CardioVision AI — See the Risk. Understand the Heart.

Inspiration

Cardiovascular disease is a major global health challenge. Machine-learning models can estimate coronary artery disease (CAD) risk, but presenting probabilities alone makes their results difficult to interpret. We wanted to bridge this gap by combining machine learning, explainable AI, and interactive 3D visualization.

What We Built

CardioVision AI is an AI-assisted cardiovascular risk visualization prototype that predicts overall CAD status and estimates stenosis probabilities for three coronary arteries: the Left Anterior Descending (LAD), Left Circumflex (LCX), and Right Coronary Artery (RCA).

Users can explore patient clinical measurements, view model predictions, inspect contributing features, and interact with a 3D heart visualization that highlights vessel-level risk.

How It Works

• Machine Learning: We trained classification models using clinical data and compared their performance using nested cross-validation and held-out testing.
• Four Prediction Tasks: Separate models estimate overall CAD status and LAD, LCX, and RCA stenosis probabilities.
• Interactive 3D Visualization: A React Three Fiber visualization displays the heart and coronary vessels, allowing users to rotate the model and explore vessel-specific predictions.
• Explainable AI: SHAP-based explanations help users understand which input features contributed to a prediction.
• Backend Integration: FastAPI serves predictions and model metrics, while the React and TypeScript frontend presents the results through an interactive dashboard.

Results

Our current evaluation produced the following nested cross-validation ROC-AUC results and held-out test ROC-AUC results:

| Target | Model | Nested CV ROC-AUC | Test ROC-AUC | | CAD | Logistic Regression | 0.950 ± 0.017 | 0.832 | | LAD | Random Forest | 0.880 ± 0.046 | 0.752 | | LCX | Random Forest | 0.751 ± 0.112 | 0.706 | | RCA | Logistic Regression | 0.715 ± 0.035 | 0.692 |

The held-out test set contains 61 patients. The results show stronger discrimination for overall CAD than for some individual vessel targets, highlighting the challenges of small clinical datasets and vessel-level prediction.

Challenges We Faced

The small dataset required careful validation to reduce overfitting and avoid misleading performance estimates. We also needed to prevent target leakage by excluding LAD, LCX, RCA, and Cath from model input features. Another challenge was representing vessel-level predictions on a 3D anatomical model without implying that the model identifies exact plaque locations.

What We Learned

This project taught us how to integrate machine learning, model evaluation, explainability, backend APIs, and interactive 3D visualization into one application. We also learned the importance of transparent evaluation, honest communication of uncertainty, and responsible presentation of medical AI predictions.

What's Next

Future improvements include expanding the clinical dataset, strengthening external validation, improving probability calibration, refining anatomical visualization, and evaluating the system with clinical experts.

Clinical Safety

CardioVision AI is an educational and research prototype, not a medical diagnostic tool. Its predictions must not replace professional medical evaluation or diagnostic imaging. The 3D display represents predicted vessel-level risk, not an actual anatomical reconstruction or confirmed blockage location.

Built With

• fastapi
• numpy
• pandas
• python
• react
• react-three-fiber
• scikit-learn
• shap
• tailwind-css
• three.js
• typescript
• vite