CardioVision3D

About CardioVision3D

❤️ Inspiration

Cardiovascular risk prediction often ends as a number on a screen. We wanted to go beyond that.

CardioVision3D was inspired by one question: What if we could not only predict cardiovascular risk, but also explain it and make it visible?

Our goal was to connect clinical data, machine learning, explainable AI, and interactive anatomy into a single experience that helps users understand overall CAD risk as well as vessel-specific LAD, LCX, and RCA risk.

💡 What it does

CardioVision3D transforms 55 structured clinical features—covering patient history, symptoms, laboratory results, ECG, and echocardiography—into four independently generated risk estimates: CAD, LAD, LCX, and RCA. Pasted markdown

But prediction is only the beginning.

We use SHAP to show which patient-specific factors influenced each prediction, and then map the vessel-level probabilities onto an interactive 3D heart, where the LAD, LCX, and RCA can be explored individually. Pasted markdown

The result is a simple pipeline:

Clinical Data → Risk Prediction → Explainability → 3D Visualization → Actionable Insight

🛠️ How we built it

We built CardioVision3D as a full-stack AI system using React, TypeScript and Vite for the frontend and FastAPI, Python, PostgreSQL and Supabase for the backend and data layer. The 3D experience uses Three.js, React Three Fiber and Drei, while SHAP provides patient-level model explanations. Pasted markdown

Instead of forcing a single algorithm onto every target, we developed separate frozen ML pipelines for CAD, LAD, LCX and RCA. Our CAD model achieved a repeated cross-validation ROC-AUC of 0.9206, with the other models independently optimized for their respective targets. Pasted markdown

We also integrated Gemini as a grounded narrative layer that converts model predictions and SHAP context into an understandable Risk Story. Importantly, Gemini does not make the medical risk predictions; those come from our frozen ML models, and the system includes a deterministic fallback if the generative service is unavailable. Pasted markdown

🚧 Challenges we faced

One of our biggest challenges was ensuring that the system was technically impressive without making claims beyond what the models actually predict. Our 3D visualization therefore represents vessel-level model probability—not exact lesion location or physical percentage stenosis.

We also had to carefully prevent target leakage, synchronize four different prediction pipelines with their SHAP explanations, map model outputs to specific coronary vessels in 3D, and build a reliable end-to-end workflow spanning authentication, patient records, inference, explainability, visualization, AI narration, and report generation. Pasted markdown

📚 What we learned

Building CardioVision3D taught us that a strong healthcare AI system is not just about achieving a good model score. Interpretability, reliability, responsible communication, and user experience matter just as much as prediction.

Most importantly, we learned how multimodal systems can bridge the gap between raw clinical data and human understanding by combining machine learning, explainable AI, generative AI, and interactive 3D visualization.

🎯 Our vision

CardioVision3D is currently a research and educational decision-support prototype, not a diagnostic system or replacement for angiography or clinical judgment. Pasted markdown

Our vision is simple:

See the Risk. Know the Heart.

Built With

• fastapi
• gemini
• ml
• postgresql
• python
• react
• scikit-learn
• supabase
• three.js