CardioScope Dashboard
Anterior View
Posterior View
Basal view
Apical View
RiskMap Anatomy

Inspiration

Cardiovascular disease is the world's leading cause of death, but most ML risk tools hand you a single number with no explanation. A clinician or student looking at "87% risk" can't tell why the model said that, or which vessel is driving it.

We built CardioScope for Track A of the Multimodal AI Hackathon 2026 to make cardiac risk predictions something you can explore, question and understand, not just read.

What it does

Enter 54 clinical inputs (demographics, risk factors, comorbidities, exam findings, symptoms, ECG, labs and echocardiography) and CardioScope returns four predictions:

Overall CAD probability
Stenosis probability for each of three major vessels: LAD, LCX and RCA

Those results are mapped onto an interactive, beating 3D heart:

🟢 ≤25% · 🟡 25–75% · 🔴 ≥75% vessel coloring
Click any vessel, label or risk card to see the top 5 SHAP factors pushing that score up or down, next to the exact measurements used
Rotate and zoom, switch anatomical camera views, toggle labels, expand the viewer, or play a synthesized heartbeat
Export the assessment as a PDF

The heart is illustrative procedural anatomy, not a patient scan. Colors show model probability, not measured blockage.

How we built it

Data: UCI Extension of Z-Alizadeh Sani dataset, 303 patients, 59 columns. After removing the 4 targets and 1 constant column, 54 inputs remain.

Models: four CatBoost classifiers, one per target. Five-fold stratified, out-of-fold results:

Target	ROC-AUC	F1	Recall
Overall CAD	0.912	0.900	0.912
LAD	0.849	0.819	0.842
LCX	0.739	0.617	0.597
RCA	0.703	0.507	0.465

Explainability: SHAP using interventional probability attribution, so contribution units are consistent across all four models.

Backend: FastAPI + Pydantic. The API loads all four models and their explainers at startup and serves POST /api/predict.

Frontend: React 19, TypeScript, Vite, Tailwind CSS 4, Zustand, Three.js and React Three Fiber. A shared vessel-ID system keeps the risk cards, 3D labels, arteries and SHAP panel synchronized.

Deployment: one Docker image serves the React build and the FastAPI backend together on Render, with real model inference, not a mock.

Challenges we ran into

Not overstating certainty. Probabilities display as <0.1% or >99.9% instead of 0% or 100%. The color bands are visualization aids, not validated clinical thresholds, and the UI says so.
Stale explanations. If a user edits the form after a prediction, the old SHAP values would no longer match the inputs. We flag results as "Inputs changed" and keep the original measurements next to the explanation until the user reruns.
Small data. With only 303 records, LCX and RCA are noticeably weaker than CAD and LAD. We report this openly instead of averaging it away.
Target leakage. Cath, LAD, LCX and RCA are never used as predictors, and tests check this.
Keeping 4 views in sync. The form, vessel cards, 3D scene and explanation panel all had to agree on one selected vessel and one set of results.

Accomplishments that we're proud of

A live, deployed app returning real predictions and SHAP explanations
Explanations tied directly to anatomy: select a vessel and see why it scored that way
Tests covering API contracts, input boundaries, feature schema, target leakage, model artifacts and SHAP consistency
A six-page technical report that documents the limitations, not just the wins
Graceful fallback: if the API is down, the app offers a clearly labeled sample assessment rather than faking a result

What we learned

Attribution is not causation. SHAP shows what the model used, and the interface has to say that.
A prediction without its limitations is misleading. Showing LCX and RCA as weaker earns more trust than hiding it.
Keeping a 3D scene, a form and an ML explanation panel consistent comes down to one shared state store.

What's next for CardioScope

External validation on larger, more diverse cohorts
Probability calibration, so a displayed 80% means something
More modalities: ECG waveforms, echocardiography images and clinical notes
Usability testing with clinicians and students

CardioScope is a research and educational prototype. It is not a medical device, not a diagnosis, and not a substitute for formal diagnostic imaging or professional medical assessment.