Single-vessel patient: LAD 92% (red) and its supply territory shaded on the 3D heart; LCX and RCA low (green).
Click an artery: the camera flies to it and shows its probability, top SHAP factors and the region it supplies.
Why this prediction? Exact SHAP waterfall from the average patient (60%) to this patient (92%).
Every measurement against its normal range, beside its contribution to the prediction.
What-if sliders: change a value and the probabilities, explanations and 3D colours update in under a second.
Normal coronaries: low risk everywhere, all green.
Works on phones too.
Landing page.
Architecture: one config-driven pipeline from data to 3D.
Nested-CV ROC-AUC per model, with ensemble and bedside baseline (95% CI).
Decision curves: every model beats “treat all” and “treat none”.

Tagline

See coronary risk in 3D and understand why: per-artery stenosis prediction from routine clinical data, mapped onto an interactive heart.

Inspiration

Cardiovascular disease is still a leading cause of death. Risk models usually end in a single percentage, which tells a clinician or a patient little about where the problem may be or why the model thinks so. We wanted a tool that turns routine clinical data into an answer you can see and question.

What it does

CardioScope takes a patient's demographics, examination, ECG, lab and echocardiography values and predicts:

the probability of overall coronary artery disease (CAD);
the probability of significant stenosis in each of the three main arteries: LAD, LCX and RCA.

How the results are shown:

Interactive 3D heart: each artery is coloured by its probability, from green through amber to red, inside a transparent torso.
- The heart muscle each artery supplies is tinted with the same risk colour.
- You can rotate, zoom and click an artery to fly to it and see its details.
Clinical dashboard:
- CAD gauge and per-vessel bars;
- an exact SHAP waterfall showing why the model predicted this risk;
- a physiology panel showing every measurement against its normal range, beside its contribution;
- a vessel detail panel;
- an "In plain words" summary.
Live updates: editing any value or moving a what-if slider updates the probabilities, the explanations and the 3D colours in under a second.
Example patients: one click loads real normal, single-vessel and multi-vessel patients.
Disclaimer: a permanent banner says it is for decision support and education only, not a substitute for angiography. We predict per-vessel status, not lesion location.

How we built it

Data:

The Extension of Z-Alizadeh Sani dataset (UCI): 303 patients, 54 model features.
features.yaml describes every feature: type, unit, normal range and group. The same file drives training, API validation and the dashboard form.

Machine learning (Python, scikit-learn, XGBoost, CatBoost):

For each of the four targets we compare logistic regression, random forest, XGBoost and CatBoost.
Nested cross-validation: repeated stratified 5-fold × 3, with hyper-parameter tuning in an inner loop. All preprocessing is refit inside every fold.
Calibration and thresholds: Platt calibration; decision thresholds tuned for recall (≥ 90 % CAD, ≥ 80 % vessels); bootstrap 95 % confidence intervals.
Leakage guard: an automated check stops LAD/LCX/RCA/Cath from ever being inputs.
Exact SHAP: native TreeSHAP for trees, exact linear SHAP for logistic regression.
Extra evaluation:
- a 10-item bedside baseline;
- a pre-specified ensemble;
- decision-curve analysis;
- a subgroup check by sex, age and diabetes.
One command trains everything, and two runs give identical results.

Backend:

FastAPI with /schema, /predict, /explain and /patients/sample.
The input model is generated from the feature config.
Deployed as a Docker service on Render.

Frontend:

React + TypeScript + React Three Fiber, deployed on Vercel.
The heart is a sculpted CC-licensed model, split into separate LAD/LCX/RCA meshes and compressed to 438 KB.
Rendering happens on demand, so it runs smoothly on an ordinary laptop with no GPU.
One shared vessels.json keeps the 3D arteries, API and model outputs consistent.

Quality: 111 automated tests (ML, API, frontend) run in GitHub Actions on every push.

Results (nested CV, mean [95 % CI])

Target	ROC-AUC
Overall CAD	0.93 [0.90–0.96]
LAD	0.84 [0.80–0.88]
LCX	0.73 [0.67–0.78]
RCA	0.73 [0.68–0.78]

Every model beats "treat all" and "treat none" in decision-curve analysis.
For LAD the full model clearly beats a bedside baseline (+0.064 AUC).
For LCX and RCA it does not, and we say so.

Challenges we ran into

Honest validation on only 303 patients. We used nested CV, paired comparisons and bootstrap CIs instead of chasing a single high score.
Mapping risk to anatomy without overclaiming. The data gives per-vessel status, not lesion position, so the 3D view colours whole arteries and their supply territories. It never shows a "lesion spot".
A calibration solver that could diverge on spread-out scores. We found it, fixed it with a line search and added a regression test.
Security software: Windows Smart App Control blocked scikit-learn on our machine, so training moved to Linux (WSL), which also matches the deployed server.

Accomplishments that we're proud of

Explanations are exact: SHAP contributions add up to each displayed probability.
One config-driven pipeline from data to 3D. Adding a feature, model or vessel is a config change.
Under-a-second live what-if updates on a free deployment.

What we learned

On small clinical datasets, simple well-calibrated models are as good as complex ones. Honest uncertainty (CIs, reliability notes, subgroup gaps) builds more trust than a single big number.

What's next for CardioScope

External validation on a second hospital's data, under ethics approval and with clinical partners.
Vessel-specific inputs, e.g. coronary CT or perfusion imaging, to improve LCX/RCA.
Patient-specific anatomy.