Medical Insurance Cost Prediction

Predict annual medical insurance charges from demographic and health features. Exposes a production-ready FastAPI service with ensemble ML models (Decision Tree, Random Forest, Ridge Regression), input validation, KS-test drift monitoring, and a full test suite.

---

Quick Start

# 1. Clone
git clone https://github.com/atharvadevne123/Medical-Insurance-Cost-Prediction
cd Medical-Insurance-Cost-Prediction

# 2. Install
make install

# 3. Run API server
make run
# → http://localhost:8000/docs

# 4. Or with Docker
make docker-up

Single prediction

curl -X POST http://localhost:8000/predict \
  -H "Content-Type: application/json" \
  -d '{"age":35,"sex":"male","bmi":27.5,"children":2,"smoker":"no","region":"northeast"}'
# → {"predicted_charge": 6420.18, "model_version": "v1-decision-tree"}

---

Dataset

Column	Type	Description
age	int	Age of primary beneficiary (0–120)
sex	categorical	male / female
bmi	float	Body Mass Index (0–100)
children	int	Number of dependents (0–20)
smoker	categorical	yes / no
region	categorical	northeast / northwest / southeast / southwest
charges	float	Target — annual insurance cost (USD)

Source: insurance_predictor/data/insurance.csv (1,338 rows, no missing values).

---

Architecture

Medical-Insurance-Cost-Prediction/
├── app/
│   ├── main.py          # FastAPI application (lifespan, endpoints, middleware)
│   ├── schemas.py       # Pydantic request/response models
│   ├── models.py        # Ensemble: DecisionTree, RandomForest, Ridge
│   ├── monitoring.py    # KS-test drift detection (DriftMonitor)
│   └── persistence.py   # joblib save/load helpers
├── insurance_predictor/
│   ├── predictor.py     # Core: load_data, preprocess, train, predict, run
│   └── data/insurance.csv
├── tests/               # pytest test suite (9 modules, 80+ tests)
├── scripts/
│   └── train_pipeline.py  # CLI: train all models, save best
├── Dockerfile
├── docker-compose.yml
├── Makefile
└── requirements.txt

Request flow: POST /predict → Pydantic validation → preprocess() → trained model → response.

---

API Reference

GET /health

Returns service health and whether the model is loaded.

{"status": "healthy", "model_loaded": true, "version": "1.0.0"}

GET /metrics

Returns model performance from the last training run.

{"train_mae": 1823.4, "test_mae": 2541.7, "test_rmse": 4312.5, "test_r2": 0.857, "model_version": "v1-decision-tree"}

GET /version

Returns API and Python version info.

POST /predict

Predict insurance charge for one individual.

Request body:

{
  "age": 35, "sex": "male", "bmi": 27.5,
  "children": 2, "smoker": "no", "region": "northeast"
}

Response:

{"predicted_charge": 6420.18, "model_version": "v1-decision-tree"}

POST /predict/batch

Predict for up to 1000 individuals.

{"records": [{"age": 25, "sex": "female", "bmi": 22.0, "children": 0, "smoker": "no", "region": "southwest"}]}

---

ML Workflow

1. Load — insurance_predictor.predictor.load_data() validates schema and raises FileNotFoundError / ValueError on bad input.
2. Preprocess — StandardScaler on continuous features, one-hot encoding on categorical features.
3. Train — Three models evaluated; best by test MAE is served.
4. Monitor — DriftMonitor compares incoming feature distributions to training baseline using the two-sample KS test.

---

Testing

make test
# Or:
PYTHONPATH=. pytest tests/ -v

Test module	What it covers
test_data_loading.py	Schema, nulls, value ranges
test_preprocessing.py	Encoding, scaling, shape preservation
test_training.py	Metrics, reproducibility, hyperparameters
test_prediction.py	Array shape, positivity, batch sizes
test_api.py	All endpoints, validation rejection
test_models.py	Ensemble model comparison
test_monitoring.py	KS drift detection
test_persistence.py	Save/load/roundtrip

---

Contributing

See CONTRIBUTING.md. Run make lint and make test before opening a PR.

---

Author

Atharva Devne