical meaning, automation scope, or supervisor behavior. Without this allocation of ownership, the technology enters production with implicit accountability and ends with decisions that have no identified owner.</p><p>The material reports that, in the case presented, the architecture reduced configuration time for new interfaces from weeks to days, replaced part of manual mapping with AI-assisted mapping and automated conformance validation, and kept changes affecting clinical data behind a human approval gate. That result belongs to one specific implementation; its value in another institution depends on data quality, workflow repetition, volume, interface-contract maturity, and governance discipline.</p><p>Production readiness requires bounded execution, correct data, verifiable rules, least-privilege permissions, risk-proportionate review, complete logging, traceable correction, and predictable behavior during failure. The number of agents and the fluency of their outputs are secondary to those requirements. For every action, the institution needs to answer where the data came from, what transformations occurred, which rule was applied, which version was active, why the case proceeded automatically or was escalated, who reviewed it, what reached the destination system, and how the result can be corrected. The comments summarize this distinction by placing thresholds, human stages, audit trails, and rollback mechanisms around the agents as the elements that sustain real-world operation.</p><p>This is the path from pilot to production: begin with a defined clinical workflow, formalize data contracts and meaning, separate deterministic from semantic functions, distribute responsibilities across specialized agents, constrain the supervisor, design a context-sensitive risk policy, give the reviewer a complete and verifiable decision package, implement auditing, observability, and correction before the first real write, validate against historical data, operate in shadow mode, release by risk class, and recalibrate continuously. Agentic AI enters healthcare when this chain functions as part of clinical care and hospital operations, with identifiable accountability at every stage.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=de83c2f838f9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Healthcare and AI Integration]]></title>
            <link>https://abacelar.medium.com/healthcare-and-ai-integration-cd7100e79c16?source=rss-7a8f1e8140cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/cd7100e79c16</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[healthcare-technology]]></category>
            <dc:creator><![CDATA[Albert Bacelar]]></dc:creator>
            <pubDate>Fri, 28 Aug 2026 17:49:54 GMT</pubDate>
            <atom:updated>2026-08-28T17:49:54.959Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RdEyPF0zC0KamBPbgdi9Pw.png" /></figure><p>Virtually every healthcare organization today is discussing artificial intelligence. The problem is that interest in the technology often emerges before the clinical or operational problem it is supposed to solve has been clearly defined. Leaders attend demonstrations, see tools that appear to perform very well in controlled environments, and quickly begin discussing procurement, integration, and implementation. Yet when the discussion turns to which clinical outcome should change, in which patient population, and at what point in the care pathway, there is often no clear answer.</p><p>This is one of the main reasons hospitals accumulate pilot projects that never become part of routine clinical practice. Before selecting a tool, the intended intervention needs to be defined precisely. Are we trying to shorten the time between admission and sepsis recognition? Identify patients at risk of clinical deterioration earlier? Reduce prescribing errors? Automate medical documentation? Prioritize examinations that require immediate review? Reduce physicians’ administrative workload? Each problem requires different data, metrics, accountable teams, and validation methods. Without this definition, virtually any commercial demonstration appears attractive, and multiple projects begin competing with one another without a clear clinical priority.</p><p>The next step comes before purchasing the technology and must involve the professionals who actually deliver care. Physicians, nurses, pharmacists, and other healthcare professionals need to assess how the system will fit into the clinical workflow. If an algorithm indicates clinical deterioration and the physician disagrees, what should be documented? If the system fails to identify a patient who later deteriorates, will that error be detected retrospectively? Who receives the alert, how quickly must it be assessed, and what happens when dozens of alerts are generated simultaneously? If the recommendation is incorporated into the medical record, will the system output, the clinician’s decision, and any disagreement between the two be documented? These definitions are as much a part of implementation as the technical integration of the software.</p><p>It is also necessary to assess whether the institution actually has the capacity to operate the solution. Having an electronic health record does not mean having data that are suitable for artificial intelligence. The data need to be available, structured when required, temporally consistent, and correctly linked to the patient and the relevant episode of care. A clinical deterioration model, for example, loses part of its usefulness if vital signs reach the system with delays, laboratory results are recorded using inconsistent units, or important information remains available only as free text. Before discussing model performance, the institution needs to understand the quality of the data that will actually reach the model during routine care.</p><p>The same applies to governance. The institution needs to define who authorizes use of the system, who monitors its performance, who investigates incidents, who has the authority to suspend its use, and who is accountable for subsequent changes made by the vendor. The organization also needs the technical capacity to integrate the tool with the electronic health record, laboratory systems, radiology, bedside monitors, or any other required data sources. Purchasing a solution before mapping these dependencies simply shifts the problem to the implementation phase, when contracts have already been signed and the cost of changing course is higher.</p><p>Clinical evaluation also does not end when the vendor presents accuracy, sensitivity, specificity, or AUROC values obtained during development. A model trained and tested in another population needs to be evaluated in the context in which it will actually be used. Epidemiological profile, disease prevalence, admission criteria, availability of diagnostic tests, demographic characteristics, and local clinical practices all affect performance. An algorithm that shows good discrimination in one hospital will not necessarily behave the same way in another institution. Local validation therefore needs to assess discrimination, calibration, false positives, false negatives, and, most importantly, the clinical consequences of those errors.</p><p>There is also an important distinction between validating a model and demonstrating clinical benefit. A system may predict a particular event with good statistical performance and still fail to produce a meaningful improvement in care. If an alert arrives too late, reaches the wrong person, generates excessive notifications, or identifies something the clinical team would already have recognized without computational support, its mathematical performance will have limited clinical value. Implementation therefore needs to evaluate the effect of the system on the care process and on the outcome that originally justified its adoption.</p><p>Regulation needs to be considered from the project design stage. In the European context, certain medical applications of artificial intelligence fall under high-risk regulatory regimes depending on their intended purpose and their relationship with medical device regulation. This requires attention to documentation, risk management, performance, human oversight, and post-deployment monitoring. The same clinical principle remains valid regardless of jurisdiction: the institution needs to continue monitoring the system after implementation. Sensitivity, specificity, alert rates, error patterns, and performance across different patient groups need to be followed over time because populations, protocols, equipment, information systems, and clinical practices change.</p><p>Another determining factor is clinical team adoption. A system may be technically sound, well integrated, and compliant with regulatory requirements, yet have little effect if clinicians do not trust its recommendations or understand its limitations. That trust is not created by a training session during the week of implementation. It begins when clinicians participate in selecting the problem, defining the workflow, determining which errors are acceptable and unacceptable, validating the system, and establishing its rules of use. Physicians need to know when to consider the output, when to question it, which data influenced the analysis, and how to proceed when their clinical assessment conflicts with the information generated by the system.</p><p>The patient is also part of this design. The institution needs to determine what information generated by artificial intelligence will be presented to the patient, how the system’s role in care will be explained, and which decisions remain under professional responsibility. Depending on the application, this discussion may also involve consent, transparency, secondary use of data, traceability, and documentation in the medical record.</p><p>For these reasons, institutional adoption of artificial intelligence needs to be managed as a change in the care delivery process. The work begins with the definition of a measurable clinical problem, followed by mapping the current workflow, assessing available data, analyzing risks, selecting the solution, validating it in the local population, integrating it into the team’s work, defining responsibilities, conducting controlled implementation, and establishing continuous monitoring. Technology represents only one part of this process.</p><p>Many organizations still focus most of their attention on the technology demonstration itself. The institutions that succeed in turning artificial intelligence into routine clinical practice will be those that structure everything that comes after the demonstration: data, clinical validation, workflow integration, governance, accountability, monitoring, and the participation of the professionals who will ultimately care for patients using that technology.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=cd7100e79c16" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How to Build AI Agents for Clinical Practice]]></title>
            <link>https://abacelar.medium.com/how-to-build-ai-agents-for-clinical-practice-e51e1317c5f6?source=rss-7a8f1e8140cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/e51e1317c5f6</guid>
            <category><![CDATA[healthcare-technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Albert Bacelar]]></dc:creator>
            <pubDate>Thu, 13 Aug 2026 16:01:04 GMT</pubDate>
            <atom:updated>2026-08-13T16:01:04.346Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Dif8yeip95fd6u2gotmNVg.png" /></figure><p>When I look at this Clinical Harness architecture, what interests me most is not the LLM. It is precisely the fact that the LLM has lost the central position it usually occupies when someone presents an artificial intelligence solution for healthcare. It is still there, of course, but encapsulated within an architecture that controls context, capabilities, execution, memory, tools, integrations, policies, observability, and efficiency. This shift is relevant because it brings the discussion of clinical agents closer to what actually happens when we move from a demonstration to a system operating within a care environment. A clinical agent is not sustained by the isolated quality of the model. It depends on the quality of the orchestration around that model. And that is exactly what this concept of a harness is trying to organize.</p><p>I would start with the two ends of the architecture, because they help us understand everything else. On the left are the inputs: healthcare professionals, patients, electronic health records, exams and reports, guidelines, indicators, schedules, and other sources. On the right are the outputs: clinical response, decision support, task creation, alerts, insights, updates. It looks like a simple representation, but there is an important architectural issue here. Between receiving clinical data and producing an action, there is a processing chain that needs to be controlled. When a healthcare professional asks an isolated chatbot a question, we have a relatively bounded interaction. When an agent receives events from the EHR, consults other sources, retrieves knowledge, executes tools, maintains state, decides on the next action, and eventually writes something back into a system, the unit of analysis is no longer the textual response. It becomes the entire execution.</p><p>That is why the harness appears at the center. I understand this harness as an orchestration layer. It determines how the different components participate in that execution, in what sequence, under which instructions, with which policies, and within which limits. At the top of the diagram, we see instructions, flow, state, routing, and policies. This matters because a clinical agent should not depend on a single large prompt attempting to handle identity, knowledge, clinical rules, memory, system access, security, and tool decisions simultaneously. When all of this is mixed together, it becomes difficult to test, version, audit, and identify the origin of a specific behavior. An orchestration architecture separates these responsibilities.</p><p>The first internal layer presented is context, where clinical RAG appears alongside knowledge graphs, patient data, the electronic health record, guidelines, protocols, and task state. This deserves attention because clinical RAG is frequently reduced to semantic search across documents. In a clinical application, that is insufficient. Relevant context is not composed merely of the section of a guideline that happens to be semantically close to the question. There is longitudinal, temporal, institutional, physiological, pharmacological, and operational context. An agent analyzing an inpatient needs to distinguish historical information from active information, recognize trends rather than isolated values, understand relationships among diagnoses, medications, procedures, and tests, and retrieve the appropriate version of a guideline or protocol. If I do not control how this context is assembled, I can have an excellent model reasoning over an inadequate representation of the case.</p><p>This is particularly important in healthcare because the electronic health record was primarily designed to document care, support communication, record clinical acts, and meet administrative and legal requirements. It does not spontaneously provide a perfect computational representation of the patient’s clinical state. There is duplication, copy-forward, conflicting information, missing data, records generated at different points in time, and clinically relevant information distributed across structured fields and free text. Therefore, when this architecture refers to clinical RAG, patient data, knowledge graphs, and task state, we are dealing with a problem much broader than document retrieval. We are dealing with context engineering applied to the clinical domain: selecting what enters the context window, establishing precedence among sources, preserving temporality, representing relationships, and preventing documentary noise from receiving the same weight as current, validated data.</p><p>Then we reach intelligence itself, where the LLM appears alongside deterministic models, scores, rules, calculations, and clinical predictions. I consider this separation technically important. There is a tendency to delegate to generative models operations that already have deterministic solutions or specific models that have been appropriately evaluated. There is no architectural reason to ask an LLM to improvise a calculation whose formula is known. Likewise, a predictive model validated for a particular population does not need to be replaced by linguistic inference. A more consistent architecture combines different forms of intelligence. The LLM interprets language, organizes unstructured information, selects tools, synthesizes results, and participates in reasoning. A deterministic function calculates. A specific classifier classifies. A predictive model estimates risk. An institutional rule verifies a criterion. The agent becomes a composition of these capabilities.</p><p>This also improves evaluation. If a numerical result came from a deterministic function, I can test that function independently. If a classification came from a specific model, I can monitor the performance and calibration of that model. If a synthesis was produced by the LLM, I apply another set of metrics. When everything is delegated to the generative model, different types of errors become mixed within the same output. In a clinical system, separating components also means separating validation surfaces.</p><p>The third column of the diagram concerns capabilities, or skills. This concept is more interesting than simply giving the model access to tools. A tool is an available operation. A skill represents an organized operational capability that can combine instructions, logic, tools, rules, and termination criteria. A clinical agent might have a specific skill for medication reconciliation, another for clinical deterioration analysis, another for preparing a shift handover, and another for assessing eligibility for a given protocol. Each capability can be tested, versioned, and authorized separately. This avoids building a universal clinical agent with unrestricted access to everything and excessively open-ended behavior.</p><p>There is also an important governance consequence. Once I can decompose an agent into capabilities, I can establish authorization at the capability level. An agent may be authorized to consult laboratory results and produce a synthesis, but not to generate an examination order. Another may prepare a task for human validation without executing it directly. A third may operate exclusively within an administrative activity. Autonomy therefore ceases to be a binary property of the agent and becomes something defined according to action, context, user, risk, and authorization level. This is much more compatible with clinical reality.</p><p>The fourth column is execution. The diagram uses the loop observe, think, act, verify, and repeat. This is where we effectively enter agentic behavior. The system receives an objective and does not necessarily know in advance the entire sequence of operations required to achieve the result. It observes the available state, selects the next action, executes that action, receives a new state, verifies the result, and decides whether to continue. This capability is useful precisely because many clinical and operational processes are not rigid sequences. The next step depends on the result of the previous one.</p><p>But this is also where the architecture becomes more sensitive. In a multi-step execution, risk is not restricted to the final generation. An error in context retrieval can alter tool selection; a tool may return incomplete data; that data may modify the state; the agent may then choose another action that is coherent with that incorrect state. In the end, the response may appear internally consistent even though the trajectory has been compromised since an earlier stage. That is why the “verify” component of this sequence deserves far more attention than it usually receives. Verification should not mean asking the same model to read its own response and determine whether it agrees with itself. Depending on the risk, verification needs to rely on external criteria, deterministic rules, schema validation, consistency checks, source confirmation, another independent component, or human intervention.</p><p>The fifth column is memory, divided into short-term and long-term memory, with feedback, recall, and governance. This distinction is also essential. An agent needs to maintain state in order to work throughout an execution. If it has just consulted a test result or executed a tool, that result needs to remain available for the next step. That is short-term operational memory. A different question is whether certain information should persist across future executions. In healthcare, persistence cannot be treated merely as a technical convenience. Clinical information already has official sources and its own rules for documentation. If the agent creates a parallel memory without defined provenance, validity, updating, and expiration, a second representation of the patient emerges and may diverge from the medical record.</p><p>Long-term memory therefore requires explicit governance. The system needs to know where the information came from, when it was recorded, whether it remains valid, what authority it carries, and in which contexts it may be retrieved. We also need to separate patient memory, interaction memory, and system learning. An operational preference observed during an interaction should not automatically acquire the status of clinical knowledge. This problem tends to grow as persistent agents begin following processes over days, weeks, or months.</p><p>Below these five layers are the tools, and beneath them the connectors, including MCP and APIs. This part of the architecture represents a concrete shift from the earliest uses of generative AI. The model stops working exclusively with what was placed in the prompt and begins accessing external resources. It can search, consult, write, send, execute, or trigger other functions, depending on the permissions granted. The connectors bridge the agent with the EHR, LIS, RIS/PACS, pharmacy, laboratories, devices, and other systems. This is where agent architecture meets the actual infrastructure of health informatics.</p><p>MCP appears here as a way to standardize how tools and resources are exposed to the agent, but I would avoid treating MCP as a solution for clinical interoperability. These are different problems. FHIR, HL7, DICOM, clinical terminologies, and system APIs continue to address aspects of information representation and exchange within the healthcare ecosystem. MCP operates at another layer, providing an interface through which model-based applications can identify and use resources made available by compatible servers. A hospital could, for example, expose certain operations through an MCP server while those operations internally communicate with FHIR APIs, proprietary services, or other components. This distinction matters because an agent integration layer should not be conceptually mistaken for a replacement for healthcare interoperability.</p><p>And the more tools we introduce into this architecture, the more relevant the principle of least privilege becomes. An agent does not need full access to the medical record simply because a particular skill uses three resources. A consultation tool does not necessarily need to offer write access. A write operation does not necessarily need to be executed without confirmation. Permissions should follow the granularity of capabilities. This reduces the risk surface and improves accountability and traceability.</p><p>The diagram places guardrails across the architecture and includes clinical policies, security and access, privacy, ethics, and compliance. I would make an important distinction here: a clinical guardrail is not synonymous with a safety prompt. If an action is prohibited, the most reliable protection belongs in the authorization or execution layer, rather than in a textual instruction that expects the model to comply with the rule. If a particular agent is not authorized to modify a prescription, the corresponding tool should not be available in that context. If a particular action requires human validation, the workflow needs to block further execution until that validation occurs. If certain data cannot leave a specific domain, that needs to be controlled by the infrastructure. In clinical systems, important policies need to be translated into technically verifiable controls whenever feasible.</p><p>Next to guardrails we have observability, and I would say this is one of the elements that most clearly separates an interesting prototype from an architecture intended for serious healthcare operation. Logs and traces, decision tracking, usage and quality metrics, cost, performance, and outcome evaluation are explicitly included in the diagram. This means storing the final response is not enough. We need to be able to reconstruct an execution. Which version of the agent was active? Which model was called? What context was retrieved? Which tools were triggered? What results did they return? Which policy was in force? Was there an error? Was there a retry? Was there human intervention? What output ultimately reached the healthcare professional?</p><p>This introduces an idea that I consider central to the auditing of clinical agents: the auditable unit becomes the trajectory. If I evaluate only the final response, I lose precisely what characterizes an agent, which is the sequence of decisions and actions that produced that response. Two agents may arrive at the same recommendation through completely different trajectories. One retrieved the correct information, executed a validated function, verified the result, and presented the source. The other reached the same result through an unstable sequence. If I observe only the output, I classify both as correct. If I observe the trajectory, I can distinguish reliability from coincidence.</p><p>This observability also changes the concept of evaluation. Evaluating a clinical agent is not a matter of applying a set of medical questions and measuring how many answers are correct. That evaluates one part of the model, not the agentic system. I need to evaluate context retrieval, tool selection, call parameterization, interpretation of returned results, compliance with policies, the decision to continue or terminate the loop, output quality, latency, failure rate, and behavior when required information is missing. Depending on the application, sensitivity, specificity, calibration, alert rates, clinician acceptance, overrides, time saved, and effects on process or clinical outcomes also become relevant. Evaluation needs to follow the architecture.</p><p>At the bottom of the diagram is efficiency: token optimization, context caching, latency, cost, and model routing. This may appear to be an engineering detail, but at hospital scale it stops being a detail. An agent that evaluates hundreds or thousands of patients, reacts to events, or repeatedly performs tasks needs to control computational consumption. Model routing is particularly interesting because it dismantles the idea that a single model is responsible for the agent. Certain tasks can be handled by smaller models, others require a model with greater reasoning capacity, and still others do not require an LLM at all. Selection can occur dynamically according to complexity, risk, acceptable latency, and cost. The agent becomes a multi-model, multi-tool architecture.</p><p>If I bring all of these layers together, what this diagram is showing is a change in the unit of design. During the first phase of generative AI, many people designed prompts. Then we began designing applications with RAG. Now, with agents, we need to design execution systems. That requires thinking about state, context, tools, permissions, memory, loops, policies, telemetry, evaluation, and integration from the outset. In healthcare, this takes on another dimension because each of these architectural decisions can interfere with a care process.</p><p>There is another aspect I would add to this architecture: identity and provenance need to be handled very carefully. In a clinical environment, knowing what was produced is not enough. We need to know on whose behalf a given operation was performed, which agent executed it, which healthcare professional was associated with the session, which source originated each piece of data, and which component transformed that information. When a system begins writing back into the medical record, creating tasks, or generating alerts, identity stops being merely an authentication detail and becomes part of the semantics of the action itself.</p><p>Another point I would add is versioning. Agents change. The prompt changes, the model changes, the policy changes, a skill changes, a tool changes, the institutional protocol changes, the knowledge base changes. Therefore, saying that “Agent X produced this recommendation” is insufficient. To technically reconstruct that execution, we need to know the exact configuration that was active at that moment. In a clinical environment, this brings agent operation closer to practices we already know from configuration management, quality systems, and controlled change processes.</p><p>That is why I find this idea of a harness useful. It shifts attention away from fascination with the model’s linguistic capability and toward the engineering required to turn that capability into a component of a clinical system. The LLM remains important, but it is governed by an architecture that determines what it receives, what it is able to do, what it must record, and what it is prevented from doing.</p><p>If I were evaluating a clinical agent project for a hospital today, I would not begin by asking which model is being used. I would ask to see this architecture. I want to know how context is assembled, how sources are prioritized, which skills exist, which tools each skill accesses, where deterministic execution occurs, how state is maintained, what persists in memory, how permissions work, where human-in-the-loop is required, which events are logged, and how an execution can subsequently be reproduced. Only after that would it make sense to discuss which model delivers the best performance within that architecture.</p><p>For me, that is the most interesting point in this Clinical Harness. The discussion around clinical agents matures when we stop treating the model as the final product and begin treating it as a component embedded within an operational architecture. Context, intelligence, capabilities, execution, and memory form the core; tools and connectors connect that core to the real environment; guardrails define the boundaries of operation; observability enables monitoring and auditing; efficiency makes operation sustainable at scale. The quality of the agent depends on the interaction among all of these layers.</p><p>And this also changes the question that hospitals, physicians, and managers should ask when evaluating an agentic solution. Knowing that it uses a particular LLM tells us very little about how it will behave within a clinical process. The more technically relevant question is to understand the harness: what mechanisms exist between data input and action production, which components are probabilistic, which are deterministic, which are auditable, where the limits of autonomy are, and how the system behaves when one of its stages fails. It is at this orchestration layer that an AI demonstration begins to become clinical agent engineering.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e51e1317c5f6" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Where Clinical AI Fails]]></title>
            <link>https://abacelar.medium.com/where-clinical-ai-fails-f4d273a54414?source=rss-7a8f1e8140cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/f4d273a54414</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[healthcare-technology]]></category>
            <dc:creator><![CDATA[Albert Bacelar]]></dc:creator>
            <pubDate>Thu, 13 Aug 2026 15:53:09 GMT</pubDate>
            <atom:updated>2026-08-13T15:53:09.439Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tCmIH7Vf99seUSR4i3hw8Q.png" /></figure><p>The evaluation of artificial intelligence in healthcare still suffers from a methodological problem: models are often tested under conditions that are far more organized than clinical practice itself. They receive clearly bounded questions, curated datasets, explicit tasks, and frequently a predefined set of acceptable answers. Performance under these conditions tells us something about whether a model recognizes medical patterns within a controlled representation of a problem, but far less about how that same system behaves when confronted with fragmented medical records, temporally conflicting information, third party data, undocumented events, guidelines validated for specific populations, or requests in which a clinical decision is embedded inside an apparently administrative task. Clinical red teaming is designed to examine precisely this gap between demonstrated competence under structured evaluation and behavior under realistic clinical ambiguity. Its purpose is not to create artificially difficult questions or to force a model into failure. It is to construct clinically plausible situations in which an inappropriate response has an identifiable consequence, then determine whether those failures form reproducible patterns.</p><p>The most informative tests are rarely built around rare diseases or unusually complex diagnostic puzzles. In many cases, the difficulty lies in something far more ordinary: information that belongs to a relative rather than the patient, an event that would be expected but was never documented, a correct guideline applied outside the population in which it was validated, a risk that becomes apparent only when information scattered across several parts of the record is connected, or the same underlying clinical decision presented through different linguistic formulations. These scenarios expose limitations that average accuracy metrics are poorly equipped to detect. A model may process every isolated component correctly while failing at the point where clinical judgment requires an understanding of provenance, timing, applicability, uncertainty, and relationships among pieces of information. The failure is therefore not necessarily a lack of medical knowledge. In many cases the knowledge is present. The problem lies in determining when, where, and under which conditions that knowledge should govern the response.</p><h3>Privacy as a problem of contextual reasoning</h3><p>One recurrent failure pattern appears when a model receives information about a patient and information about third parties within the same context. A medical record may contain detailed family history, information disclosed by relatives, external test results, communications with caregivers, or material that is clinically relevant to understanding the case but not necessarily appropriate for inclusion in every downstream document. If the model is asked to prepare a referral letter, discharge summary, medical report, or communication to another institution, it may treat every available piece of information as legitimate source material. The failure occurs before any obvious disclosure: the system does not consistently determine who owns the information, whether disclosure is authorized, whether its use is compatible with the purpose of the document, or whether part of the available context should remain excluded from the output.</p><p>This matters because privacy in clinical AI cannot be reduced to removing names, identifiers, or other direct personal information. The system must reason about <strong>information provenance and permission to use it</strong>. Access to a fact does not establish permission to reproduce it. A useful red teaming scenario therefore does not ask the model whether confidentiality is important. Such a question tests conceptual knowledge. A stronger test embeds sensitive third party information inside an otherwise routine clinical task and observes whether the system identifies the restriction without being prompted to do so. That difference is methodologically important because many real privacy failures do not begin with an explicit request to disclose confidential information. They arise when a legitimate task is performed without distinguishing between information that is available and information that is appropriate for the intended purpose.</p><h3>Fabricated clinical events and the conversion of plausibility into documentation</h3><p>Another recurrent pattern appears when a model fills documentary gaps with events that would be plausible within the clinical sequence. A record contains the indication for a procedure, preprocedural evaluation, and subsequent evolution, but no documented evidence of consent. When generating a summary, the model states that consent was obtained. Another record contains a recommendation for specialist referral, but no evidence that the referral occurred; the generated text describes it as completed. Similar failures occur when a family discussion is reported as having taken place, a prescribed medication is described as administered, an examination is treated as performed, or a planned intervention is converted into a completed event.</p><p>This class of failure deserves particular attention because the fabricated event often fits the clinical narrative perfectly. It does not look absurd. It looks expected. That is precisely what makes it difficult to detect during human review. In clinical documentation, however, the difference between “planned,” “likely,” “recommended,” and “documented as completed” carries clinical, legal, operational, and regulatory consequences. Clinical red teaming should therefore include cases in which an event is highly plausible but deliberately absent from the source material, then assess whether the system preserves that absence or silently converts probability into fact. The distinction is fundamental: linguistic coherence is not documentary evidence. A system producing clinical text must therefore be evaluated not only on whether the statements it makes are medically sensible, but also on whether it preserves the epistemic status of each piece of information.</p><p>The same problem extends beyond explicit fabrication. Models may collapse uncertainty by replacing “possible,” “suspected,” or “under investigation” with definitive diagnostic language. They may merge treatment intention with treatment completion or convert a provisional assessment into a final diagnosis because later sections of the record appear compatible with it. These transformations are subtle because they improve narrative continuity while degrading evidentiary fidelity. Clinical evaluation therefore needs to examine whether the model preserves distinctions between observation, interpretation, intention, recommendation, execution, and confirmation.</p><h3>The right guideline applied to the wrong population</h3><p>Clinical models often demonstrate substantial knowledge of guidelines, thresholds, risk scores, and treatment protocols. Yet this creates another failure mode: a rule may be recalled correctly and still be applied incorrectly because the patient does not belong to the population for which that rule was validated. A threshold established for one disease is transferred to another condition. An adult protocol is applied to a pediatric patient. A prediction model derived from stable outpatients is used in critically ill patients. A recommendation from a particular population is applied to patients who would have been excluded from the underlying studies.</p><p>The output may contain the correct number, the correct score, the correct recommendation, and even the correct source. The error lies in the link between evidence and patient. Traditional evaluations tend to reward retrieval of the rule itself. If a question asks for the cutoff associated with a score, identifying that cutoff is considered successful performance. Clinical red teaming changes the unit of evaluation. It constructs a case in which that same cutoff appears near a population or clinical state in which it should not govern the decision, then measures whether the system recognizes the incompatibility.</p><p>This is a test of <strong>applicability</strong>, not recall. Medical evidence is bounded by study population, clinical setting, inclusion criteria, exclusion criteria, intervention characteristics, endpoint definitions, and intended use. A model that retrieves a recommendation without identifying its domain of validity may produce an answer that appears technically precise while being clinically inappropriate. These are among the most difficult failures to identify because surface level factual checking may confirm nearly every component of the response. The numerical value is correct; the clinical context in which it is being used is not.</p><p>For this reason, red teaming should deliberately manipulate population characteristics while preserving other elements of the case. Age, pregnancy, renal function, disease severity, comorbidity, care environment, immunological status, procedural context, and prior treatment can each determine whether a rule remains applicable. A model that continues to produce the same recommendation across clinically meaningful changes in these variables is demonstrating a failure of contextual validity even when its factual knowledge remains intact.</p><h3>Fragmented information and failures of clinical synthesis</h3><p>Real medical records rarely present risk in a single sentence. Clinically relevant information is distributed across physician notes, nursing documentation, medication lists, laboratory results, imaging, prior admissions, specialist recommendations, and records generated at different points in time. A model may interpret every individual component correctly while failing to integrate their combined meaning. A drug appears reasonable in isolation. A laboratory abnormality appears modest. A comorbidity alone does not contraindicate treatment. Yet the interaction among deteriorating renal function, current dose, patient age, concomitant medications, and laboratory trajectory substantially changes the risk profile.</p><p>This type of failure exposes an important difference between local comprehension and clinical synthesis. If all relevant information is placed in the same paragraph and the relationship among variables is made explicit, a model may perform well. The more demanding test distributes those variables across the record and requires the system to construct the relationship itself. Clinical red teaming should therefore reproduce the information architecture of actual clinical documentation rather than presenting artificially compressed cases. The question is no longer whether the model understood each datum independently. The question is whether it identified which data needed to be connected in order to recognize the clinically relevant state.</p><p>Temporal structure makes the problem even more difficult. A creatinine value from three days earlier does not carry the same meaning as the current value. A medication listed on admission may subsequently have been discontinued. A diagnostic hypothesis documented at presentation may have been rejected after further investigation. A recommendation made before a procedure may no longer be valid afterward. When models treat all information in the context window as equally current, they flatten the longitudinal structure of the case and create false combinations of data that never actually coexisted.</p><p>Clinical red teaming should therefore include temporal conflicts deliberately. Older values should remain present alongside newer ones. Historical medications should coexist with current prescriptions. Preliminary diagnoses should remain visible after later revision. The evaluation then asks whether the model identifies which information has been superseded, which remains active, which is temporally dependent, and which requires confirmation before being used in a decision. In longitudinal clinical care, chronology is not metadata. It is part of the clinical meaning of the data.</p><h3>When the wording changes the clinical answer</h3><p>One of the most consequential findings in clinical red teaming is that a model’s behavior may change according to how the same underlying decision is framed. When asked directly, “Is this treatment safe?”, the model may search for contraindications, mention uncertainty, identify missing information, and recommend caution. Present the same clinical situation as “calculate the dose,” “write the prescription,” “prepare the discharge plan,” or “draft the referral letter,” and some of those safeguards may disappear. The clinical decision has not materially changed. The linguistic wrapper around the decision has.</p><p>This exposes a form of instability that conventional evaluation rarely measures. A clinical system should preserve relevant constraints across semantically equivalent tasks. If a contraindication matters when safety is queried explicitly, it should also constrain an operational request that would enact the same decision. Otherwise, the quality of the clinical safeguard becomes dependent on whether the user happens to formulate the request in a way that triggers it.</p><p>Red teaming can test this systematically by constructing families of prompts that contain the same clinical facts and the same underlying decision but express the task differently. The outcome of interest is not textual consistency. It is <strong>clinical invariance</strong>. Material safety conclusions should remain stable when the surface form of the request changes.</p><p>The same logic applies to privacy, evidence applicability, medication safety, documentation, and diagnostic uncertainty. Asking “Is there a privacy concern here?” tests whether the model knows about privacy. Asking the system to create a document containing third party data without mentioning privacy tests whether that knowledge influences its behavior spontaneously. Asking whether a score was validated in a given population tests methodological knowledge. Embedding that score in an incompatible population without announcing the mismatch tests whether the system detects the problem on its own. The second type of evaluation is much closer to the conditions under which these systems will be used.</p><h3>Clinical red teaming as a method for identifying systematic failure</h3><p>The value of clinical red teaming lies less in accumulating individual errors than in identifying <strong>reproducible classes of failure</strong>. A single incorrect response may reflect noise, ambiguity, or an isolated model behavior. A repeated pattern in which the system discloses third party information during documentation tasks, converts planned events into completed ones, applies validated rules outside their intended population, misses risks distributed across the medical record, or loses safety constraints when the phrasing changes represents something different. It identifies a behavioral property of the system.</p><p>Once a failure class has been identified, evaluation becomes more structured. Cases can be varied systematically. The same privacy problem can be tested across discharge summaries, referrals, insurance reports, research documentation, and interhospital communication. The same applicability problem can be tested by modifying age, disease state, care setting, or exclusion criteria. The same synthesis problem can be tested by changing the distance between relevant pieces of information in the record, the number of irrelevant distractors, or the temporal relationship between observations. The objective is to determine the conditions under which the failure appears, disappears, or becomes more frequent.</p><p>This methodology also requires defining the clinical consequence before the model is tested. It is insufficient to state that the output differed from a reference answer. The evaluator needs to establish what information should have constrained the response, what inference exceeded the available evidence, which clinical rule was misapplied, and what consequence would follow if the output were accepted. That is where clinical expertise becomes central to the design of the evaluation. The difficult part is frequently not producing an adversarial prompt. It is determining whether the failure is clinically meaningful and whether the scenario represents a plausible mode of use.</p><p>For the same reason, the strongest clinical red teaming cases are often mundane rather than spectacular. Summarizing a chart, preparing a discharge document, calculating a dose, reconciling medications, reviewing laboratory trends, selecting a protocol, or preparing a referral are ordinary clinical activities. The adversarial element lies in the arrangement of information, the missing documentation, the conflicting chronology, the population mismatch, or the hidden safety constraint. These are precisely the conditions under which systems that perform well on medical examinations may behave unpredictably in practice.</p><h3>Where clinical AI actually fails</h3><p>The central question is therefore not whether an AI model “knows medicine.” Modern systems increasingly demonstrate that they contain substantial medical knowledge. The more difficult question is whether they know <strong>which information is usable, which fact is documented, which inference remains uncertain, which recommendation applies to this patient, which observations must be combined, and which constraints remain valid when the user’s request changes form</strong>.</p><p>Privacy requires the model to distinguish access from authorization. Documentation requires it to distinguish evidence from plausibility. Evidence based medicine requires it to distinguish knowledge of a recommendation from knowledge of its domain of validity. Clinical reasoning requires it to integrate distributed and temporally dependent information. Safety requires these distinctions to survive changes in task formulation.</p><p>As models become more fluent and clinically knowledgeable, the most consequential failures are increasingly unlikely to look absurd. They are more likely to appear inside well written, coherent, technically persuasive outputs in which one unsupported assumption, one misplaced rule, one missed temporal relationship, or one improperly reused piece of information changes the clinical meaning of the answer. Clinical red teaming is designed to identify those conditions before they are encountered in patient care, reproduce them under controlled evaluation, determine whether they represent systematic behavior, and convert each recurring pattern into a testable requirement for clinical AI systems.</p><p>The relevant measure of a clinical model is therefore not merely how often it produces the correct answer when the problem has already been organized for it. The harder and more clinically meaningful question is <strong>under which conditions does the system fail to recognize what an attentive clinician would identify as unsupported, inapplicable, improperly disclosed, internally inconsistent, or unsafe?</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=f4d273a54414" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Invisible Risk Inside the Hospital]]></title>
            <link>https://abacelar.medium.com/the-invisible-risk-inside-the-hospital-8a6423df6de2?source=rss-7a8f1e8140cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/8a6423df6de2</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[dr2-thinktech]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[healthtech]]></category>
            <dc:creator><![CDATA[Albert Bacelar]]></dc:creator>
            <pubDate>Thu, 06 Aug 2026 22:19:56 GMT</pubDate>
            <atom:updated>2026-08-06T22:19:56.040Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Dr2 thinktech — artificial intelligence" src="https://cdn-images-1.medium.com/max/1024/1*pzP9IstY867k5nGBaxDdMw.png" /><figcaption>Dr2 thinktech — artificial intelligence</figcaption></figure><p>Your hospital already has enough technology to see more clearly than it sees today.</p><p>The discomfort starts there.</p><p>Because when hospital leadership says it needs more technology, it is often trying to name as a technical shortage what, inside the operation, has already become institutional tolerance. The electronic medical record exists, but relevant information arrives late. The alert exists, but it has become part of the background. Double-checking exists, but it coexists with exceptions so frequent that they no longer feel like exceptions. The dashboard exists, but the data feeding it was born incomplete, delayed, or adjusted after the fact. The wristband identifies the patient, the prescription goes through the system, the protocol was trained, the workflow was designed; still, at the end of the shift, someone reconstructs a line of care from memory, a parallel message, a support note, a quick phone call, and a tired confidence in the team’s good intentions.</p><p>This is the point almost no hospital likes to place on the table: the technology was implemented, but the clinical reality kept circulating outside it.</p><p>When that happens, the problem no longer belongs to IT and begins to expose a failure of operational command. A system does not govern behavior by itself. A platform does not correct a permissive culture. A mandatory field does not produce reliable data when the team has learned that filling in anything clears the pending task. An alert does not protect the patient when everyone knows it will be ignored without consequence. An indicator does not guide decisions when it measures an administrative fiction rather than care as it actually happened at the bedside. The hospital, then, sustains two institutions at the same time: the official one, clean, auditable, presented in meetings; and the real one, made of shortcuts, adaptations, normalized exceptions, and small violations that seem harmless until they line up inside an adverse event.</p><p>The most honest image of this scenario is not a room without computers. It is the opposite: screens lit, paid systems, contracted modules, published protocols, and beside them, a clipboard holding the information that determined care before it entered the medical record.</p><p>The distance between these two scenes is where risk lives.</p><p>Medication risk does not begin only in the final error of dose, route, or patient. It begins earlier, when medication reconciliation becomes a formal task, when an allergy was entered in the wrong field, when a verbal suspension did not become a traceable order, when checking happened later, when the alert competed with dozens of irrelevant warnings and lost clinical authority. Communication failure does not begin only because someone “didn’t notify.” It begins when the hospital accepts multiple channels as though they all carried the same clinical value. Lost information does not arise by chance. It arises when the official workflow demands documentation, but the routine rewards speed outside the workflow. Decisions based on guesswork do not arise from individual incompetence. They arise when the institution collects data as bureaucracy and then expects it to sustain management, safety, and clinical intelligence.</p><p>The technical name for this phenomenon is systemic underuse of technology. The practical name is more uncomfortable: the hospital invested in control, but preserved improvisation as a method.</p><p>And improvisation, inside a complex organization, rarely presents itself as disorder. It usually comes well dressed. It comes as “this was faster.” It comes as “I’ll document it later.” It comes as “everyone does it this way.” It comes as “the system gets in the way.” It comes as “in the emergency department there isn’t time.” Some of these phrases carry part of the truth, because poorly designed workflows do overload teams, bad screens steal clinical time, and disconnected systems push professionals toward shortcuts. The concession ends where institutional responsibility begins: when the hospital sees the deviation, lives with it, depends on it, and still keeps calling it digital transformation.</p><p>The mature question for a medical director, a CIO, or an operations leader is not how many systems the institution has contracted. The question is how much of the clinical truth passes through the system while it is still early enough to protect the patient.</p><p>Time is the word that dismantles the fantasy.</p><p>Data entered later loses clinical strength. A checklist completed later documents intention, not a safety barrier. A copied progress note preserves form, not reasoning. A mandatory field completed without criteria produces volume, not intelligence. Retrospective documentation supports a defensive record until the first case in which someone needs to reconstruct, minute by minute, who knew what, when they knew it, what decision they made, and which evidence was available at that moment. The hospital that accepts this pattern manufactures a gray zone: it appears to have traceability, but has reconstruction; appears to have governance, but has tolerance; appears to have data, but has operational residue.</p><p>That is why buying more technology at this stage often expands the problem instead of solving it.</p><p>More systems placed over an undisciplined operation create more escape routes. More alerts placed over a culture that does not distinguish clinical priority from administrative warning increase desensitization. More fields placed over a team already documenting under pressure produce defensive completion. More dashboards placed over fragile data give an appearance of precision to decisions that still emerge from contaminated material. The institution calls this digital maturity, but what grew was the surface area of technology, not the capacity to use it as an instrument of care.</p><p>The turning point does not begin with the next contract. It begins when leadership decides to measure adherence the way it measures infection, waiting time, length of stay, denied claims, readmission, and adverse events.</p><p>Real adherence is not attendance at training. It is observed behavior inside the workflow. It is prescription in the system before administration. It is checking at the point of care, not at the end of the shift. It is an alert with clinical hierarchy, ownership, expected response, and audit. It is an exception recorded as an exception, not as an invisible arrangement. It is data born where care happens, at the time when the decision is still alive. It is leadership that does not turn every operational difficulty into an excuse and does not turn every deviation into individual blame. Between blind punishment and permissiveness, there is management. A safe hospital lives in that narrow lane.</p><p>What no one likes to admit is that technology exposes the institution before it helps it.</p><p>It shows where the process does not fit the routine, where the protocol was written for the committee and not for the shift, where the team created shortcuts because the care design failed, where leadership stopped demanding compliance because the demand would expose its own omission, where data is treated as an administrative task even though it is the raw material of safety, clinical decision-making, audit, and operational intelligence. A system, when used well, is not one more screen. It is the place where the hospital leaves a trace of the care it claims to deliver.</p><p>Without that trace, the institution depends on memory.</p><p>And memory, in a hospital, works under sleep deprivation, interruption, urgency, emotional pressure, shift handoffs, overlapping tasks, and legal risk. No team, however competent, sustains care safety through memory indefinitely. Technology should reduce that load. When it becomes staged usage, it does the opposite: it adds bureaucracy to improvisation and still convinces management that risk is under control.</p><p>The difference between having a system implemented and having a system incorporated appears when the hospital loses margin.</p><p>During a packed shift, in the ICU without a bed, with a high-risk medication, during an internal transfer, with a patient without a companion, with a critical exam result that needs to reach someone, with an allergy entered in the wrong place, with a prescription changed near handoff, with a sepsis protocol started late, with a discharge organized from incomplete information. These are the points where technology stops being a topic of innovation and becomes the moral infrastructure of care: either it organizes responsibility, or it serves as scenery for disorganization.</p><p>The dangerous hospital of our time is not necessarily the one without technology.</p><p>It is the one that has enough technology to know better, but tolerates operating worse.</p><p>The action, therefore, needs to be less glamorous and more adult: audit real use, redesign workflows with those who execute them, remove useless alerts, protect time for clinical documentation, integrate systems that fragment the journey, define owners for exceptions, turn clinical data into a care asset, and create consequence for the difference between the declared workflow and the practiced workflow. Without that, every new module enters the same institutional organism that has already learned to bypass the previous one.</p><p>In the end, the question is not whether the hospital uses technology.</p><p>The question is where the truth of care happens.</p><p>When it happens inside the system, there is room for management, learning, safety, and improvement. When it happens on parallel paper, in informal groups, in the memory of the shift, or in documentation written later to close a pending task, the hospital is not digitized. It is decorated with technology.</p><p>And decoration does not protect patients.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=8a6423df6de2" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Systems Thinking in Healthcare]]></title>
            <link>https://abacelar.medium.com/systems-thinking-in-healthcare-507942e79ca9?source=rss-7a8f1e8140cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/507942e79ca9</guid>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[healthcare-technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Albert Bacelar]]></dc:creator>
            <pubDate>Thu, 06 Aug 2026 22:18:47 GMT</pubDate>
            <atom:updated>2026-08-06T22:18:47.977Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L0M50LblwcyRzC0aA-eK-g.png" /><figcaption>Dr2 thinktech — artificial intelligence</figcaption></figure><p>The sovereignty of systems thinking is not born in a planning room, nor does it survive inside a polished presentation for the board; it appears, without asking permission, when the hospital realizes it has bought systems, hired teams, created indicators, digitized forms, installed dashboards, automated steps, and still the patient continues to move through the institution as if crossing corridors that do not speak to one another.</p><p>That is the dividing line.</p><p>On one side, the organization that adds technologies and believes integration will emerge from accumulation. On the other, the institution that understands a less comfortable truth: in healthcare, every local decision charges interest across the entire system.</p><p>The fragmented hospital rarely sees itself as fragmented. It sees itself as hardworking. There are competent people in every department, meetings, targets, protocols, platforms, spreadsheets, committees, and legitimate urgency at almost every end of the operation. The operating room tries to keep rooms moving. The emergency department tries to clear stretchers. The ICU tries to release beds. Nursing tries to close staffing gaps. Billing tries to correct denials. Quality tries to reduce adverse events. IT tries to keep systems breathing. Leadership tries to keep the numbers viable without losing care safety. No one wakes up intending to harm care.</p><p>And still, the whole becomes ill.</p><p>The cause does not sit in lack of effort. It sits in the absence of institutional cognitive architecture, the inability to convert workflow, data, risk, accountability, and consequence into one operational language.</p><p>When each department optimizes its own task without seeing the full journey, the institution earns small local victories and accumulates systemic losses. The laboratory reduces its internal processing time, but the result does not reach the clinical decision at the right moment. Patient flow organizes the queue, but does not converse with discharge prediction. Pharmacy controls stock, but fails to read prescribing patterns that anticipate shortage. Bed management accelerates movement, but overlooks clinical frailty, readmission risk, and the real capacity of the next shift. The electronic health record stores everything, but guides very little. The dashboard displays numbers, but does not govern decisions.</p><p>The hospital becomes rich in screens and poor in synthesis.</p><p>One scene summarizes this type of failure: a patient ready to leave the ICU at 10 a.m., a ward unable to receive with predictability, a family without clear information, a critical bed blocked, an unstable patient waiting in the emergency department, a medical team negotiating priority over the phone, while the systems record correct fragments of the wrong story. Nothing there looks like collapse when viewed in isolation. The collapse lives between the parts.</p><p>That intermediate space holds much of the hospital’s waste.</p><p>When digitization enters without redesign, it does not repair that space; it accelerates the circulation of error. A poor paper form consumes time. A poor digital form consumes time, scales irritation, creates false traceability, and still gives the institution the feeling of modernization. The process remains crooked, now with a login, a password, and a support ticket.</p><p>Technology does not rescue a workflow no one had the courage to map.</p><p>Artificial intelligence in healthcare faces the same risk. When treated as a tool acquisition, it becomes one more layer over the same fragmented arrangement: one agent for access, another for summaries, another for prioritization, another for billing, another for reporting, each responding to a partial pain, each producing an apparent gain, each widening the distance between automation and governance. The question that defines the value of clinical AI is not which model answers better. The question is where the decision fails, who carries the consequence, which data sustains the inference, which workflow changes after the response, and which indicator proves that care improved without transferring invisible burden to the team.</p><p>Without that, AI becomes peripheral sophistication.</p><p>Systems thinking in a hospital environment does not mean thinking big. It means thinking connected. It means knowing that a delayed discharge does not belong to the ICU; it belongs to the chain of care, hospitality, pharmacy, transportation, patient flow, family, clinical team, authorization, clinical risk, and demand prediction. It means knowing that an emergency department queue does not begin at the emergency entrance; it begins in primary care, outpatient access, exam availability, hospital back-up capacity, bed governance, and the internal culture of decision-making. It means knowing that burnout is not born only from workload; it is born from workload added to friction, useless repetition, duplicate documentation, broken communication, and the feeling that the institution demands excellence with tools that make care harder.</p><p>The system speaks before it breaks. Most organizations listen late.</p><p>Institutional maturity appears when the hospital stops asking “which solution should we buy?” and starts asking “which decision must become more predictable, safer, and measurable?” That shift changes the conversation. The subject leaves the technology showcase and reaches the right table: clinical workflow, data governance, clinical risk, team adoption, auditability, continuity of the patient journey, operational impact, and executive responsibility.</p><p>The purchase stops being the first move. Diagnosis takes the place it should always have occupied.</p><p>A diagnosis of clinical AI and institutional orchestration begins with the map of decision-making, not with a catalog of tools. Where is the decision born? At which point does it slow down? Who documents it? Who validates it? Who suffers when it fails? Which data arrives too early, too late, or detached from context? Which human task deserves computational support, and which task demands presence, judgment, and clinical conversation? Which indicator measures care, and which one merely measures activity?</p><p>That distinction separates a digital hospital from an intelligent hospital.</p><p>A digital hospital records. An intelligent hospital anticipates, coordinates, and learns without humiliating the team through rework. A digital hospital accumulates systems. An intelligent hospital defines interoperability as a management discipline, not as a technical promise. A digital hospital celebrates implementation. An intelligent hospital measures real adoption, workflow deviation, safety, response time, clinical variation, team satisfaction, and consequence for the patient.</p><p>The sovereignty of systems thinking is the capacity to avoid being governed by one’s own fragmentation.</p><p>There is a frequent trap in healthcare leadership: mistaking available information for institutional comprehension. The hospital has business intelligence, indicators, reports, spreadsheets, meetings, and dashboards. Even so, relevant decisions still depend on informal messages, a coordinator’s memory, phone calls between departments, and bedside heroism. The institution has data, but no decision circuit. It has records, but no coordination. It has technology, but no shared cognition.</p><p>The uncomfortable question is not whether data exists. The question is whether it reaches the exact point where it changes conduct.</p><p>When that fails, the system manufactures blindness with the appearance of control. The board sees averages. Operations lives variation. Leadership tracks a monthly indicator. The team faces minute-by-minute instability. The patient experiences the failure as waiting, misalignment, repeated history-taking, delayed response, and a disoriented journey. In the meeting room, the problem receives an administrative name. In the corridor, it gains a body.</p><p>And healthcare always charges through the body.</p><p>Cognitive integration changes this logic because it connects three layers that usually move apart: the real workflow, reliable data, and governed decision-making. Real workflow is what happens outside the elegant process map. Reliable data preserves context, origin, timing, quality, and purpose. Governed decision-making defines criteria, limits, responsibility, audit trail, and consequence. Without these three layers, any automation operates on unstable ground.</p><p>That is why interoperability cannot be reduced to systems that “talk.” Systems exchange information; mature institutions align meaning. A field filled in one module must carry the same clinical, operational, and administrative interpretation in another. An alert must distinguish risk from nuisance. An automated recommendation must arise from validated data, respect clinical context, and leave enough traceability for review. An AI agent must work inside explicit boundaries: support, prioritize, summarize, signal, route, never replace responsibility without governance.</p><p>The problem is not the machine making an error. The problem is the institution not knowing where the error would enter, who would notice it, how it would be contained, and what learning would return to the process.</p><p>This is the point many projects ignore. The seduction lies in the demo that works. The risk lives in routine. In a demo, the case is clean, the user is attentive, the scope looks controlled. In routine, the patient arrives incomplete, the data comes late, the shift changes, the connection drops, the team creates workarounds, the alert causes fatigue, and the exception becomes habit. Healthcare technology must survive a Tuesday shift at 3:12 a.m., when no one wants to admire the solution; everyone needs it to stay out of the way of the decision.</p><p>Systemic maturity proves itself at that hour.</p><p>The exponential healthcare mentioned in so many institutional speeches will not be produced by isolated speed. Speed without coordination increases collision. Precision without context narrows the field of vision. Automation without validation transfers risk. Personalization without integration fragments the journey even more. Prediction without operational capacity creates statistical impotence: the institution knows earlier, yet remains unable to act in time.</p><p>To anticipate without coordinating is a sophisticated form of helplessness.</p><p>The probable future of healthcare management will demand less fascination with tools and more sovereignty over decision flows. Hospitals that learn to orchestrate data, processes, and teams will reduce waste without crushing people, apply AI without staging modernity, improve patient experience without cosmetic treatment of bottlenecks, and sustain efficiency without turning care into an assembly line. Hospitals that remain fragmented will keep buying solutions for symptoms produced by their own institutional design.</p><p>The difference will become visible in small details, the kind rushed governance tends to ignore. The patient will not repeat the same information five times. The team will not document the same data in three places. The ICU will not announce discharge after the operational window has already been lost. The emergency department will not function as a storage room for previous failures. The manager will not decide from a delayed photograph. The physician will not be forced to choose between caring and feeding the system. The indicator will not exist to punish later, but to correct earlier.</p><p>That is the real shift: moving from an institution that reacts to an institution that sees the chain of consequence while it is still forming.</p><p>Systems thinking does not soften the complexity of healthcare. It prevents complexity from being managed in pieces. In practice, it means selecting a relevant problem, mapping the living workflow, qualifying the data that sustains the decision, defining governance, testing at controlled scale, measuring clinical and operational impact, adjusting with the team, auditing system behavior, and expanding only when safety accompanies the gain. It looks less shiny than buying a large platform. It delivers more.</p><p>Because a hospital does not break only when technology is missing. A hospital breaks when it loses the capacity to understand itself.</p><p>The sovereignty of systems thinking is that recovered capacity: seeing the patient not as an episode, data not as a record, AI not as spectacle, the team not as a replaceable resource, the indicator not as executive decoration, and operations not as backstage. Everything converses. Everything charges. Everything returns to care.</p><p>The organization that understands this governs with intention.</p><p>The one that does not will keep calling fragmentation growth, rework dedication, dashboards intelligence, automation progress, and improvisation strong culture.</p><p>In the end, the question is not whether the institution has solutions. Almost all of them do.</p><p>The question is whether it has enough system for those solutions not to increase the problem.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=507942e79ca9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A IA respondeu corretamente. A métrica deu nota zero.]]></title>
            <link>https://abacelar.medium.com/a-ia-respondeu-corretamente-a-m%C3%A9trica-deu-nota-zero-9fde4adf375c?source=rss-7a8f1e8140cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/9fde4adf375c</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[healthcare-technology]]></category>
            <dc:creator><![CDATA[Albert Bacelar]]></dc:creator>
            <pubDate>Tue, 04 Aug 2026 22:08:35 GMT</pubDate>
            <atom:updated>2026-08-04T22:08:35.503Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*iER1i8TOs9VQ-z05GvBn1w.png" /></figure><p>Uma resposta clínica aberta admite mais de uma redação correta. O paciente descreve um sintoma, o modelo organiza as hipóteses, orienta a procura por atendimento e explica os sinais de alarme. O conteúdo respeita o conhecimento médico disponível. Ainda assim, basta usar palavras diferentes das que aparecem na resposta de referência para o BLEU registrar uma pontuação próxima de zero.</p><p>Foi o que Ting Fang Tan e colaboradores encontraram ao avaliar respostas produzidas por GPT-4o, Claude 4 Sonnet e DeepSeek em oftalmologia, uso de medicamentos e anestesia. Respostas consideradas adequadas por especialistas receberam pontuações BLEU entre 0,00 e 0,04. O resultado expõe um erro de medição que atravessa parte dos estudos sobre inteligência artificial em saúde: a métrica registra semelhança entre textos e seu número acaba interpretado como qualidade clínica.</p><p>O BLEU calcula a sobreposição de palavras e pequenas sequências entre a resposta gerada e um texto de referência. Se o gabarito afirma “procure atendimento médico imediatamente” e o modelo escreve “dirija-se agora a um serviço de emergência”, a orientação clínica permanece. A coincidência lexical cai. Para a métrica, a distância aumentou; para o paciente, a conduta recebida continua equivalente.</p><p>Essa diferença compromete a interpretação dos resultados. Uma pontuação BLEU baixa não significa, por si, diagnóstico incorreto, recomendação insegura ou explicação inadequada. Significa que o texto produzido se afastou da formulação escolhida como referência. Quando pesquisadores convertem esse afastamento em erro clínico, deixam de avaliar o conteúdo que interessa e passam a premiar obediência verbal ao gabarito.</p><p>O problema cresce nas perguntas abertas. Em questões de múltipla escolha, existe uma alternativa correta previamente delimitada. Nas respostas discursivas, profissionais distintos organizam o mesmo raciocínio com escolhas lexicais, níveis de detalhe e sequências argumentativas diferentes. Uma única resposta de referência representa uma formulação aceitável, não todo o espaço de respostas clinicamente válidas.</p><p>Foi para enfrentar essa limitação que os autores propuseram o S.C.O.R.E., uma rubrica de avaliação por especialistas composta por cinco dimensões independentes, graduadas de 1 a 5.</p><p>Safety examina se a resposta contém informação enganosa associada a dano físico ou psicológico.</p><p>Consensus and Context verifica a exatidão e o alinhamento com o consenso clínico aplicável ao contexto apresentado.</p><p>Objectivity identifica discriminação ou decisões apoiadas em critérios injustos.</p><p>Reproducibility mede a consistência entre três respostas geradas para a mesma pergunta.</p><p>Explainability avalia a justificativa fornecida, o raciocínio apresentado e as informações que sustentam a orientação.</p><p>A independência entre os eixos preserva a localização da falha. Uma resposta alcança boa explicabilidade e baixa segurança; outra respeita o consenso, porém muda de conduta a cada geração. Uma nota agregada esconderia parte dessa diferença. O S.C.O.R.E. mantém cada deficiência visível.</p><p>O estudo também encontrou uma variação que as classificações gerais de modelos costumam omitir. O GPT-4o obteve o melhor resultado em oftalmologia, Claude 4 Sonnet liderou nas perguntas sobre medicamentos e DeepSeek ficou à frente em anestesia. O modelo vencedor mudou junto com a especialidade.</p><p>Esse achado requer leitura conjunta com o método empregado. Os parâmetros haviam sido ajustados nas perguntas oftalmológicas e depois aplicados, sem modificação, aos três modelos e às demais áreas. O GPT-4o liderou justamente no domínio usado para o ajuste. Os autores interpretaram esse resultado como indício de que o efeito da calibração ficou detectável na avaliação.</p><p>A própria rubrica apresentou comportamento semelhante. A consistência dos avaliadores se sustentou na oftalmologia, especialidade usada no ajuste. Em medicamentos e anestesia, a concordância foi baixa. O S.C.O.R.E. corrige o objeto avaliado, pois desloca a atenção da coincidência textual para atributos clínicos. Sua aplicação fora do domínio calibrado ainda exige definições operacionais, treinamento dos avaliadores e nova verificação de consistência.</p><p>Esse limite não reduz o valor do trabalho. Define seu alcance. Foram utilizadas cinco perguntas por especialidade, cada área contou com um avaliador e as respostas foram geradas três vezes. Trata-se de uma prova de conceito. A amostra não sustenta uma classificação definitiva entre GPT-4o, Claude e DeepSeek, nem valida a rubrica para toda a Medicina.</p><p>O resultado mais relevante aparece antes da comparação entre os modelos. Se uma resposta clinicamente correta recebe BLEU 0,00, o número não está medindo aquilo que o leitor acredita. E, quando a régua mede outra coisa, aumentar o tamanho do benchmark repete o erro com maior precisão estatística.</p><p>Referência: Tan TF, Elangovan K, Ong J, Ke Y, Lee A, Shah N, Sung J, Wong TY, Xue L, Liu N, Wang H, Kuo CF, Chesterman S, Yeong ZK, Ting DSW. <em>A S.C.O.R.E. framework for evaluating open-ended responses from large language models in healthcare</em>. Cell Reports Medicine. 2026. <a href="https://doi.org/10.1016/j.xcrm.2026.102883">DOI: 10.1016/j.xcrm.2026.102883</a>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9fde4adf375c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A regular ChatGPT is better than the medical AI agents you are being sold.]]></title>
            <link>https://abacelar.medium.com/a-regular-chatgpt-is-better-than-the-medical-ai-agents-you-are-being-sold-649c5096cac1?source=rss-7a8f1e8140cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/649c5096cac1</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[healthcare-technology]]></category>
            <category><![CDATA[healthcare]]></category>
            <dc:creator><![CDATA[Albert Bacelar]]></dc:creator>
            <pubDate>Tue, 04 Aug 2026 00:06:32 GMT</pubDate>
            <atom:updated>2026-08-04T00:06:32.086Z</atom:updated>
            <content:encoded><![CDATA[<h3>A regular ChatGPT is better than the medical AI agents you are being sold. A blinded test proved it.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B1JECTK3WKv3_u0TBHovpA.png" /></figure><p>OpenEvidence and UpToDate Expert AI are sold to physicians and hospitals with a very clear promise: because they were built for medicine and consult recognized clinical databases, they should provide better answers than a regular ChatGPT. The reasoning sounds convincing. A medical platform understands the language of the field, searches specialized content, and charges for that specialization. The price assumes a real advantage.</p><p>Researchers at NYU Langone decided to measure that advantage. They compared OpenEvidence and UpToDate Expert AI with 3 general-purpose models: GPT-5.2, Gemini 3.1 Pro, and Claude Opus 4.6. In simple terms, they compared medical AI tools sold as specialized products with models designed for anyone and capable of answering questions about virtually any subject.</p><p>The test had 3 parts. In the first, the systems answered 500 medical questions similar to US medical licensing exams. In the second, they responded to another 500 cases used to assess quality, safety, clarity, and alignment with physician judgment. In the third, they received 100 real questions asked by physicians while treating patients.</p><p>The most important part came next. The answers to those 100 real questions were removed from the platforms, mixed together, and given to 12 physicians. The reviewers did not know whether each answer came from GPT, Gemini, Claude, OpenEvidence, or UpToDate Expert AI. They saw the text and assigned a score. Altogether, the study included 1,800 independent physician evaluations.</p><p>The general-purpose models won all 3 comparisons.</p><p>On the medical questions, Gemini achieved 97.4% accuracy, GPT-5.2 reached 94.2%, and Claude scored 90.2%. OpenEvidence scored 89.6%, while UpToDate Expert AI scored 88.4%.</p><p>In the test assessing the quality of open-ended answers, GPT-5.2 scored 88%. Gemini scored 79.3%, and Claude scored 77%. OpenEvidence reached 62.6%, while UpToDate Expert AI scored 61.3%. The difference between the general-purpose GPT model and the 2 clinical platforms exceeded 25 percentage points.</p><p>The real-world questions produced the same result. On a scale from 1 to 4, Gemini received an average score of 3.62, GPT-5.2 scored 3.54, and Claude scored 3.52. OpenEvidence received 3.24, while UpToDate Expert AI scored 3.17. Google’s automated AI summary scored 3.27, a result close to those of the paid medical platforms.</p><p>Here, “regular ChatGPT” means a general-purpose model without a specialized clinical platform layered on top of it. The study used advanced versions of these models with external search enabled. The central comparison remains unchanged: the general-purpose systems produced better answers than tools marketed with the promise of medical superiority.</p><p>The result directly challenges the sales pitch. OpenEvidence and UpToDate Expert AI provide organized access to clinical content, physician-oriented interfaces, and integration with recognized sources. Those features have practical value. The problem begins when organization, branding, and access to a medical database are presented as evidence of better answers.</p><p>A reliable source inside the system does not guarantee that the AI will find the correct passage, understand the question, and use the information appropriately. The platform searches for documents, selects a set of results, and builds its answer from them. When the search retrieves incomplete material or information that is only loosely related to the case, the answer is compromised from the outset, even when the underlying database has an excellent reputation.</p><p>The authors consider this mechanism one possible explanation for the weaker performance of the clinical platforms. The internal architecture of OpenEvidence and UpToDate Expert AI is closed, so the study could not determine exactly where the loss of quality occurred. The measurable finding remains: when the brands were hidden and physicians evaluated the content alone, GPT, Gemini, and Claude received higher scores.</p><p>This changes how these products should be purchased. A hospital should not adopt a tool merely because its name includes words such as “medicine,” “clinical,” or “evidence.” Before signing a contract, the platform should answer the same questions as a general-purpose model, with all identifying information removed and the responses evaluated by the professionals who will use the system.</p><p>Without that comparison, the institution is paying for the promise of specialization. In this study, that promise lost to a regular ChatGPT.</p><p><strong>Reference:</strong> Vishwanath K, et al. <em>General-purpose large language models outperform specialized clinical AI tools on medical benchmarks</em>. <em>Nature Medicine</em>. 2026. DOI: 10.1038/s41591–026–04431–5.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=649c5096cac1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[I Took GitHub’s First Agentic AI Developer Exam as a Physician. Here Is What It Actually Tested.]]></title>
            <link>https://abacelar.medium.com/i-took-githubs-first-agentic-ai-developer-exam-as-a-physician-here-is-what-it-actually-tested-4f83c6be790e?source=rss-7a8f1e8140cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/4f83c6be790e</guid>
            <category><![CDATA[healthcare-technology]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Albert Bacelar]]></dc:creator>
            <pubDate>Fri, 31 Jul 2026 15:35:01 GMT</pubDate>
            <atom:updated>2026-07-31T15:35:01.383Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*kfJT74ozbU7jHyXM2RNjQw.png" /></figure><p><em>Why agentic AI is becoming a discipline of architecture, governance, evaluation, and accountability, and why medicine may be one of the fields that needs this skill most.</em></p><p>On May 18, 2026, I sat in front of my computer to take the beta version of the GH-600 exam, the assessment behind the new <strong>GitHub Certified: Agentic AI Developer</strong> credential.</p><p>There was no immediate result.</p><p>Because this was a beta exam, Microsoft and GitHub needed time to analyze the questions, validate their statistical performance, establish the final scoring model, and process the results of the first group of candidates.</p><p>On July 27, the email arrived:</p><p><strong>You passed your Developing in Agentic AI Systems exam.</strong></p><p>A second message followed:</p><p><strong>You’ve earned your GitHub Certified: Agentic AI Developer certification.</strong></p><p>The credential matters to me for reasons that go beyond adding another badge to a profile. I am a physician, intensive care specialist, data scientist, developer, professor, and entrepreneur. My professional life has always moved across domains that are usually treated as separate: clinical decisions, data architecture, software development, organizational management, and risk.</p><p>Agentic AI places those domains inside the same system.</p><h3>The examination experience</h3><p>The exam was remotely proctored. An Indian human evaluator monitored the session throughout the assessment, from the initial identification and room inspection to the completion of the test.</p><p>The supervision was continuous. The desk, room, computer, webcam, microphone, and examination environment had to comply with the testing rules. The evaluator remained responsible for checking whether the candidate was alone, whether the workspace contained unauthorized material, and whether any behavior could compromise the integrity of the examination.</p><p>That detail deserves attention.</p><p>Agentic AI is frequently discussed through demonstrations that last a few minutes: an agent opens a browser, writes code, creates a document, calls an API, or completes a sequence of tasks. The GH-600 exam approached the subject from a different perspective. The candidate was being assessed under strict controls while answering questions about how to design other systems that must also operate under strict controls.</p><p>The form of the examination reflected part of its content: identity, permissions, observation, rules, traceability, and accountability.</p><h3>Being part of the first cohort</h3><p>I took the exam while it was still in beta, which placed me in the first global cohort evaluated for this certification.</p><p>A beta candidate faces a particular situation. The exam already measures a defined body of knowledge, although the final scoring model has not yet been completed. Results are withheld while the exam provider evaluates question quality, difficulty, discrimination, and consistency.</p><p>That means candidates leave the test without knowing whether they passed.</p><p>In my case, the interval between taking the exam and receiving the result lasted more than two months. During that period, the certification did not appear in my Microsoft Learn profile, the score was unavailable, and the examination history was difficult to locate through the platform’s changing interfaces.</p><p>The delay made the eventual result more meaningful. The credential was not generated by completing a course or watching a sequence of lessons. It required passing a supervised examination in a new technical field whose boundaries are still being defined.</p><h3>What the exam actually evaluated</h3><p>The certification is called <strong>Agentic AI Developer</strong>, although the scope extends well beyond writing an agent prompt or connecting a language model to a tool.</p><p>The assessment covered six domains:</p><ol><li>Preparing agent architecture and software development lifecycle processes</li><li>Implementing tool use and environment interaction</li><li>Managing memory, state, and execution</li><li>Performing evaluation, error analysis, and tuning</li><li>Orchestrating multi-agent coordination</li><li>Implementing guardrails and accountability</li></ol><p>These domains describe a professional responsibility that is broader than conventional software development.</p><p>An agentic system must receive an objective, determine or follow a plan, interact with external resources, preserve relevant state, evaluate intermediate results, manage failure, and operate within defined limits. When multiple agents are involved, the system must also coordinate responsibilities, resolve conflicts, and prevent duplicated or contradictory actions.</p><p>A system that writes a plausible answer is one thing.</p><p>A system that changes a repository, accesses files, calls external services, creates pull requests, executes workflows, stores memory, and influences business or clinical operations belongs to another risk category.</p><p>The examination focused on that second category.</p><h3>Tool use changes the nature of the risk</h3><p>Language models become operationally significant when they gain access to tools.</p><p>A model without tools produces output. A model with tools affects systems.</p><p>It may read a private repository, change code, query a database, send a message, update a ticket, invoke a cloud service, create a deployment, or retrieve information from an external environment.</p><p>The engineering question therefore changes. The developer must define which tools the agent can access, under which identity, with what permissions, for how long, and under which conditions.</p><p>The agent also needs a reliable method for handling tool errors. APIs return incomplete data. Credentials expire. Files change. Network calls fail. External systems provide unexpected schemas. A tool may succeed technically while producing a result that violates the business objective.</p><p>Tool integration requires more than connectivity. It requires contracts, validation, observability, and controlled failure.</p><h3>Memory is a design decision</h3><p>The word “memory” is often used loosely in discussions about AI agents.</p><p>In production systems, memory must be decomposed.</p><p>There may be short-term conversational context, working memory for the current task, durable memory across sessions, retrieved organizational knowledge, execution state, user preferences, and audit records. Each category serves a different purpose and carries a different risk.</p><p>What should the agent remember?</p><p>Who authorized that information to be stored?</p><p>How long should it remain available?</p><p>Which users or agents may retrieve it?</p><p>How will incorrect or outdated memory be corrected?</p><p>In healthcare, these questions are immediate. A clinical system cannot treat an old assumption, an unverified note, and a validated medical record as equivalent forms of memory. Provenance matters. Time matters. Context matters. The authority of the source matters.</p><p>The same principle applies to legal systems, financial processes, software repositories, and corporate decision-making.</p><h3>Evaluation must examine behavior, not appearance</h3><p>Agentic systems can fail while producing polished output.</p><p>A response may appear coherent while the agent has used the wrong source, ignored a constraint, selected an inappropriate tool, repeated an action, concealed uncertainty, or stopped before completing the objective.</p><p>Evaluation must therefore examine the execution path.</p><p>Did the agent understand the task?</p><p>Did it select the correct tool?</p><p>Did it use the correct parameters?</p><p>Did it verify the result?</p><p>Did it respect permissions?</p><p>Did it recognize a failure?</p><p>Did it recover correctly?</p><p>Did it know when human intervention was required?</p><p>Error analysis in agentic systems resembles the analysis of a process more than the grading of a paragraph. The final output is one artifact within a chain of decisions.</p><p>This was one of the most relevant aspects of the certification. It treats agent development as an engineering discipline in which quality depends on the entire trajectory from instruction to action.</p><h3>Multi-agent coordination is an organizational problem</h3><p>Adding more agents does not automatically improve a system.</p><p>Multiple agents create new coordination costs. They may compete for the same resource, overwrite each other’s changes, repeat the same work, rely on inconsistent context, or create circular delegation.</p><p>A multi-agent architecture needs explicit responsibilities.</p><p>One agent may plan. Another may execute. A third may review. Another may monitor policy compliance. That separation can improve control when the boundaries are clear and the exchange of information is structured.</p><p>The architecture must also define who has authority to stop the process.</p><p>That point connects agentic systems to real organizations. Companies already struggle with unclear ownership, fragmented information, duplicated work, and decisions made without accountability. Reproducing those defects through autonomous software creates faster confusion.</p><p>Multi-agent design is partly software architecture and partly institutional design.</p><h3>Guardrails are part of the product</h3><p>In many AI projects, safeguards are added after the main system has been built.</p><p>Agentic AI makes that sequence difficult to defend.</p><p>Guardrails affect what the system is allowed to do, which data it may access, what actions require approval, how outputs are validated, and how incidents are investigated. These controls shape the architecture from the beginning.</p><p>A production agent may need permission boundaries, approval gates, sandboxed execution, rate limits, policy checks, action logs, rollback mechanisms, and human escalation.</p><p>Accountability also requires attribution. When an action occurs, the organization must know which agent initiated it, which model and configuration were used, which tools were called, what data informed the decision, and which person or policy authorized the operation.</p><p>Without that record, the organization has automation without governance.</p><h3>Why this matters in medicine</h3><p>My medical background changed the way I interpreted the exam.</p><p>Clinical work trains physicians to operate in environments where incomplete information, time pressure, uncertainty, and asymmetric consequences coexist. A minor error in one context may be reversible. The same error in a high-risk clinical process may cause irreversible harm.</p><p>Medicine also teaches that competence does not consist of generating an answer. It involves selecting the right action for the right patient at the right time, while considering contraindications, competing priorities, available resources, and the consequences of delay.</p><p>Agentic AI introduces similar requirements into software systems.</p><p>A clinical agent that identifies possible sepsis, for example, should not be evaluated solely on whether it produces the word “sepsis.” Its performance depends on when the condition was detected, which data supported the alert, whether the system differentiated missing data from normal data, which protocol was activated, who received the notification, whether the recommendation was acknowledged, and whether the process generated an auditable record.</p><p>The agent exists inside a care pathway.</p><p>This is why physicians should participate in agentic AI development. Domain expertise is required to define meaningful objectives, acceptable failure modes, escalation rules, and boundaries of autonomy.</p><p>Technical teams can build the mechanism. Clinical teams must help define what responsible action means.</p><h3>A physician with an Agentic AI Developer certification</h3><p>I have not found a public record of another Brazilian physician holding this certification. That does not establish that I am the first, though it suggests that I may be among the earliest physicians in Brazil to earn it.</p><p>The more relevant point is what this combination represents.</p><p>Healthcare increasingly depends on software, data pipelines, decision-support systems, automation, interoperability, and AI-assisted workflows. Physicians who understand these systems can participate in their design instead of encountering them only after deployment.</p><p>At the same time, developers entering healthcare need more than technical fluency. They need to understand clinical context, professional responsibility, data sensitivity, workflow constraints, and the difference between a technically valid response and a clinically appropriate action.</p><p>The intersection is where much of the work now resides.</p><h3>The certification is a marker, not a finish line</h3><p>The GitHub Certified: Agentic AI Developer credential confirms competence across architecture, tool integration, memory, evaluation, orchestration, guardrails, and accountability.</p><p>It also signals the emergence of a new professional profile.</p><p>Organizations will need people who can connect models to real systems while preserving control. They will need developers who understand governance, domain specialists who understand software, and leaders who can assess when autonomy creates value and when it creates unacceptable exposure.</p><p>For me, the certification joins medicine, data science, software engineering, and organizational design within the same professional path.</p><p>The next generation of AI systems will not remain inside chat windows. They will enter workflows, repositories, hospitals, companies, legal operations, and decision processes.</p><p>The question is no longer whether agents can act.</p><p>The question is whether we can design their actions with enough precision, evidence, restraint, and accountability to deserve trust.</p><p><a href="https://medium.com/u/8df3bf3c40ae">GitHub</a></p><p><a href="https://medium.com/u/9e2fd63dfd89">Geekologi™</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4f83c6be790e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[A physician who uses artificial intelligence without understanding its limits delegates a decision…]]></title>
            <link>https://abacelar.medium.com/a-physician-who-uses-artificial-intelligence-without-understanding-its-limits-delegates-a-decision-677122a16527?source=rss-7a8f1e8140cf------2</link>
            <guid isPermaLink="false">https://medium.com/p/677122a16527</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[ai-in-healthcare]]></category>
            <category><![CDATA[healthcare]]></category>
            <category><![CDATA[healthcare-technology]]></category>
            <dc:creator><![CDATA[Albert Bacelar]]></dc:creator>
            <pubDate>Fri, 31 Jul 2026 15:34:55 GMT</pubDate>
            <atom:updated>2026-07-31T15:34:55.838Z</atom:updated>
            <content:encoded><![CDATA[<h3>A physician who uses artificial intelligence without understanding its limits delegates a decision that still remains their responsibility</h3><figure><img alt="Artificial Intelligence — Albert Bacelar" src="https://cdn-images-1.medium.com/max/1024/1*zcdEL_ABkujKmad2nnIZoA.png" /></figure><p>Medical education still treats artificial intelligence as an operational skill. Professionals are taught how to access a tool, structure prompts, interpret a response, and incorporate the result into their work. They learn how to activate the system before learning how to examine the path that led to its output.</p><p>This mismatch matters because algorithmic error rarely begins when a recommendation appears on a screen. By the time a physician receives an inaccurate risk score, diagnostic hypothesis, or clinical alert, many of the conditions that produced that result have already been established. They were embedded in the composition of the training dataset, the way the data were recorded, the reference standard selected, the validation design, and the environment in which the system was deployed.</p><p>The article <em>Integrating AI Literacy into Medical Education: Preparing Future Clinicians for an AI-Driven Healthcare System</em>, published in <em>Medical Science Educator</em>, organizes AI education into six domains. The framework includes technical foundations, evidence appraisal, ethical and social implications, clinical use, workflow integration, and continuous improvement. None of these domains requires physicians to become programmers. All of them require critical reading, methodological judgment, and an understanding of the care environment.</p><p>This point should reshape the discussion about AI literacy in healthcare.</p><p>During medical training, we learn how to assess clinical trials, identify bias, distinguish association from causation, examine outcomes, interpret measures of diagnostic performance, and determine whether the findings of a study apply to the population we treat. When an algorithm enters the discussion, that discipline is often replaced by a superficial assessment: the system appears advanced, it was trained on a large dataset, and it reports a high performance metric. Its recommendation then acquires a level of authority that clinical studies themselves rarely receive.</p><p>An accuracy score, in isolation, says little about a system’s safety. It is necessary to know who was included in the training population, which patients were excluded, how missing data were handled, what the prevalence of the outcome was, which variables served as imperfect proxies for clinical phenomena, and whether the model was validated in institutions other than the one in which it was developed.</p><p>When this information does not reach the end user, the physician receives an answer without receiving the conditions that support its validity.</p><p>Data quality is also often discussed in vague terms. In healthcare, data are never produced in a neutral setting. They carry the routines of the service, time pressure, the quality of clinical documentation, patterns of test ordering, specialist availability, billing rules, and the decisions made by the care team before anything was recorded. A hospital dataset contains both the patient’s disease and the institution’s behavior in response to it.</p><p>When a hospital orders a particular test more frequently, admits patients according to its own criteria, or records diagnoses differently, the algorithm learns those practices. Part of what appears to be clinical intelligence is the statistical repetition of a local habit. When the model is transferred to another institution, performance declines because the new environment no longer matches the relationships learned during training.</p><p>External validation is meant to address this problem. Yet many studies rely on closely related datasets, similar populations, or institutions operating within the same ecosystem. The model crosses an administrative boundary without crossing a clinically meaningful one. The result is described as external validation, although the model remains sheltered from the differences it will encounter in routine care.</p><p>Implementation adds another layer of risk. A technically sound model changes the behavior of those who use it. It shifts priorities, alters the time devoted to each patient, generates alerts, moves steps within the workflow, and influences professionals with different levels of experience. A highly sensitive system causes harm when it produces more alerts than the team can reasonably manage. A well-calibrated model loses its value when its output arrives after the clinical decision has already been made. A correct recommendation, displayed without context, encourages automatic acceptance.</p><p>For this reason, evaluating AI in healthcare requires examining what happens after deployment. How many recommendations were accepted? Under what circumstances did clinicians disagree? Which patient groups accounted for most errors? Did performance change over time? Did the team begin to ignore certain alerts? Did the system alter clinical decisions in ways that had not been anticipated?</p><p>These questions belong to medical practice, healthcare management, and clinical governance. They should not remain confined to technology teams.</p><p>Medical education must prepare professionals to follow the entire lifecycle of an algorithm. This includes understanding the purpose for which it was created, examining the evidence used to justify its adoption, recognizing differences between the study population and the population receiving care, documenting disagreements, reviewing adverse events, and discontinuing use when real-world performance diverges from published results.</p><p>Physicians do not need to understand every mathematical detail of a model. They need to understand enough to recognize when a statistically plausible answer becomes clinically inappropriate.</p><p>This competence will be rarer than access to the tools themselves. The use of artificial intelligence is likely to become commonplace, including among professionals with no specific training in the field. The ability to audit its use will remain limited as long as curricula continue to treat technology as an instrument rather than as part of the decision-making environment.</p><p>This is where the professional distinction will emerge. On one side will be physicians who receive the system’s output and incorporate it into care. On the other will be physicians who examine the provenance of the data, the reference population, the limits of validation, the effect on workflow, and the compatibility of the recommendation with the patient in front of them.</p><p>Artificial intelligence introduces a new source of information into medicine. It also introduces a new source of error. Those who learn merely how to use the tool understand half of the work. The other half begins when the result appears too convincing to question.</p><p>CHANG, Linda; SREEDHAR, Radhika. Integrating AI literacy into medical education: preparing future clinicians for an AI-driven healthcare system. <em>Medical Science Educator</em>, [S. l.], v. 36, n. 1, p. 17–22, 2026. DOI: 10.1007/s40670–025–02599-y. (<a href="https://link.springer.com/article/10.1007/s40670-025-02599-y?utm_source=chatgpt.com">Springer</a>)</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=677122a16527" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>