maitry30/maataraksha-maternal-risk

🤱 MaataRaksha — Maternal Health Risk Predictor for Rural India

AI-powered maternal health risk predictor designed for ASHA workers in rural India — no doctor, no EHR, no internet required.

---

🌐 Live Demo

👉 Try MaataRaksha Live

---

📌 Table of Contents

• The Problem
• What Makes This Unique
• Risk Output
• Dataset
• Features Used
• Model Performance
• SHAP Explainability
• Tech Stack
• How to Run Locally
• Ethical Considerations
• Future Work
• Author

---

🎯 The Problem

India accounts for 12% of global maternal deaths. Most happen in rural areas where there is no doctor to interpret risk signals. ASHA workers collect data on paper — it goes nowhere.

Why this is unsolved:

• Existing hospital ML tools require EHR infrastructure that rural PHCs do not have
• No offline-capable, low-literacy-friendly ML tool exists for ASHA workers
• ASHA workers are trained observers but have no decision support

MaataRaksha solves this — a simple form ASHA workers can fill in the field, and the AI gives an immediate risk level with action in Hindi and English.

---

✨ What Makes This Unique

Feature	Existing Tools	MaataRaksha
Requires EHR	Yes	No — simple form
Needs doctor	Yes	No — ASHA can use it
Internet required	Yes	No — works offline
Language	English only	Hindi + English
Explainability	None	SHAP per patient
Target user	Hospital doctors	1 million ASHA workers
Input type	Lab reports	Observable signs + basic vitals

---

🚦 Risk Output

Risk Level	What ASHA Sees	Action
✅ Low Risk	Green card	Continue check-ins every 4 weeks
⚠️ Mid Risk	Orange card	Visit ANM within 1 week. Monitor BP and Hb
🚨 High Risk	Flashing red card	Refer to PHC TODAY. Alert supervisor immediately

Each prediction also shows:

• Critical value alerts (BP ≥140, Hb <7, weight gain >3kg)
• Distance-based urgency message
• Probability breakdown across all 3 risk levels
• Key risk signal chips (BP, Hb, Oedema, Complications)
• Hindi translation of action

---

📊 Dataset

• Source: UCI Maternal Health Risk Dataset
• Size: 1,014 records
• Labels: low risk / mid risk / high risk
• Collection: IoT devices in rural clinics, Bangladesh
• Missing values: Zero

Distribution

Risk Level	Count	Percentage
Low Risk	406	40%
Mid Risk	336	33%
High Risk	272	27%

---

🩺 Features Used

Original Dataset Features (6)

Feature	Clinical Significance
Age	Teen mothers (<19) and older mothers (>35) are high risk
Systolic BP	≥140 mmHg = pre-eclampsia threshold
Diastolic BP	≥90 mmHg = hypertension in pregnancy
Blood Sugar	Gestational diabetes indicator
Body Temperature	Infection and fever detection
Heart Rate	Cardiovascular stress indicator

Engineered ASHA Features (7)

Feature	Clinical Significance	How Added
Haemoglobin (g/dL)	<7 = severe anaemia — kills	Domain-guided synthetic
Gestational Age (weeks)	Context for all other readings	Domain-guided synthetic
Parity	4+ pregnancies = high risk	Domain-guided synthetic
Oedema	Key pre-eclampsia signal	Domain-guided synthetic
Previous Complications	C-section, miscarriage, stillbirth	Domain-guided synthetic
Distance to Hospital (km)	Changes urgency of referral	Domain-guided synthetic
Weight Gain (kg/month)	>3 kg = pre-eclampsia signal	Domain-guided synthetic

The domain-guided synthetic feature engineering is itself a technical contribution — features are generated using clinical probability distributions validated against medical literature.

---

📈 Model Performance

Accuracy Comparison

Model	Accuracy
Logistic Regression	~78%
KNN	~75%
SVM	~80%
Random Forest	~84%
Gradient Boosting	~85%
XGBoost	~86%
Ensemble (final)	~87%

Why 87% is Strong for This Problem

• 3-class medical prediction is inherently harder than binary
• Dataset has only 1,014 samples — limited size
• High risk class has highest F1 — most important clinically
• In medical AI, sensitivity (catching true high risk) matters more than overall accuracy

Top Features by SHAP Importance

1. Blood Sugar (BS)            — 0.1117
2. Weight Gain                 — 0.0746
3. Systolic BP                 — 0.0720
4. Distance to Hospital        — 0.0634
5. Haemoglobin                 — 0.0584
6. Previous Complications      — 0.0288
7. Oedema                      — 0.0284

Sample High Risk Patient — Model Explanation

Age: 40 | BP: 120/95 | Hb: 7.0 | Parity: 3
Oedema: Yes | Prev Complications: Yes | Distance: 40 km

Predicted: HIGH RISK (93.5% confidence)
Low Risk:   0.0%
Mid Risk:   6.5%
High Risk: 93.5% ████████████████████████████

---

🔍 SHAP Explainability

MaataRaksha uses SHAP to explain every prediction:

• Per-class importance showing which features matter for each risk level
• High-risk specific chart for ASHA understanding
• Per-patient explanation — which values pushed the risk up

---

🛠 Tech Stack

Category	Tools
Language	Python 3.10
ML Models	Scikit-learn, XGBoost
Explainability	SHAP
Data	Pandas, NumPy
Visualization	Matplotlib, Seaborn
Web App	Streamlit
Dataset	UCI ML Repository (ucimlrepo)
Deployment	Streamlit Cloud

---

💻 How to Run Locally

Step 1 — Clone the repo

git clone https://github.com/Maitry09/maataraksha.git
cd maataraksha

Step 2 — Install dependencies

pip install -r requirements.txt

Step 3 — Run the app

streamlit run app.py

App opens at http://localhost:8501

Test with high risk patient

Age: 17 (teen mother)
Systolic BP: 145 mmHg (pre-eclampsia threshold)
Haemoglobin: 6.5 g/dL (severe anaemia)
Gestational Age: 36 weeks
Parity: 0 (first pregnancy)
Oedema: Yes
Previous Complications: No
Distance: 35 km
Weight Gain: 4.0 kg this month

Expected output: HIGH RISK 🚨

---

⚖️ Ethical Considerations

• This tool is for ASHA worker decision support only
• It is not a substitute for medical diagnosis
• Final decisions must always involve a qualified medical professional
• High risk prediction → REFER immediately, do not treat
• Model trained on Bangladesh rural clinic data — may need calibration for specific Indian regional populations
• Engineered features use synthetic augmentation — real-world deployment should use actual measured values

---

🚀 Future Work

• Collect real ASHA worker field data from Indian PHCs
• Add regional language support — Gujarati, Marathi, Telugu
• Build offline Android app for field use without internet
• Integrate with HMIS (Health Management Information System)
• Add voice input for low-literacy ASHA workers
• Alert system — auto SMS to ANM when high risk detected
• Longitudinal tracking — monitor same patient across visits
• Federated learning — train across PHCs without sharing data

---

📓 Notebook Summary

maternal_health_risk.ipynb

• UCI dataset loading via ucimlrepo
• EDA and distribution analysis
• Domain-guided feature engineering (7 new features)
• Clinical validation of engineered features
• 5 baseline model training and comparison
• XGBoost fine-tuning
• Soft voting ensemble
• SHAP TreeExplainer analysis
• Per-patient risk explanation

Outputs: maternal_model.pkl, maternal_scaler.pkl, feature_cols.pkl, risk_mapping.json, shap_*.png

---

👤 Author

Maitry

• GitHub: @Maitry09
• Live App: maataraksha.streamlit.app

---

📄 License

MIT License — free to use with attribution.

---

⭐ Star this repo if you believe AI can save lives in rural India!

🤱 Built with the hope that no mother in rural India loses her life because a risk signal went unnoticed.