Agentic AI Foundations for Healthcare & Life Sciences

Healthcare has run predictive models for years; agentic systems are a different proposition, because software that holds a goal and acts through clinical systems can cause harm that no credit memo reverses. This module establishes what an agent actually is, the four-layer architecture every clinical agent shares, and how to place a candidate workflow on a reversibility map before anyone writes code.

• 1.Why a Clinical Agent Is Not a Clinical Model15m
• 2.The Four-Layer Clinical Agent: Perception, Reasoning, Action, Oversight15m
• 3.Lab: A Clinical Agent Opportunity Canvas and Reversibility Map25m

Establishes how medical language models are actually built — domain pretraining, continued pretraining and adaptation — and what the published Med-PaLM, MEDITRON, BioGPT and BioMedGPT results do and do not demonstrate. The module's core argument is that examination accuracy is a weak proxy for reasoning about an undifferentiated patient, and it gives you the evidence to say so.

• 1.Three Routes to a Medical Model: BioGPT, MEDITRON and the Med-PaLM Lineage10m
• 2.The Vignette Gap: Why 86% on MedQA Is Not 86% on an Undifferentiated Patient10m
• 3.Lab: Build a Clinical LLM Suitability Dossier for a Named Care Pathway25m

The distinction between supporting a clinician and deciding for one is the line that determines whether software is a regulated device, and FDA moved that line on 6 January 2026. This module works through the statutory criteria, grades clinical agent autonomy against what tool-calling actually changes, and produces a defensible per-capability position statement.

• 1.Where the Line Falls: The Four Statutory Criteria for Non-Device Clinical Decision Support15m
• 2.Graded Autonomy: From Retrieval to Action and the Point Where the Default Inverts15m
• 3.Lab: Classify a Clinical Agent's Autonomy and Draft Its Device-Status Position Statement25m

Coordination across disciplines is where tasks fall between owners, and splitting a workflow across agents manufactures new boundaries of exactly that kind. This module reads SBAR, the ward round and the multidisciplinary team as working designs, then takes a position on when a supervisor agent genuinely improves on one agent with more tools.

• 1.What SBAR, the Ward Round and the MDT Already Know About Handoffs10m
• 2.Supervisor, Blackboard, or One Agent With More Tools?10m
• 3.Lab: A Handoff Contract and Orphan-Task Register for Discharge Coordination25m

Clinical agent projects rarely fail at the model. They fail at the interface, where HL7 v2 feeds, mismatched FHIR profiles, expiring pagination tokens and local code systems consume the schedule. This module covers the integration surface an agent actually meets and how to assess it before committing to a build.

• 1.HL7 v2, FHIR R4 and the Interface Estate an Agent Inherits10m
• 2.SMART on FHIR, CDS Hooks and Bulk Export: How Agents Reach the Record15m
• 3.Lab: A FHIR Data Feasibility Assessment for a Clinical Agent Use Case20m

Medical retrieval is not general retrieval: guidelines conflict between bodies, evidence carries a certainty grade and a shelf life, retracted papers stay in the corpus, and drug databases are licensed in ways that constrain what you may embed and cache. This module covers corpus design, licence-aware architecture, and where citation-grounded generation actually moves clinical risk.

• 1.Why Medical Retrieval Is Not General Retrieval: Conflicting Guidelines, Evidence Grades and Retracted Sources15m
• 2.Drug Databases, Licence Terms and the Limits of Citation-Grounded Generation15m
• 3.Lab: Build an Evidence Provenance Register and Conflict Policy for a Formulary Support Agent25m

Privacy law is where published material on health AI is most often out of date, and this module is the course's single source of truth on it. It sets out what HIPAA requires today rather than what a proposed rule would have required, what Article 9 and the EHDS change for European deployments, and why the Abu Dhabi DOH Responsible AI Standard is the instrument that decides Gulf architecture.

• 1.HIPAA for Health Agents: PHI Boundaries, De-identification and the Rule That Is Not Law15m
• 2.GDPR Article 9, the EHDS and the DOH Abu Dhabi Responsible AI Standard15m
• 3.Lab: A Jurisdictional Privacy Control Matrix for a Discharge-Summary Agent20m

The regulatory boundary for clinical software moved twice in eighteen months: the Predetermined Change Control Plan guidance became final in December 2024, and the Clinical Decision Support guidance was rewritten on 6 January 2026. This module establishes which agents are devices, which pathway applies, and how change is authorised across US, EU and Gulf regimes.

• 1.When a Clinical Agent Becomes a Regulated Device: Classification and the 2026 CDS Rewrite15m
• 2.Predetermined Change Control Plans and the EU Conformity Route for Adaptive Models15m
• 3.Lab: Building a Regulatory Pathway Determination Pack for a Clinical Agent25m

Bias in clinical AI is four distinct failures - label, measurement, representation and deployment - and each needs a different remedy. This module works through the documented cases, establishes that competing fairness metrics cannot all be satisfied, and treats the choice between them as a clinical judgement. It then confronts what an imported model meets in a majority-expatriate Gulf population.

• 1.How Bias Enters a Clinical Model: Cost Proxies, Pulse Oximeters, eGFR and Skin Tone15m
• 2.Choosing Which Fairness to Enforce: Impossibility Results and Gulf Population Transportability15m
• 3.Lab: Building a Subgroup Performance and Fairness Dossier for a Clinical Agent25m

Electronic prescribing already ran this experiment: published override rates for drug safety alerts range from 49% to 96%. This module treats clinical agent safety as safety engineering rather than AI ethics, working through the hierarchy of controls, forcing functions, Swiss cheese independence and alert budgets, and argues that human-in-the-loop as usually implemented transfers liability rather than reducing risk.

• 1.Automation Bias, Alarm Fatigue and Why the Override Button Is Not a Safeguard15m
• 2.Forcing Functions, Graceful Degradation and the Independence of Defensive Layers15m
• 3.Lab: Building a Patient Safety Case and Override Instrumentation Plan25m

Saliency maps are the default deliverable, and the published evidence on them is unkind. This module separates the three artefacts that hide under "explainability" — a conformity record for regulators, a decision aid at the point of care, and an assurance object for procurement — and replaces admiration of an explanation with a test it can fail.

• 1.Three Explanations, Three Audiences: Conformity, Decision and Assurance15m
• 2.What the Evidence Says About Saliency Maps, and Rudin's Case for Interpretable Models15m
• 3.Lab: Specifying Radiologist-Facing Output and a Discordance-Weighted Acceptance Test25m

An agent connected to the record adds attack surface a predictive model never had: instructions arriving inside clinical text, tool credentials that aggregate into standing privilege, retrieval as an egress path, and a supply chain of weights and indexes nobody inventories. This module maps those threat classes onto the OWASP 2026 lists, IEC 81001-5-1, FDA premarket expectations and the NIST AI RMF, and builds a defensible control register.

• 1.What an Agent Adds to the Attack Surface: Injection, Tool Privilege and Retrieval15m
• 2.Framework Coverage and Its Gaps: IEC 81001-5-1, FDA Premarket Cybersecurity and the NIST AI RMF15m
• 3.Lab: A Threat Model and Control Register for an EHR-Connected Clinical Agent25m

Federated learning is the most oversold privacy technique in health AI, and this module separates what it delivers from what it only appears to. It covers FedAvg and the published multi-institution evidence, the inversion and membership inference attacks that make 'no data left the building' an incomplete claim, the accuracy cost of the real mitigations, and why a cross-emirate federation still sits under DOH, DHA and UAE localisation rules.

• 1.What Federated Averaging Actually Moves: Cross-Silo Topology, Non-IID Data and the Published Evidence15m
• 2.Gradient Inversion, Membership Inference and Two Mitigations That Both Cost Accuracy15m
• 3.Lab: A Federated Feasibility and Disclosure Assessment for an Abu Dhabi and Dubai Sepsis Model25m

Most clinical AI evidence fails at the same three points: predictive value is quoted at the wrong prevalence, calibration is never reported, and utility is asserted rather than tested. This module works the arithmetic of sensitivity, specificity, predictive value and net benefit, and sets the validation, reporting and drift-monitoring standards a deploying hospital should demand.

• 1.Operating Characteristics and the Prevalence Trap: Sensitivity, Specificity and Predictive Value10m
• 2.Discrimination, Calibration and Net Benefit: What AUC Cannot Tell You15m
• 3.Lab: An Evaluation and Drift-Monitoring Dossier for a Sepsis Early-Warning Agent25m

Everything built across the previous fourteen modules converges into one submission a hospital AI governance committee could vote on. The module assembles the reversibility map, device-status determination, privacy matrix, safety controls and evaluation plan into a deployment plan for a DOH-licensed Abu Dhabi hospital, and treats what the design refuses to automate as seriously as what it delivers.

• 1.Assembling the Submission: Eleven Artefacts, Four DOH Pillars, One Decision Sheet15m
• 2.The Refusal Register: What This Deployment Will Not Automate, and Why15m
• 3.Lab: Assemble and Defend an Agentic AI Deployment Plan for a DOH-Licensed Hospital25m