jahidbappi/sehat

Open-source, offline-first TB & pneumonia chest X-ray screening for low-resource clinics — calibrated, fairness-audited, decision-support ML # Sehat > AI disease screening for the four billion people a radiologist never visits. **Sehat** (সেহাত / صحت / सेहत — "health") is an open-source platform that lets a rural clinic with a cheap laptop and no internet connection screen chest X-rays for **tuberculosis** and **pneumonia** in seconds. It is decision support — it flags suspected cases for human confirmation. It never claims to diagnose. ## The problem Tuberculosis kills roughly 1.3 million people every year. It is curable. The bottleneck is not treatment — it is diagnosis. Chest X-ray screening requires a radiologist, and most of the world does not have one. Pneumonia is the leading infectious killer of children under five. Same bottleneck. The science to help already exists. The public datasets exist. What does not exist is an open, production-grade, offline-first system that a clinic can actually run. That is the gap Sehat fills. ## What Sehat does - **Screens** chest X-rays for TB and pneumonia with calibrated probability scores, not just binary guesses. - **Runs offline** on low-end hardware via quantized ONNX/TFLite exports. - **Proves it generalizes**: models are evaluated on hospitals they never trained on (site-held-out validation) and across sex/age subgroups — because medical ML that only works in the lab helps no one. - **Is honest about its limits**: every release ships with a model card, a data card, and subgroup performance reports. ## Architecture ``` public datasets ──► unified manifest ──► training ──► evaluation ──► model registry (NIH, NLM TB) (versioned, DVC) (Lightning) (fairness, (MLflow) calibration) │ ┌─────────────────────────────────────┤ ▼ ▼ edge export (ONNX/TFLite) cloud serving (FastAPI) │ │ offline clinic PWA NGO / hospital API ``` ## Quickstart ```bash # Install (data-layer tooling) pip install -e ".[dev,data]" # Download a public TB dataset …