Loading · جارٍ التحميل
Documentation
Thabat supports early detection of change in thinking for adults aged 50 to 75 in Abu Dhabi primary care. It compares each person with their own usual range and routes confirmed change to the family doctor. It does not diagnose, stage or treat any condition, and a clinician makes every referral decision. All known limits
Under the IMDRF software-as-a-medical-device framework, Thabat informs clinical management of a serious condition, which places it in category II. In the UAE it would be assessed as medical device software by the Emirates Drug Establishment, and run under DoH Abu Dhabi standards for AI in healthcare and ADHICS. The language model only words results. It never scores, so it sits outside the decision logic.
| Situation | What the system does | Who decides |
|---|---|---|
| Stroke signs, head injury on blood thinners, self-harm thoughts | Stops the check, shows 998, alerts the clinic | Emergency services and the clinic |
| Two checks in a row below the usual range | Holds the result and drafts a message | Family doctor approves, edits or holds |
| Family member reports change | Sends the record for review whatever the score | Family doctor |
| Not a usual day (sleep, new medicine, unwell) | The check is not counted, and a repeat is booked. | Automatic, visible to the clinician |
| Too little signal (missing items, new device) | Lowers the signal level and says what is missing | Clinician weighs it |
Every set has a SHA-256 fingerprint. Run make eval to reproduce every figure. Evaluation lab
The product path uses only open licences (Apache-2.0, MIT, BSD, OFL) or our own code, which is proprietary (all rights reserved). Non-commercial models (BiMediX2, MedAgentSim) are for evaluation only. Model registry
Paper and digital MoCA (Montreal Cognitive Assessment), Cognivue, CANTAB, Cogstate, and Inception's clinic-side brain health lab, which Thabat complements in the community. Table
| Field | Meaning | Unit |
|---|---|---|
| composite | Mean of memory, speed and switching, each scaled to norms | SD |
| baseline | Mean of the first three counted checks | SD |
| rci | Reliable change: (check − baseline − practice) ÷ SEdiff | z |
| cusum | Running drift below baseline in noise units | score |
| not counted | Check not counted because of a readiness answer | yes/no |
| signal | Full, reduced or low, from missing items | level |
| cohort seed | Simulation seed for the validation cohort | 20261113 |
| Item | Plan |
|---|---|
| Sites | Two family health centres in Al Ain and one pharmacy kiosk in Mussafah; 12 months |
| Owners | Public Health Centre programme lead. Clinical lead (family physician); DoH digital health for Sahatna and Malaffi, the shared health record |
| KPIs | Reviewed within 30 days. Documented outcome. False referrals per 1,000. Lost to follow-up. Median days to assessment. Completion gap by group |
| Success criteria | 85% of referrals reviewed within 30 days. False referrals under 10 per 1,000 checks. Completion gap between groups under 10 points. No urgent case missed in audit |
| Phase 1 validation | Agreement with the MoCA (Montreal Cognitive Assessment) at the IFHAS screening programme. Local Gulf norms. Usability with 40 people aged 50 to 75 and their families |
| Data source | Sahatna check events joined to Malaffi encounters by the Public Health Centre data team |
| Reviewed by | Clinical lead and the Public Health Centre programme lead before the ethics submission |
The clinical adviser is a family physician from a pilot site, appointed by the clinical lead before the pilot starts. They sign the clinical-expert statement for the change rule, the deferral rules and the person-facing wording, and chair a monthly safety review of urgent stops and held results.
| Risk | Control | Residual |
|---|---|---|
| Health data leaves the UAE | All stores and the production language model in the UAE region. Only facts, never names or IDs, reach the model | Low |
| Family sees more than allowed | Per-item scopes checked on the server. Revocation ends access at once | Low |
| Re-identification from programme counts | Groups under 10 hidden. No names on programme screens | Low |
| A result reaches the person before review | Doctor-band results are held until a clinician approves | Low |
| Tampering with the record of who did what | Append-only, hash-chained audit log, checked on every read | Low |
| Threat | Where | Control |
|---|---|---|
| Spoofing | Sign-in | UAE PASS only for people. Staff SSO. Re-confirmation before export |
| Tampering | Results and audit | Server-side scoring. Hash-chained audit |
| Repudiation | Clinician decisions | Every decision needs a note and is audited with the actor |
| Information disclosure | Family and programme views | Scope checks on the server. Small-cell suppression |
| Denial of service | Read-aloud and language model | Rate limits. Switches that fall back to approved templates |
| Elevation of privilege | Console roles | Role checks on every staff route and API |
Today the norms come from published literature. In the pilot, each site keeps its own data and a federated update builds Gulf norms by age, schooling and language without moving records between sites (federated learning). Norms version and site count show on every result.
Sahatna launch, UAE PASS, Malaffi HL7 and FHIR, IFHAS import, wearables, Inception lab FHIR feed, MBZUAI research export. Contracts
We use your check answers to give you results and a plan and to share them with your family doctor. Family and research use only with your separate consent. You can download your data or ask us to delete it.
We aim for WCAG 2.2 AA, with 18 px minimum text, 48 px touch targets, high contrast, read-aloud and full Arabic right-to-left support. Tell us about any barrier through Help.
This service supports screening. It does not diagnose or replace your doctor. In an emergency call 998.