❤️ Heart Risk AI - Advanced Cardiovascular Risk Predictor

This project provides an advanced AI-powered tool to predict cardiovascular risk. It includes a full MLOps pipeline, from data preprocessing and advanced feature engineering to a real-time interactive web application that allows users to choose between two different model versions.

Features

• Dual Model Support: Switch between a baseline model and an advanced, fine-tuned version.
• Interactive UI: A user-friendly Streamlit application for real-time predictions.
• Advanced Machine Learning: The v2 model uses a sophisticated stacked ensemble, advanced feature engineering, and probability calibration.
• Detailed Analysis: The app provides a breakdown of risk factors and personalized recommendations.

Quick Start & Setup

1. Clone the repository:

   git clone https://github.com/JuanInfante122/heart-risk-model.git
   cd heart-risk-model

2. Set up the environment:

   make setup
   make install

3. Configure Hugging Face Token: This application loads models from a private Hugging Face repository. You need to provide an access token.

   • Create a file at .streamlit/secrets.toml.
   • Add your Hugging Face token to this file as follows:

     HUGGING_FACE_HUB_TOKEN = "hf_YOUR_TOKEN_HERE"

4. Run the application:

   make run

Model Evolution & Results

This project showcases a clear evolution from a baseline model to a more advanced and robust predictor. Below is a summary of the results for both versions.

Metric	Baseline Model (v1)	Advanced Model (v2)
AUC-ROC	80.0%	79.4%
Sensitivity	85.0%	85.2%
Specificity	55.0%	53.4%
F1-Score	-	73.5%
Accuracy	-	69.3%

The v2 model was optimized to maintain high sensitivity while improving overall performance and robustness.

Methodology

• Dataset: Cardiovascular Disease Dataset from Kaggle (70,000 entries).
• Goal: Binary classification to predict 10-year cardiovascular risk.

Baseline Model (v1)

• Models: A simple ensemble of Random Forest and XGBoost.
• Feature Engineering: Basic features derived from the original dataset.
• Validation: Standard 80/20 stratified split.

Advanced Model (v2)

The second version of the model represents a significant leap in methodology:

• Advanced Feature Engineering: Includes polynomial features, statistical aggregations, and medical risk scores (e.g., Framingham, pulse pressure).
• Intelligent Feature Selection: Combines multiple techniques (Univariate stats, RFE, Tree-based importance, L1 regularization) to select the most impactful features.
• Advanced Sampling: Evaluates multiple strategies (SMOTE, ADASYN, etc.) to handle class imbalance, selecting the best-performing one for training.
• Hyperparameter Tuning: Uses Optuna for sophisticated and efficient hyperparameter optimization.
• Stacked Ensemble: Creates a powerful stacked ensemble model using the best-performing base models as estimators and a logistic regression model as the meta-learner.
• Probability Calibration: Applies isotonic regression to calibrate model probabilities, making them more reliable.

Tech Stack

• Python 3.8+
• Scikit-learn, XGBoost, LightGBM, CatBoost
• Pandas, NumPy
• Streamlit for the interactive web app
• Optuna for hyperparameter tuning

Medical Disclaimer

This tool is for educational and research purposes only. It is not a substitute for professional medical advice, diagnosis, or treatment.

Contributing

Contributions are welcome! Please follow the guidelines in CONTRIBUTING.md.

License

MIT License (see LICENSE).