🩺 Pima Diabetes Prediction — Cost-Aware Pipeline

A compact, production-minded workflow to predict diabetes risk from routine clinical measurements with probability calibration and an explicit threshold policy.

Case study: CASE_STUDY.md

---

What this repository includes

• Main notebook: diabetes-prediction-from-eda-to-production.ipynb
• Optional reference notebook: pima-indians-diabetes-database.ipynb
• Exported artifacts under ./artifacts/
• Optional scoring script: scripts/predict.py

---

Dataset

Source: Kaggle “Pima Indians Diabetes Database” (diabetes.csv).

Expected columns (typical):

• Pregnancies, Glucose, BloodPressure, SkinThickness, Insulin, BMI, DiabetesPedigreeFunction, Age, Outcome

Local (recommended)

1. Download the dataset CSV.
2. Place it at: data/raw/diabetes.csv

Kaggle

The notebook also supports: /kaggle/input/pima-indians-diabetes-database/diabetes.csv

---

Getting started

1) Install

python -m venv .venv
# Windows: .\.venv\Scripts\activate
# macOS/Linux: source .venv/bin/activate

pip install -r requirements.txt

2) Run the notebook

Open and run:

• diabetes-prediction-from-eda-to-production.ipynb

The notebook will:

• clean and audit the dataset
• train models and calibrate probabilities
• select an operating threshold (policy)
• export artifacts/pima_best_pipeline.joblib

---

Artifacts

The exported bundle includes:

• trained pipeline
• operating threshold
• run metadata

See artifacts/README.md.

---

Score a CSV (optional)

After exporting artifacts, you can score a CSV:

python scripts/predict.py --csv data/raw/diabetes.csv --out artifacts/scored.csv

The output adds:

• diabetes_proba
• diabetes_pred

---

Methodology notes

• Probability calibration makes thresholds usable for decisions.
• Threshold policy is selected on validation and exported with the artifact.

---

License

MIT (code). Dataset licensing depends on the dataset source where you download it.