AI Agents for Clinical Decision Support & Diagnostics

This module opens the course by grounding diagnostic agents in the real machinery of radiology and pathology: PACS, DICOM worklists, the reporting cycle and whole-slide imaging. It sets out the published external validation record honestly, and takes a position on triage-and-prioritise against autonomous read.

• 1.The Imaging Substrate: PACS, DICOM Worklists and the Report as Deliverable10m
• 2.The External Validation Record: Why Imaging AI Travels Badly Between Sites15m
• 3.Lab: Build a Site Readiness Dossier for a Chest X-Ray Triage Agent25m

Module 1 treated images, where the input is a discrete study a radiologist requested. Continuous physiological signals behave differently: they never stop, most of what arrives is artefact, and the ward is already saturated with alarms. This module covers acquisition physics, the 12-lead versus single-lead distinction, and why alarm burden, not detection accuracy, is the operational constraint on any monitoring agent.

• 1.Sampling Rates, Lead Sets and Artefact: What a Waveform Actually Contains10m
• 2.Alarm Burden, Ward Telemetry and the Consumer Wearable Divide10m
• 3.Lab: Build an Alarm Ledger and Net-Alert Budget for a Telemetry Agent20m

Diagnosis is a decision under uncertainty, not a classification task, and an agent that returns a confident ranked list without asking what a clinician would have asked has solved a different problem. This module covers illness scripts, pre-test probability, likelihood ratios and decision thresholds, then builds a question-first differential agent tested against anchoring and premature closure.

• 1.Illness Scripts, Semantic Qualifiers and Why a Ranked List Is Not a Differential10m
• 2.Pre-Test Probability, Likelihood Ratios and the Questions the Agent Never Asked15m
• 3.Lab: Build a Question-First Differential Agent with an Anchoring and Closure Audit25m

Sepsis prediction is the one area of clinical AI with a public, peer-reviewed failure and a published counter-example. This module sets NEWS2, MEWS and the Sepsis-3 definition as the baselines any agent must beat, reads the Epic Sepsis Model external validation and the alert-burden debate accurately, and produces the silent-run protocol the module 15 capstone will be built against.

• 1.Sepsis-3, NEWS2 and MEWS: The Baseline a Prediction Agent Must Beat15m
• 2.What the Epic Sepsis Model External Validation Taught the Field15m
• 3.Lab: A Silent-Run Protocol and Alert Budget for a Ward Sepsis Agent25m

Staging is a classification act governed by a rulebook that now versions per disease site; treatment matching is eligibility logic over guidelines before it is machine learning; response assessment is specified arithmetic. This module builds oncology agents that respect that structure, and argues they belong in tumour board preparation rather than in the room.

• 1.Staging as a Versioned Rulebook: AJCC Version 9, UICC TNM and the Synoptic Report15m
• 2.Treatment Matching as Eligibility Logic: NCCN Categories, ESCAT and the Trials Problem15m
• 3.Lab: Building an MDT Preparation Pack with a RECIST 1.1 Response Timeline25m

Coding automation fails on a distinction rather than on accuracy: SNOMED CT is a terminology for capture and reasoning, ICD is a classification for counting and paying, and the conversion between them loses information deliberately. This module covers where ICD-11 genuinely stands against ICD-10-CM billing reality, and why an ambient documentation agent that improves completeness is also, unavoidably, a revenue instrument.

• 1.Terminology Versus Classification: SNOMED CT, LOINC and the ICD-11 Reality Check15m
• 2.Ambient Documentation Agents, E/M Levels, Risk Adjustment and the Upcoding Problem15m
• 3.Lab: Building an Autocoding Assurance Pack with Provenance, Query Compliance and a Code-Shift Baseline25m

Interaction checking exists in every prescribing system and clinicians override the overwhelming majority of its alerts. This module treats screening as a subtraction problem: severity grading, pharmacokinetic versus pharmacodynamic mechanism, organ-function modifiers, and the patient-specific context generic checkers cannot reach — inside the licensing limits of commercial drug knowledge bases.

• 1.Why Ninety Per Cent of Interaction Alerts Are Overridden15m
• 2.Pharmacokinetic, Pharmacodynamic and Patient-Specific Layers15m
• 3.Lab: Build a DDI Screening and Suppression Register25m

Genomic agents fail where sequencing succeeds: interpretation, not variant calling, is the bottleneck. This module works through the ACMG/AMP framework, ClinVar's conflicting submissions, CPIC pharmacogenomics and the variant of uncertain significance as the dominant practical problem, alongside the consent and family-disclosure duties that genetic data uniquely carries in the Gulf.

• 1.Variant Calling Is Largely Solved; Interpretation Is the Bottleneck10m
• 2.Pharmacogenomics, Secondary Findings and the Relatives Who Never Consented10m
• 3.Lab: Build a Variant Evidence Dossier Agent with an Explicit Non-Resolution Contract20m

Eligibility criteria are prose written for clinicians, and most of them are ambiguous, temporally unanchored or simply unanswerable from the record. This module treats representation rather than matching as the real problem, works through ClinicalTrials.gov's quality limits and the pre-screening versus screening accountability line, and confronts the inequity an efficient screener produces on thin data.

• 1.Prose Criteria, Computable Predicates and the Criteria That Are Neither15m
• 2.Pre-Screening, Screening and the Signature: Registry Data, Accountability and Regulatory Position15m
• 3.Lab: Build a Criterion Computability Register and Screening Equity Audit for Three Live Trials25m

An agent that produces a plausible synthesis without a reproducible search has written a literature-flavoured essay, not evidence. This module anchors autonomous synthesis in systematic review method — PRISMA 2020 and PRISMA-S, dual independent screening, RoB 2, ROBINS-I and GRADE — and treats retraction, preprint and citation-integrity handling as engineering requirements rather than editorial hygiene.

• 1.Reproducible Search Strategy: PRISMA-S, Recall and the Protocol That Comes First10m
• 2.Dual Screening, RoB 2 and ROBINS-I: Where Agent Judgement Is Defensible15m
• 3.Lab: Build a Citation Integrity Gate and Reproducible Search Ledger25m

APACHE, SAPS and SOFA were validated for case-mix adjustment and unit-level benchmarking, yet are routinely read as individual prognoses at the bedside. This module separates the two uses, reads MIMIC and eICU for what they can and cannot support, confronts the self-fulfilling-prophecy problem, and produces a written use boundary for any output touching escalation or end-of-life care.

• 1.Severity Scores as Case-Mix Adjustment: What APACHE, SOFA and SAPS Were Validated For15m
• 2.MIMIC, eICU and Treatment Leakage: Why an ICU Outcome Model Can Validate Itself15m
• 3.Lab: A Score Automation Spec and Prognostic Use Boundary for an ICU Agent25m

Mental health is the one clinical domain where several legislatures have already answered the autonomy question, and the answer was no. This module treats PHQ-9 and GAD-7 as what they are, explains why statistical suicide-risk prediction fails at the individual level, and builds the service design and escalation architecture that Nevada AB 406, the Illinois WOPR Act and California SB 243 actually permit.

• 1.PHQ-9 and GAD-7: What a Validated Score Measures and What It Cannot Decide15m
• 2.Risk Stratification, Base Rates and the Statutes That Have Already Answered This15m
• 3.Lab: Specify a Screening-to-Clinician Escalation Pathway and Its Safety Case25m

Where a clinical agent's output lands decides whether it changes care, and no amount of model quality compensates for the wrong channel. This module treats autonomy as a measured property of a deployment rather than a label, covers alert override, cognitive load, accountability for following or overriding, and trust calibration, and argues that a system followed almost always is autonomous whatever the interface claims.

• 1.Where the Output Lands: Inbox, Worklist, Chart or Interruptive Alert15m
• 2.Copilot as Governance Posture: Measuring Autonomy by Behaviour, Not Interface15m
• 3.Lab: Build an Autonomy Ledger and Placement Contract for a Deterioration Agent25m

Validation methodology for the whole course sits here. The module fixes which reporting standard applies at which stage, why retrospective performance is a screening result rather than evidence, and how silent-mode running, drift triggers and a Predetermined Change Control Plan turn a deployed clinical agent into something that can be improved without being re-litigated.

• 1.Which Reporting Standard Applies: SPIRIT-AI, CONSORT-AI, TRIPOD+AI and DECIDE-AI15m
• 2.Silent-Mode Running, Drift Triggers and the Predetermined Change Control Plan15m
• 3.Lab: Write a Silent-Mode Protocol and Change Control Plan for a Deterioration Agent25m

The capstone turns fourteen modules of artefacts into one document a hospital clinical governance committee could vote on. It builds the agent against a comparator the March 2026 Surviving Sepsis Campaign guidelines have just re-endorsed, and treats the stopping rule and the refusal register as the parts that decide the outcome.

• 1.The Capstone Brief: A Guideline-Endorsed Incumbent and Where the Burden of Proof Sits10m
• 2.Designing the Build: EHR Integration, Workflow Placement and the Stopping Rule10m
• 3.Lab: Assemble and Defend the Sepsis Agent Deployment Case, Stopping Rule and Refusal Register25m