Trust centre

Assumptions and limits

What this build does not show, and what would change our answers.

Not measured here

Accuracy in practice. Current numbers come from the validation cohort and authored test cases. Pilot phase 1 measures it in clinics.
Validity in Gulf Arabic speakers and low-education groups (norms are literature priors).
Agreement with the MoCA (Montreal Cognitive Assessment) at the IFHAS screening programme, and with specialist assessment.
Real usability with people aged 50–75 (simulated personas only).
Production access to Malaffi (the shared health record), Sahatna and UAE PASS needs each owner's onboarding and approval.

Test-retest reliability 0.82 and practice effect 0.30 SD (published ranges).
A confirm check 2–4 weeks after a flag is acceptable to people and clinics.
Family physicians can absorb about 2.6k extra reviews a year at current thresholds.

IFHAS MoCA accuracy study (1,000 people, two centres, sized from Buderer). Pilot protocol
Education-banded Gulf norms from the first 3,000 people.
SUS and teach-back with 40 people aged 50–75, Arabic and English.

Risk	Signal we watch	Response	Owner
Too many referrals	Referrals above 900 per 10,000 checks for 2 weeks	Governance review of the threshold. Capacity simulator rerun	Programme lead
Missed urgent message	Any urgent case found later in clinician review	New rule within 24 hours. Held-out set extended. Incident report	Clinical safety officer
Unfair results across groups	Wrong flags per 1,000 differ by more than 5 between groups	Local norms re-fitted. Group-specific review before release	Data science lead
Explanation drifts from the result	Verifier blocks above 2% of explanations in a week	Switch to the fixed template. Prompt and model reviewed	Platform admin
Device or app change shifts scores	Mean score jump after a release or device change	Sessions held from baselines. Bridging study before reuse	Platform admin

Evidence grade. Gait and sleep links to later decline come from cohort studies (moderate). They are not diagnostic accuracy for one person.
Device validity. iPhone walking speed is valid in older adults; Walking Steadiness was built on a US cohort. Neither is checked in Emiratis aged 50 and over.
Who owns one. About a third of UAE internet users own a smartwatch, skewing younger and urban. Passive walking speed is iPhone only. Same thresholds and dates with or without a device.
Local habits. Car-based days, indoor summers and Ramadan shift baselines, so Ramadan nights get their own sleep baseline and summer walking is compared with last summer.
Thresholds (0.10 m/s, 5 h, 35 dB HL) come from Western cohorts. They stay fixed until phase 1 tests them against IFHAS MoCA and reports the false re-check burden.
Never: a device signal alone moves a band, starts a referral, or makes a check later.

Risk

Signal we watch

Response

Owner

Too many referrals

Referrals above 900 per 10,000 checks for 2 weeks

Governance review of the threshold. Capacity simulator rerun

Programme lead

Missed urgent message

Any urgent case found later in clinician review

New rule within 24 hours. Held-out set extended. Incident report

Clinical safety officer

Unfair results across groups

Wrong flags per 1,000 differ by more than 5 between groups

Local norms re-fitted. Group-specific review before release

Data science lead

Explanation drifts from the result

Verifier blocks above 2% of explanations in a week

Switch to the fixed template. Prompt and model reviewed

Platform admin

Device or app change shifts scores

Mean score jump after a release or device change

Sessions held from baselines. Bridging study before reuse

Platform admin