Inspiration

In rural India, Community Health Workers (ASHA workers) often screen patients during ~5-minute doorstep visits under intermittent connectivity and frequently without point-of-care glucometers or HbA1c analyzers. Most clinical AI demos fail in the field because they assume stable internet, mandatory lab tests, black-box risk scores, and English-only interfaces.

Sanjeevni Healthcarer (built by Team Aparajit) bridges this gap with an offline-first, bilingual (English / हिन्दी), explainable clinical screening & triage platform engineered for safety, transparency, and real-world rural constraints.

Research Prototype Disclaimer: Sanjeevni Healthcarer is an investigational clinical decision-support prototype, not a diagnostic medical device, and never replaces examination by a qualified physician.


What It Does

  • 5-Minute Guided Doorstep Visit (ASHA Worker role): Walks workers through Basic Vitals → Blood Pressure → PHQ-9 Mental Health → Triage & Referral with voice read-aloud (EN / हिन्दी), No-Lab Mode toggle, and WhatsApp follow-up sharing.
  • Dual-Path Clinical Risk Engine (Full-Lab & No-Lab): Automatically routes to a dedicated No-Lab XGBoost model whenever glucose/HbA1c readings are unavailable—clearly flagging predictions as unconfirmed estimates requiring confirmatory glucometer/lab testing.
  • Non-Downgradable Clinical Safety Layer: Deterministic clinical rules sit above ML predictions (shared/triage_thresholds.json), strictly enforcing GREEN -> YELLOW -> RED escalations based on ADA 2024 (Glucose $\ge 300\text{ mg/dL}$, HbA1c $\ge 12.0\%$), AHA/ACC ($BP \ge 180/120\text{ mmHg}$), and PHQ-9 crisis thresholds (Item 9 $> 0$ or total $\ge 20$). Rules can only escalate, never downgrade.
  • PHC Doctor Triage Queue & HL7 FHIR R4 Export: Sorts patients strictly RED → YELLOW → GREEN, tracks longitudinal vitals, and exports LOINC-coded HL7 FHIR R4 Bundles (Patient, Observation, RiskAssessment).
  • Explainable AI & Counterfactuals: Sub-2s SHAP TreeExplainer drivers with bilingual plain-language explanations, plus DiCE counterfactuals constrained strictly to modifiable lifestyle features (bmi, HbA1c_level, smoking_history)—never altering immutable attributes like age or gender.
  • Offline-First PWA: Service Worker (sw.js) caches the app shell, IndexedDB queues visits with client-generated UUID idempotency, and an offline balanced LogisticRegression scorer shares identical cutoffs via shared/triage_thresholds.json.
  • PHQ-9 & Bilingual Hindi RAG Guide: Built-in depression screener with Indian crisis helplines (Tele-MANAS 14416, KIRAN, 112) and Unicode Devanagari RAG over grounded ADA/AHA/PHQ-9 documents.

Data Science, Calibration & Fairness Governance

Trained and evaluated on 100,992 records (80/20 stratified split, 8.5% positive prevalence, RANDOM_SEED = 42):

Model Path ROC-AUC Sensitivity Specificity Precision Brier Score (Uncal → Cal) ECE (Uncal → Cal)
Full-Lab XGBoost + Platt 0.9773 91.61% 90.02% 46.02% 0.0598 → 0.0258 0.0961 → 0.0232
No-Lab XGBoost + Platt 0.8324 78.45% 72.69% 21.06% 0.1697 → 0.0667 0.2612 → 0.0076
Offline Balanced Logistic (Full-Lab) 0.9616 — — — 0.0781 —
Offline Balanced Logistic (No-Lab) 0.8209 — — — 0.1756 —

Honest Fairlearn Subgroup Audit (100,974 records)

Rather than hiding bias, our live Model Governance Card exposes subgroup metrics directly in the UI:

  • Gender Parity: Full-Lab DP Diff 0.0406, EO Diff 0.0228 (Female recall 91.5%, Male recall 93.7%).
  • Age Disparity Limitation: Under-30 recall drops sharply without labs (66.1% Full-Lab vs 3.6% No-Lab), while 60+ FPR rises (27.9% Full-Lab vs 76.1% No-Lab). Our UI explicitly warns clinicians to confirm younger symptomatic patients and 60+ alerts with lab testing.

How We Built It

  • Backend & ML: Python, FastAPI, XGBoost, scikit-learn (Platt calibration + balanced LogisticRegression), SHAP TreeExplainer, DiCE, Fairlearn, SQLAlchemy + SQLite, JWT RBAC (worker, doctor, admin) with zero-PII audit logs.
  • Frontend: React 19, TypeScript, Vite 8 PWA (sw.js + IndexedDB), Tailwind CSS, i18next (full EN / हिन्दी parity), Recharts, Framer Motion, Three.js.
  • Deployment & CI: GitHub Actions (pytest + npm run build), Docker + Gunicorn/Uvicorn on Render (Backend), and Vercel (Frontend).

Challenges We Faced

  1. Unfairness & Miscalibration in Raw Tree Models: Raw XGBoost probabilities showed high Expected Calibration Error (0.2612 in No-Lab); Platt scaling reduced No-Lab ECE to 0.0076 and Full-Lab Brier to 0.0258.
  2. Online/Offline Triage Drift: Preventing divergence between the Python backend and browser offline mode required unifying thresholds and logistic weights in a single shared/triage_thresholds.json contract.
  3. Hindi RAG Tokenization: Standard ASCII regex tokenizers dropped Devanagari scripts; we implemented a Unicode-aware Devanagari tokenizer routing Hindi queries to localized clinical docs with strict refusal ("जानकारी नहीं है") outside grounded guidelines.

Demo Credentials (Quick Select on Login Page)

(Note: Backend runs on Render free tier and may take ~30–50s on the initial cold request.)

Built With

Share this project:

Updates

Submission history