Testing without taking the data
BODH lets developers evaluate models against real, anonymised Indian health data without that data ever leaving its original location. The model goes to the data rather than the data going to the model.
That mechanism solves two problems at once. It protects patient privacy and keeps data within national jurisdiction. And it answers the question that matters most for any health AI deployed in India: does this actually work on Indian patients, or only on the population its training data came from? Models trained on Western cohorts are known to perform poorly against Indian disease profiles and clinical presentations, and no amount of internal testing detects that.
What RISE submits for benchmarking
Every AI component that could influence a clinical or resource-commitment decision is in scope. The programme's position is that a component which has not been benchmarked does not go into a live response.
| Component | Reference | What is benchmarked |
|---|---|---|
| Trend analysis | MOD-5.1 | Cluster detection accuracy and false-positive rate against historical outbreak and displacement events. |
| Need forecasting | MOD-5.2 | Forecast accuracy at 12, 24 and 72-hour horizons, and calibration of the published confidence intervals. |
| Air ambulance demand forecasting | MOD-5.2 / AA-05 | Predicted versus actual aircraft requirement by disaster type, terrain and reported injury severity. |
| Disaster progression simulation | MOD-5.3 | Scenario fidelity against recorded event progressions. |
| Clinical triage support | BAEMS BM-01 | Triage priority concordance with clinician assignment, and under-triage rate — the measure that matters most, since under-triage kills and over-triage only costs. |
| Resource allocation optimisation | MOD-3.2 | Allocation quality and equity of distribution across population segments. |
| ML routing and dispatch optimisation | 2027 roadmap | Benchmarked before deployment, not after. Planned for the 2027 innovation stage. |
Bias is the finding we expect to have to act on
Disaster data is systematically skewed. Events in accessible areas are better recorded than events in inaccessible ones; populations with phones generate more signal than populations without; the areas that most need air evacuation are precisely the areas least represented in historical records.
A model trained on that data will underestimate need in exactly the places the platform exists to reach. We expect benchmarking to show it. The commitment is to publish that finding and correct for it rather than to report the aggregate accuracy figure and move on.
BODH submission is committed, not achieved. See the conformance register at CMP-08 for the position on each component.