Trust centre

Responsible AI and standards

Each item has a status and a link to its evidence in the app.

DECIDE-AI: 17 AI-specific reporting topics (paraphrased)

Intended use, patients and usersScreening support for 50 to 75 in primary care. Not diagnosisEvidenceMet
AI system and versionEngine, router rules and narrator versions shown on every screenEvidenceMet
Setting and workflow integrationSahatna tile, family-physician worklist, referral to Malaffi (the shared health record)EvidenceMet
Participants and recruitmentPublic Health Centre invitation by age band. Pilot sites namedEvidencePartly
User trainingNurse and family-physician training in the pilot kitEvidencePartly
How outputs are shown and usedState in words first. Recorded, calculated and AI layers kept apartEvidenceMet
Human oversight of AI outputApprove, edit or hold with a coded reason and a noteEvidenceMet
Outcomes: utility, safety, human factorsPilot KPIs and harms defined in advanceEvidenceMet
Safety, errors and adverse eventsUrgent stop audited. Monthly safety reviewEvidenceMet
Human factors and usabilityScripted persona runs now. Usability with 40 people in phase 1EvidencePartly
Changes to the AI during the studyEvery change versioned, re-scored with make eval, signed offEvidenceMet
Protocol deviationsRecorded in the audit log and the monthly safety reviewEvidencePlanned
Missing and poor-quality dataSignal level drops. Missing never treated as zeroEvidenceMet
Subgroups and fairnessLanguage, schooling and help-needed rows with denominatorsEvidenceMet
Comparison with current careFixed cut-off, 5% drop and single-check change comparedEvidenceMet
Learning curvePractice check first. First-four-checks view for cliniciansEvidenceMet
Ethics, involvement and opennessEthics submission before the pilot. Code and eval sets reproducibleEvidencePartly

TRIPOD+AI: key reporting items (paraphrased)

Title and abstract identify the model and its useDocs and model cardsEvidenceMet
Objectives and intended useStated with limitsEvidenceMet
Data sources and participantsSimulated cohort with seed and data dictionaryEvidenceMet
Outcome definitionTrue decline as defined in the cohort assumptionsEvidenceMet
Inputs and how they are measuredFour activities plus readiness gatesEvidenceMet
Sample size5,000 people. Subgroups under 30 in decline markedEvidenceMet
Missing data handlingLive 24-case input suiteEvidenceMet
Analysis methodsReliable change with practice adjustment, confirm rule, CUSUM (cumulative sum)EvidenceMet
Performance with confidence intervalsWilson intervals on every proportionEvidenceMet
Fairness across groupsCoverage and specificity per groupEvidenceMet
External validationAgreement with the MoCA (Montreal Cognitive Assessment) at the IFHAS screening programme, pilot phase 1EvidencePlanned
Code and data availabilitymake eval reproduces every numberEvidenceMet

STARD-AI (diagnostic accuracy)

Reference standardSpecialist assessment in the pilotEvidencePlanned

DoH Responsible AI standard (clause matrix)

TransparencyProvenance chips on every block. Model + version shownEvidenceMet
Human oversightApprove / edit / reject with a required noteEvidenceMet
Data protection and residencyHosted in the UAE regionEvidenceMet
FairnessEducation- and language-adjusted norms. Equity panelEvidencePartly
AccountabilityHash-chained audit log. Kill switchesEvidenceMet

ADHICS (health information security)

Access controlRole-based access. Role guard. Re-auth for sensitive actions (planned)EvidencePartly
Audit and loggingAppend-only, exportableEvidenceMet
Incident managementRunbooks for narrator off, rule change and Malaffi outageEvidencePartly
Continuity and backupNightly backup and a monthly restore drill in productionEvidencePlanned
Third-party and cloudVendor list with residency per data classEvidenceMet

DPIA

Privacy impact assessmentDraft prepared; DPO review in pilot set-upEvidencePartly