In health software, accuracy is an architectural property, not a marketing claim. An is engineered so that every conclusion can be traced to its data, its method and its measured confidence.
A credible health product requires more than data connectors. An's stack combines three families of technology:
Apple HealthKit, Google Health Connect, wearable-data aggregation and home measurement devices feed a single record per person, normalized to FHIR R4 — the interoperability standard behind Google Cloud Healthcare, Azure Health Data Services and AWS HealthLake. Partner infrastructure connects without rework.
Risk estimation requires anchors: population statistics (the WHO Global Health Observatory and World Bank health indicators); published clinical risk frameworks, such as SCORE2 for cardiovascular risk and FINDRISC for diabetes; drug reference data (OpenFDA); and the clinical guidelines that individual alerts cite as their basis.
Our own engine, in three layers: validated risk scores computed deterministically; deviation detection against each person's longitudinal baseline; and calibrated machine learning, trained and evaluated on open, labeled datasets, which reports an uncertainty range with every estimate — never a bare number.
The engine is evaluated on public, labeled datasets (including NHANES and PhysioNet resources) under versioned data and a fixed methodology. Sensitivity, specificity and calibration are published for each release, so that independent parties can reproduce the results.
Each report identifies its basis: the signals used, the published framework or guideline applied, and a confidence and data-coverage score. Insufficient data automatically limits the strength of the conclusion.
During the pilot, the engine operates in shadow mode: its alerts are compared with actual check-up outcomes and clinician review. Precision and recall are measured on Vietnamese users before any broader performance claim is made.
A conclusion must be corroborated across multiple signals and over time. A single noisy reading never becomes an alert. Where the engine cannot be sufficiently certain, the case is escalated for human review.
We describe our components and their evidence precisely. MedGemma's published model card reports benchmark results such as 69.1% on MedQA; benchmark performance is not clinical accuracy, and An therefore validates on its own population before making stronger claims. An's durable asset is the consented, longitudinal health dataset that this architecture accumulates over time.