CardioVision AI

Dashboard
New Assessment
Health Insight
3D Heart Explorer
History

Inspiration

CardioVision AI was inspired by the need to bridge the gap between statistical machine learning predictions and intuitive spatial anatomical understanding in cardiovascular care. Traditional predictive models generate abstract risk percentages, but they often fail to give clinicians and patients a clear visual representation of where pathology is developing inside the heart.

What it does

CardioVision AI is a web-based clinical decision support tool that predicts:

Overall Coronary Artery Disease (CAD) risk
Vessel-specific stenosis ($\ge 50\%$ narrowing) for the:
- Left Anterior Descending (LAD)
- Left Circumflex (LCX)
- Right Coronary Artery (RCA)

It takes clinical, demographic, ECG, lab, and echocardiographic inputs and instantly maps the predicted risk probabilities onto an interactive 3D heart model. Each artery is color-coded from low risk (green) to severe stenosis risk (red).

To ensure transparency, an Explainable AI (XAI) dashboard powered by SHAP and LIME breaks down feature importances, alongside mandatory clinical safety disclaimers.

How we built it

The project uses a decoupled architecture:

Layer	Technologies
ML pipeline	Python 3.12, trained on the UCI Z-Alizadeh Sani dataset, with strict target exclusion protocols to prevent data leakage
Backend	FastAPI, serving real-time predictions and feature attributions
Frontend	React 19, Vite, Tailwind CSS v4, Recharts, Zustand
3D visualization	React Three Fiber (Three.js) with custom shaders applied to isolated vessel nodes within open-source `.gltf` meshes

Challenges we ran into

Strictly segregating direct target features (LAD, LCX, RCA, Cath) from the input parameters.
Isolating sub-meshes on the 3D heart without distorting the surrounding geometry.
Optimizing Zustand state management and SHAP computations to deliver responsive, real-time 3D updates without requiring high-end GPUs.

Accomplishments that we're proud of

Achieving a zero target leakage ML pipeline
Fluid spatial risk rendering on 3D artery nodes
Transparent, interpretable predictions in an accessible browser environment

What we learned

The clinical correlation between regional wall motion abnormalities (RWMA) and localized ischemia
Advanced 3D WebGL rendering techniques
Rigor in preventing data leakage

What's next for CardioVision AI

Automated pipelines to convert patient-specific CT/MRI angiograms into custom 3D meshes
Extending vessel segmentation to branch arteries such as the Posterior Descending Artery (PDA)
Integrating EHR and HL7 FHIR standards for automated data ingestion
Adapting the 3D WebGL canvas into WebXR for immersive VR pre-procedural planning

Built With

ai
explainable-ai
fastapi
machine-learning
numpy
pandas
python
react
react-three-fiber
rest-api
scikit-learn
shap
tailwind.css
three.js
typescript
vite

Submitted to

Multimodal AI Hackathon 2026

Created by

Kushagra Kumar: ML & Data Pipeline Lead
Built the Python 3.12 machine learning pipeline. He cleaned and scaled the UCI Z-Alizadeh Sani data with pandas and numpy, and enforced zero target leakage by excluding LAD, LCX, RCA, and Cath from the inputs. He trained and evaluated multi-output models (scikit-learn, XGBoost, LightGBM) for overall CAD and per-vessel stenosis, using ROC-AUC and precision-recall metrics. He saved the trained pipeline with joblib for the backend.

Ashutosh Sahu: Backend API & Explainable AI Lead
Built the FastAPI backend served with uvicorn. Pydantic v2 schemas define strictly typed request and response contracts. He integrated SHAP and LIME so every prediction returns feature attributions along with the vessel probabilities, and he configured CORS so the frontend could connect to the API.

Kshitij Pandey: 3D Graphics & Spatial Visualization Lead
Built the interactive 3D heart with React Three Fiber and Three.js. He isolated the vessel nodes in the open-source .gltf mesh and applied custom shaders that color each artery from green to red by predicted risk. He synced the UI and 3D canvas with Zustand and profiled performance with r3f-perf so it runs without a high-end GPU.

Vedant Mishra: Frontend UI, Integration & Documentation Lead
Built the React 19, Vite, and Tailwind CSS v4 interface: the multi-step assessment form, the results dashboard, and the history page. He visualized SHAP contributions and risk trends with Recharts and connected the UI to the FastAPI backend. He also integrated the pieces into one working app and wrote the project documentation and the 6-page technical submission.

ASHUTOSH SAHU
Kshitij Pandey
Kushagra Kumar
Vedant Mishra