Patient Engagement, Virtual Care & AI Health Companions

An agent that supports a clinician and one that speaks to a patient are different products, because in the second case no clinical judgement sits between the output and the person. This module sets the ethical perimeter for the course: where simulated empathy is honest, why disclosure raises candour rather than suppressing it, and what EU AI Act Article 50 requires from 2 August 2026.

• 1.The Missing Clinician: Support Agents Versus Patient-Facing Agents15m
• 2.Simulated Empathy, Disclosure and the Article 50 Duty15m
• 3.Lab: A Therapeutic Boundary Charter for a Patient-Facing Health Companion25m

Module 1 established therapeutic boundaries for patient-facing agents. This module applies them to the three conditions where remote monitoring is most commercially attractive, separating the agent that reminds from the agent that titrates, and treating escalation thresholds, review capacity and the unread reading as the design problems that decide whether a service delivers care or liability.

• 1.Reminder, Coach, Titrator: Where the Regulatory Line Falls15m
• 2.Remote Monitoring Data Flows, Escalation Thresholds and the Unread Reading15m
• 3.Lab: Build an Escalation and Review Specification for a Three-Condition Monitoring Service25m

Non-adherence is usually a rational response to cost, side effects, complexity or doubt about necessity, and a reminder addresses none of them. This module sets out barrier taxonomies, what NLP can and cannot extract from patient language, validated adherence measurement and its failure modes, and the surveillance and disclosure questions that decide who sees the resulting number.

• 1.Why Non-Adherence Is Rational: Barrier Taxonomies and the Limits of the Reminder15m
• 2.Measuring What You Cannot See: PDC, Self-Report, Electronic Monitoring and Who Reads the Result15m
• 3.Lab: Build a Barrier-Resolution Routing Pack for a Chronic-Care Adherence Agent20m

Nevada, Illinois and California have already legislated on AI in mental health, and the answer was that a machine may not be the clinician. This module works out the design space that survives that answer — psychoeducation, structured self-monitoring, between-session practice support and routing to a human — and treats safety rails as a tested escalation architecture rather than a disclaimer.

• 1.The Line Three Legislatures Drew: Prohibited Care, Permitted Support15m
• 2.Manualised Programmes, Open-Ended Conversation and the Adherence Problem15m
• 3.Lab: A Scope-and-Escalation Specification with a Safety Case for a Between-Sessions Companion25m

Enhanced recovery pathways move the observation problem out of a staffed ward and into an unsupervised home, where most deterioration is a new event rather than a continuing one. This module examines what home-collected data can carry, what wound photography cannot, and why the escalation route to a clinician is the product rather than the detector.

• 1.Enhanced Recovery, Early Discharge and the Complications That Present at Home10m
• 2.Wound Photographs, Patient-Reported Outcomes and the Limits of Home-Collected Data10m
• 3.Lab: Building an Escalation Map and Readmission Measurement Plan for One Surgical Pathway25m

Coaching services fail at maintenance rather than onboarding, and eating is the one domain where a highly engaging product can injure the person using it. This module works from COM-B and behaviour change technique taxonomies to a design position: specify behaviours rather than body outcomes, treat Ramadan as the organising event of the Gulf year, and place the prohibited mechanics under clinical governance.

• 1.Diagnosing the Behaviour Before Building the Agent: COM-B, BCT Taxonomies and the Outcome Trap15m
• 2.Habit, Attrition and the Gulf Calendar: Designing for Twelve Months Rather Than Twelve Weeks15m
• 3.Lab: Building a Nutrition Coaching Agent Design Dossier with a Prohibited-Mechanics Register and Escalation Protocol25m

This is the highest-stakes patient-facing domain in the course, for two reasons no earlier module shares: the person typing is frequently not the patient, and antenatal deterioration leaves an escalation window short enough that a triage layer becomes the harm. It sets the consent, children's data and escalation positions that follow, and treats developmental screening as an instrument with a pathway obligation attached.

• 1.When the User Is Not the Patient: Parental Consent, Children's Data and the Evolving Rights of the Child15m
• 2.Antenatal Escalation: Pre-eclampsia, Reduced Fetal Movement and the Cost of a Triage Layer15m
• 3.Lab: A Screening Pathway Contract for a Child Development Agent25m

Geriatric agents operate where clinical multifactorial risk, home surveillance and fluctuating decisional capacity meet. This module covers validated falls assessment and the passive sensing versus wearables trade-off, cognitive screening and its administration conditions, consent modelled as a continuing state rather than a signature, and the caregiver as a second user with her own burnout.

• 1.Falls as a Multifactorial Risk: Assessment, Passive Sensing and the Acceptability Trade-Off15m
• 2.Cognitive Decline, Capacity and Consent as a Continuing State15m
• 3.Lab: Dual-User Protocol — Caregiver Load Contract and Companionship Boundary Statement25m

Symptom checkers carry the longest published evaluation record of any patient-facing health AI, and it is not a flattering one. This module reads that record honestly, treats the safety-versus-utility operating point as a governance decision requiring a named owner rather than a threshold a data scientist picks, and argues that care navigation, not diagnosis, is where the deliverable value sits.

• 1.What the Symptom Checker Evaluation Record Actually Shows15m
• 2.The Safety-Utility Operating Point: Triage Frameworks, Over-Triage Cost and Named Ownership15m
• 3.Lab: Operating-Point Charter and Care-Navigation Routing Map for a Virtual Triage Front Door25m

Arabic is not one language, and a model evaluated on Modern Standard Arabic tells you little about an Emirati patient's experience. This module covers dialect variation, code-switching, bidirectional rendering, numeral forms and transliterated name matching, and treats the multilingual expatriate population of the Gulf as an explicit scoping decision rather than an accident.

• 1.Which Arabic? Modern Standard, Gulf Dialect and the Code-Switched Clinic15m
• 2.What Serving a Language Costs: Population Inventory, Bidirectional Text and Name Matching15m
• 3.Lab: Language Service Scope Declaration and a Gulf Clinical Arabic Evaluation Set25m

Patient agents are only as good as the record they can see and the channel they are allowed to use. This module covers FHIR R4 patient-facing APIs and SMART patient launch, what Malaffi, NABIDH and Riayati actually expose, the write-back question governance committees keep refusing, and the liability that attaches to patient-generated data nobody has agreed to review.

• 1.Reading the Record: FHIR R4, SMART Patient Launch and What the UAE Exchanges Actually Expose15m
• 2.Write-Back, Patient-Generated Data and the Duty That Arrives With the Payload15m
• 3.Lab: The Integration Contract — Device Grade Matrix, PGHD Register and a Write-Back Decision Record25m

Consent doctrine assumes risks that can be listed and a clinician who can be blamed; an autonomous patient agent supplies neither. This module separates consent to treatment from consent to data processing, treats consent as a continuing state rather than an enrolment signature, and works through how liability is actually allocated between health system, vendor and clinician of record while the courts stay silent.

• 1.Disclosing a System Whose Behaviour Cannot Be Enumerated15m
• 2.Two Legal Acts, Three Possible Defendants: Data Processing Consent and Liability Allocation15m
• 3.Lab: Layered Consent Instrument and Liability Allocation Matrix for a Diabetes Companion20m

Oncology holds the strongest randomised evidence that patient-reported symptom monitoring improves outcomes, and also the clearest limits on what that evidence licenses. This module reads STAR, Denis, eRAPID and PRO-TECT precisely, fixes the instruments and the escalation pathway that make PRO collection worth anything, and draws a hard boundary around prognosis, dying and distress.

• 1.What STAR, Denis, eRAPID and PRO-TECT Actually Showed About Symptom Monitoring15m
• 2.PRO-CTCAE, Distress Screening and the Line an Oncology Agent Must Not Cross15m
• 3.Lab: Symptom Tracking and Escalation Specification for a Day-Oncology Unit25m

Evaluation is where patient agent programmes are most often flattered by their own instruments. This module separates process, surrogate and patient-important outcomes, sets out pragmatic, cluster and stepped-wedge designs for services that cannot randomise individuals, treats the digital placebo and differential attrition as design problems rather than caveats, and makes equity a pre-specified endpoint.

• 1.Process, Surrogate and Patient-Important: Specifying an Outcome You Can Defend10m
• 2.Pragmatic, Cluster and Stepped-Wedge Designs and the Digital Placebo Problem15m
• 3.Lab: Attrition Dossier and Equity Pre-Specification for a Diabetes Companion Evaluation25m

The capstone assembles the course into one signable artefact: a diabetes companion for a DOH-licensed facility in Abu Dhabi, bound by the Responsible AI Standard and by health data localisation under Federal Law No. 2 of 2019. It treats a majority-expatriate cohort as the design problem, and adds the document every earlier module implied — a refusal register naming what the companion will not do.

• 1.Reading the Brief: What Abu Dhabi's Constraint Set Decides Before You Design15m
• 2.Who the Companion Serves, What It Refuses, and Why Both Are Architecture15m
• 3.Lab: Assemble the Diabetes Companion Design Dossier and Refusal Register for a DOH-Licensed Facility25m