Diabetes Risk Prediction 🏥

A compact, explainable ML project that predicts an individual's risk of developing diabetes based on health metrics. The project includes an exploratory notebook, a trained LightGBM model, and a Streamlit app for interactive predictions.

---

Table of Contents

• Overview
• Quick Start
• Project Structure
• Model & Features
• How the App Works
• Retraining / Experiments
• Deployment & Usage
• Notes & Disclaimer
• License & Contact

---

Overview ✅

This repository contains a machine learning pipeline and demo application to predict diabetes risk from structured health and lifestyle features. The project was developed and evaluated in diabetes_prediction.ipynb and the best-performing model (LightGBM) is saved as diabetes_model_LightGBM.pkl for inference inside a Streamlit app (app.py).

---

Quick Start 🚀

1. Create an environment and install dependencies:

python -m venv .venv
.\.venv\Scripts\activate    # Windows
pip install -r requirements.txt

2. Run the Streamlit app locally:

streamlit run app.py

3. Open the displayed URL in your browser, fill in health inputs, and click Predict Diabetes Risk.

---

Project Structure 🔧

• app.py — Streamlit app for interactive predictions and risk explanations.
• diabetes_prediction.ipynb — Notebook with EDA, feature engineering, model training and evaluation.
• diabetes_model_LightGBM.pkl — Trained LightGBM classifier used by app.py (model file).
• feature_names.json — Ordered list of features expected by the model.
• model_metadata.pkl — Saved metadata about the model/training (if present).
• submission.csv — Example output/predictions on the test partition.
• requirements.txt — Python dependencies used by the project.

---

Model & Features 📊

• Model: LightGBM classifier (saved as diabetes_model_LightGBM.pkl)
• Reported performance: Best AUC ≈ 0.7148 (3-fold CV, see diabetes_prediction.ipynb)
• Feature list: See feature_names.json. Important derived features include:
  • bmi_age — interaction: BMI × age
  • waist_bmi — waist-to-hip ratio × BMI
  • pulse_pressure — systolic − diastolic
  • chol_hdl_ratio, trig_hdl_ratio — lipid ratios
  • health_score — composite lifestyle score (diet, activity, sleep, screen-time)
  • is_senior, is_obese — binary flags

These features are created in create_features() inside app.py and in the notebook during preprocessing.

---

How the App Works 🧠

• The app loads diabetes_model_LightGBM.pkl using joblib.
• User inputs (demographics, vitals, labs, lifestyle) are converted into a single-row DataFrame in the exact order expected by the model, engineered via create_features(), and then passed to model.predict_proba().
• The UI shows the predicted probability and a risk level category derived from thresholds in the app:
  • Low Risk: probability < 0.3
  • Moderate Risk: 0.3 ≤ probability < 0.5
  • High Risk: 0.5 ≤ probability < 0.7
  • Very High Risk: probability ≥ 0.7

⚠️ Important: This is a demonstrative model for educational/informational purposes only. It is not a substitute for professional medical diagnosis.

---

Retraining & Experiments 🧪

• All training, hyperparameter evaluation (multiple algorithms were compared), and model selection are performed in diabetes_prediction.ipynb. The notebook uses scikit-learn pipelines and evaluates models using ROC AUC with cross-validation.
• To retrain: open the notebook, adjust preprocessing or model hyperparameters, re-run the training cells and save the best model. The notebook contains cells that save:
  • diabetes_model_LightGBM.pkl
  • feature_names.json
  • model_metadata.pkl

---

Deployment & Usage Tips ⚙️

• For local demo: streamlit run app.py is sufficient.
• To deploy: host the repository on a VM or use Streamlit Cloud / other platforms that support Streamlit apps. Ensure diabetes_model_LightGBM.pkl is present in the app directory.
• Monitor inputs closely: the model expects numerical ranges similar to typical adult health metrics; out-of-distribution inputs may produce unreliable probabilities.

---

Notes & Disclaimer ✍️

• Data privacy: No personal data is stored by this demo app; it runs locally and predictions are ephemeral.
• Medical disclaimer: The predictions are probabilistic estimates from an ML model and should not be used for clinical decisions.

---

License & Contact 📨

If you use or adapt this project, please add proper attribution. For questions or contributions, open an issue or contact the maintainer.

Enjoy exploring the model!