CardioVision 3D: Diagnostic Assessment System

Inspiration

Coronary Artery Disease (CAD) remains the primary contributor to global cardiovascular mortality. Conventional non-invasive triaging approaches frequently yield binary, flat classification outputs that fail to provide spatial anatomical insights. Furthermore, standard academic benchmark models often suffer from target leakage by including post-procedure ground-truth variables (such as catheterization outcomes). CardioVision 3D was developed to provide an end-to-end, clinically safe decision-support system that maps multi-vessel stenosis risk directly to an interactive 3D cardiac digital twin using non-invasive clinical biomarkers.

What It Does

• Multi-Vessel Stenosis Stratification: Simultaneously predicts global CAD risk and localized stenosis probabilities across the Left Anterior Descending (LAD), Left Circumflex (LCX), and Right Coronary Artery (RCA) trees with 95% epistemic confidence intervals.
• Interactive 3D Coronary Artery Twin: A WebGL/Plotly-powered anatomical representation that dynamically maps predicted stenosis risks to corresponding vessel segments using standardized clinical color gradients.
• Explainable AI (SHAP): Decomposes individual biomarker contributions (such as chest pain presentation and ejection fraction) using a tree explainer to eliminate black-box opacity.
• Counterfactual Intervention Simulation: Models the therapeutic effect of Guideline-Directed Medical Therapy (GDMT), allowing clinicians to simulate reductions in blood pressure and lipid profiles to visualize quantified arterial risk reduction.
• Clinical Export & EHR Integration: Enables single-click plain-text cardiology consultation report generation and raw biomarker CSV export.

How We Built It

• Anti-Leakage Machine Learning Pipeline: Trained on the Z-Alizadeh Sani cohort, strictly isolating and eliminating Cath and vessel ground truths to prevent artificial accuracy inflation.
• Multi-Target Estimator Engine: Class-balanced, tuned XGBoost estimators benchmarked against Logistic Regression, Random Forest, and LightGBM via 5-fold cross-validation.
• Frontend & Spatial Visualization: Built using Streamlit, Plotly 3D WebGL rendering, and responsive custom CSS clinical card styling.

Challenges We Overcame

• Preventing Target Leakage: Ensuring catheterization markers did not contaminate the training space while maintaining strong sensitivity and AUROC across vessel territories.
• Parametric 3D Coronary Geometry: Constructing anatomical spline trajectories for the coronary branches and dynamically interpolating risk-based color gradients in real time.

Clinical & Regulatory Safety

CardioVision 3D is designed strictly as an assistive clinical decision-support and educational reference system under SaMD Class IIa guidelines. It serves to augment, rather than replace, licensed cardiological judgment.

Built With

• healthcare
• plotly
• python
• scikit-learn
• shap
• stramlit
• xgboost