Inspiration
PROBLEM: An ASHA (frontline health) worker in rural India visits pregnant women with only a BP cuff, a thermometer and a glucometer. She has to decide, on the spot, which mothers are high-risk and need urgent referral to a clinic — and delayed referral is one of the biggest drivers of maternal deaths. Today she has no tool for that decision at all: no app, no score, and often no signal.
What it does
SOLUTION: AshaCare AI tells the worker, from 6 basic vitals she can already measure, which mothers are high-risk and need urgent referral — and why, in plain language she can read out. Single-case triage, or a whole village roster uploaded as CSV and ranked by risk so she knows who to visit first. The priority list also downloads as a plain-text summary for printing, so it works on paper in a village with no signal.
How we built it
HOW IT WORKS:
1) Worker enters 6 vitals (age, systolic/diastolic BP, blood sugar, body temperature, heart rate) — or uploads a roster CSV.
2) An XGBoost classifier trained on the UCI Maternal Health Risk dataset (1,014 real cases) outputs Low / Mid / High risk with class probabilities (held-out accuracy 0.8522, high-risk recall 0.964).
3) Explanations are measured, not invented: each vital is replaced by the training median and the drop in the model's own predicted probability is that vital's contribution, written as a plain sentence against the usual range, plus a one-paragraph plain summary per case.
4) Referral urgency is explicitly rule-based (labelled as such in every response): High → URGENT, refer within 24 hours; Mid → SOON, follow up within 7 days; Low → ROUTINE. A single extreme vital escalates the case to IMMEDIATE emergency referral whatever the model said (systolic ≥ 180 or diastolic ≥ 120 mm Hg, blood sugar ≥ 20 mmol/L, temperature ≥ 103.5 °F).
5) Bad input is never guessed: out-of-range, missing or non-numeric vitals are rejected for a single case and skipped-and-listed for a roster. It is a screening aid only, not a diagnosis — high-risk cases must be referred to a doctor.
Built With
- fastapi
- pytest
- python
- scikit-learn
- streamlit
- xgboost
