Coronary Risk Explorer

Interactive 3D anatomical heart & clinical decision-support predicting multi-vessel coronary artery stenosis (LAD/LCX/RCA) and CAD from 54 routine variables with real-time SHAP explainability.

Comment

Coronary Risk Explorer

Inspiration

Cardiovascular disease remains the leading cause of death globally. In outpatient triage, physicians evaluate dozens of non-invasive variables—including ECG tracings, symptoms, echocardiograms, and blood work—to decide whether a patient needs invasive catheterization.

However, raw risk percentages fail to provide clinicians or patients with an intuitive, spatial understanding of where disease may be developing.

We built Coronary Risk Explorer to connect non-invasive predictive AI, clinical interpretability, and real-time 3D anatomical visualization into an interactive clinical decision-support tool.

What It Does

Coronary Risk Explorer predicts overall Coronary Artery Disease (CAD)—represented by the dataset's Cath label—and the presence of ≥50% stenosis in the three major coronary arteries:

  • LAD — Left Anterior Descending
  • LCX — Left Circumflex
  • RCA — Right Coronary Artery

The model uses 54 routine clinical parameters.

Interactive 3D Anatomical Mapping

The application visualizes the heart, thorax, and coronary tree using anatomical meshes derived from BodyParts3D (CC BY 4.0).

Artery branches are dynamically recolored according to calibrated risk using custom WebGL shaders featuring:

  • Pulsatile blood-flow visualization
  • Risk-based glow effects
  • Heartbeat synchronization
  • Patient-specific pulse-rate animation

This transforms an abstract probability into an intuitive spatial representation of potential disease.

Fast In-Browser Inference

The system provides real-time scoring directly in the browser:

  • Approximately 60 ms end-to-end scoring latency
  • No dedicated GPU required
  • Approximately 1.7 ms GPU frame time on integrated Intel Arc graphics

Exact Explainability

The application computes real-time SHAP feature attributions, decomposing the prediction into domain-level contributions:

  • ECG
  • Echocardiography
  • Laboratory measurements
  • Medical history

The interface also provides standard physiological adult ranges to help contextualize individual measurements.

Clinical Safety & Decision Thresholds

The application includes several safeguards designed for responsible clinical decision support:

  • Threshold explorer
  • Rule-out preset
  • Balanced preset
  • Cath-vs-vessel consistency warnings
  • Prominent clinical disclaimers

How We Built It

ML Pipeline

The machine-learning pipeline was developed across 8 reproducible notebooks using the UCI Extension of the Z-Alizadeh Sani benchmark, containing 303 patients.

We evaluated:

  • Logistic Regression
  • Support Vector Machines
  • Random Forest
  • XGBoost
  • LightGBM

Models were evaluated using repeated 5×5 stratified cross-validation, with hyperparameters optimized using nested Optuna TPE.

A 61-patient holdout set was strictly sealed until final evaluation.

Final performance:

Prediction ROC-AUC
Cath / overall CAD 0.854
LAD stenosis 0.794

Confidence intervals were calculated using 2,000-resample bootstrap estimation.

Calibration

We applied out-of-fold Platt scaling so that predicted probabilities better reflect real-world event frequencies rather than merely providing ranking scores.

Backend

The backend is implemented as a FastAPI microservice with:

  • Pydantic validation schemas
  • Per-response CSP nonces
  • Rate limiting
  • Automated API testing
  • Leakage-prevention tests
  • SHAP additivity tests

The project contains 39 automated tests covering the above safety and correctness requirements.

Frontend & 3D

The frontend combines:

  • Three.js
  • WebGL
  • Custom GLSL shaders
  • GSAP ScrollTrigger
  • Responsive UI

The application was designed with zero build-step overhead for the frontend experience.

Challenges We Ran Into

Small-Sample Clinical Constraints

Working with only 303 patients required unusually strict validation practices.

We used:

  • Nested cross-validation
  • Stratification
  • Sealed holdout evaluation
  • Leakage prevention
  • Bootstrap confidence intervals

This was necessary to reduce the risk of reporting overly optimistic performance.

Anatomical Reconstruction

The BodyParts3D models represent ventricular cavities primarily as blood volumes rather than complete anatomical outer surfaces.

To create the visualization required by the product, we algorithmically reconstructed the outer ventricular surface using:

  • Local wall thickness
  • Anatomical geometry
  • Coronary proximity

This allowed us to produce a more complete interactive representation of the heart.

Low-Latency Real-Time SHAP

Computing exact SHAP explanations for individual predictions introduced a latency challenge.

We optimized single-threaded TreeExplainer scoring to deliver exact feature attributions in approximately 60 ms, avoiding the additional overhead of a worker pool.

Accomplishments We're Proud Of

  • Met all 5 requirements of Track A with a production-hardened web application.
  • Built an honest, clinically grounded ML system with:
    • Bootstrap confidence intervals
    • Probability calibration
    • Calibration slopes close to 1.0
  • Created an interactive 3D WebGL anatomical experience capable of maintaining approximately 60 FPS on standard integrated graphics.
  • Combined predictive modeling, interpretability, clinical context, and anatomical visualization into a single workflow.

What We Learned

Evaluating clinical machine learning requires prioritizing validation methodology, calibration, and explainability over simply choosing the most complex or highest-performing black-box model.

In a small clinical dataset, methodological rigor can be more important than squeezing out another few points of model performance.

What's Next

The next stage is exploring multimodal integration with angiographic video datasets, including datasets such as AngioCAD, to enable:

  • Segment-level coronary analysis
  • Multi-vessel validation
  • Integration of non-invasive predictions with angiographic evidence
  • More granular anatomical risk localization
  • Potential longitudinal disease monitoring

Built With

  • achieving-0.854-roc-auc-(cath)-and-0.794-roc-auc-(lad)-with-2
  • and-39-automated-tests-covering-api-contracts
  • and-lightgbm-across-repeated-5x5-stratified-cv
  • and-shap-additivity.-frontend-&-3d:-three.js-webgl-rendering
  • custom-glsl-shaders
  • docker
  • fastapi
  • gsap
  • gsap-scrolltrigger-animations
  • html5
  • javascript
  • leakage-prevention
  • optuna
  • per-response-csp-nonces
  • python
  • random-forests
  • rate-limiting
  • scikit-learn
  • shap
  • svms
  • three.js
  • tuning-hyperparameters-via-nested-optuna-tpe.-a-61-patient-holdout-was-strictly-sealed-until-final-evaluation
  • webgl
  • xgboost
Share this project:

Updates

Submission history