CardioAI: Coronary Risk Prediction & 3D Visualization

• 

  Model performance: ROC curve, confusion matrix, model comparison and feature importance

• 

• 

  After turning off “typical chest pain”: CAD drops to 34% and the arteries turn green

• 

  Dashboard

• 

  LAD artery selected, with its “Why this LAD prediction?” breakdown

Inspiration

Cardiovascular disease is the leading cause of death worldwide. Risk calculators give a single percentage, but they don't show where in the heart the problem is likely to be, or why the model thinks so. We wanted a tool that turns clinical numbers into a picture a clinician or patient understands at a glance.

What it does

• Takes 54 routine clinical inputs: demographics, risk factors, vital signs, symptoms, ECG, labs and echocardiography.
• Predicts overall CAD and stenosis in each of the three major coronary arteries (LAD, LCX, RCA).
• Colours each artery on an interactive 3D heart (green → amber → red) by its own predicted probability. Positive vessels pulse.
• Lets users rotate and zoom, hover for details, and click a vessel to inspect it. The heart beats at the patient's recorded pulse rate.
• Explains every prediction with SHAP values in percentage points, a plain-English summary, and a table comparing each measurement to the cohort's normal range.
• Updates everything in real time (~0.3 s) as inputs change.
• Includes a Model performance page with ROC curves, confusion matrices and model comparison.
• Shows a permanent clinical-safety disclaimer: decision support and education only, not a substitute for diagnostic imaging.

How we built it

• Data: Extension of the Z-Alizadeh Sani dataset (UCI, 303 patients). LAD, LCX, RCA and Cath are excluded from the inputs to prevent target leakage. This is enforced in code and by a unit test.
• Models: For each of the 4 targets, Logistic Regression, Random Forest, XGBoost and a soft-voting ensemble are compared with repeated stratified 5-fold CV. The best model by ROC-AUC is kept. Its decision threshold is tuned on out-of-fold predictions (Youden's J), and it is scored once on a 20% held-out test set.
• Explainability: SHAP Linear/Tree explainers, rescaled so the contributions add up exactly to the predicted probability minus the cohort average.
• Backend: FastAPI (Python).
• Frontend: React + React Three Fiber (Three.js/WebGL). The heart and coronary anatomy are generated procedurally. The arteries follow the real anatomical grooves, so the model outputs always map to the right vessel. It runs smoothly without a dedicated GPU.

Results (held-out test set, 61 unseen patients)

Challenges we ran into

• With only 303 patients, the results have high variance. We used repeated CV and kept the test set untouched so the numbers are honest.
• Predicting the specific vessel from non-imaging data is hard by nature, especially for RCA.
• We wanted the 3D anatomy to be anatomically meaningful, light enough for laptops, and free of mesh licensing issues.
• SHAP values in log-odds are not intuitive for clinicians, so we converted them to probability points.

Accomplishments that we're proud of

• An end-to-end system: from a raw Excel file to live, explainable 3D predictions.
• Strict leakage prevention and an honest validation methodology.
• Explanations a non-technical person can read.

What we learned

• Validation design matters more than model choice on small clinical datasets.
• Explainability needs a translation layer before it is useful to clinicians.

What's next

• Validate on external, multi-centre datasets.
• Add imaging inputs, such as CT angiography, for lesion localisation within each vessel.
• Use high-fidelity anatomical meshes (BodyParts3D).

Built With

• fastapi
• python
• react
• scikit-learn
• shap
• three.js
• xgboost