CardioInsight AI – Interpretable Cardiovascular Risk

• 

Inspiration

Cardiovascular Disease (CVD) remains the leading cause of death worldwide, yet many people do not understand why they are at risk or what changes could actually reduce that risk. Most existing prediction tools provide only a risk score without explanation.

This project was inspired by the need for an accessible, transparent, and actionable AI system that not only predicts cardiovascular risk but also helps users understand the driving factors behind it and how they can improve their health outcomes.

What it does

CardioInsight AI predicts a person’s 10-year cardiovascular disease risk and goes beyond traditional models by:

Explaining which health factors contribute most to the predicted risk

Showing how much the risk could reduce if certain factors (like blood pressure, BMI, cholesterol, or smoking) are improved

Providing actionable insights instead of just a numerical score

Supporting both full clinical data and limited feature inputs, making it usable in real-world scenarios

The system also evaluates population-level risk using a real cardiovascular dataset and supports interactive individual risk analysis.

How we built it

We built CardioInsight AI using a two-model architecture:

Model A (Clinical Model): Uses a comprehensive set of medical features for high-accuracy risk estimation in clinical or research settings.

Model B (Lightweight Model): Uses a smaller, commonly available feature set to ensure usability when full medical data is unavailable.

Both models are trained using XGBoost, optimized for class imbalance, and calibrated using probability calibration to ensure reliable risk estimates.

To improve transparency and trust:

We used SHAP (Explainable AI) to identify which features increase or decrease risk

Implemented what-if simulations to estimate risk reduction if a modifiable factor improves

Deployed the system as an interactive Streamlit web application

Challenges we ran into

Handling class imbalance in real medical datasets without sacrificing recall for high-risk patients

Ensuring probability outputs are well-calibrated, not just accurate

Making complex ML explanations understandable to non-technical users

Designing medically safe what-if simulations without unrealistic assumptions

Aligning different datasets (training vs evaluation) while maintaining consistency

Accomplishments that we're proud of

Built an explainable CVD risk prediction system, not just a black-box model

Successfully implemented actionable risk-reduction insights

Designed a dual-model approach for both clinical and real-world usability

Deployed a fully working interactive web application

Maintained reproducibility with clean notebooks and saved model artifacts

What we learned

In healthcare AI, interpretability is as important as accuracy

Calibrated probabilities matter more than raw predictions in medical decision support

Small, well-chosen features can still produce meaningful predictions

Explainable AI greatly improves user trust and adoption

Building for real-world constraints is just as important as model performance

What's next for CardioInsight AI – Interpretable Cardiovascular Risk

Clinical validation using larger and more diverse datasets

Personalized lifestyle recommendations based on demographic profiles

Integration with wearable or routine health data

Improving fairness and bias analysis across population groups

Exploring collaboration with healthcare researchers for real-world impact

Built With

• github.
• google-colab
• jupyter
• numpy
• pandas
• python
• scikit-learn
• shap
• streamlit
• xgboost