CardioScope3D: AI Cardiac Analysis & 3D Visualization

Inspiration

Cardiovascular diseases remain a significant healthcare challenge, and understanding coronary artery conditions requires analyzing complex patient data. We were inspired to explore how machine learning, explainable AI, and interactive 3D visualization could be combined to make cardiac data easier to investigate and understand. This led to CardioScope3D, a research-oriented platform designed to explore coronary artery disease-related predictions and visualize cardiac anatomy in an interactive environment.

What it does

CardioScope3D is an AI-powered web application that explores four prediction targets: Coronary Artery Disease (CAD), Left Anterior Descending (LAD), Left Circumflex (LCX), and Right Coronary Artery (RCA).

The system uses patient-related clinical features to generate model-based probability estimates. It combines these outputs with an interactive dashboard and 3D visualization to provide a more intuitive way to explore cardiac data. SHAP-based explainability is intended to help investigate how input features contribute to individual model predictions.

The platform is designed for research and educational purposes, rather than as a replacement for professional medical diagnosis.

How we built it

We developed CardioScope3D using a modular full-stack architecture that integrates a web interface, a Python backend, machine learning models, and 3D visualization.

Frontend: React.js and Vite for the user interface.
3D Visualization: Three.js for interactive cardiac visualization.
Backend: Python and FastAPI for API development and model inference.
Dataset: The Z-Alizadeh Sani dataset, containing 303 patient records and 55 input features.
Machine Learning: Scikit-learn and XGBoost, with Random Forest selected for CAD and LAD and XGBoost selected for LCX and RCA in the current training pipeline.
Explainable AI: SHAP for investigating feature contributions to model predictions.
Model Persistence: Joblib for saving and loading trained model artifacts.

The workflow takes patient-related inputs, validates and processes the features, passes them to the appropriate machine learning models, and presents the resulting estimates through the application interface.

Challenges we ran into

One of our main challenges was maintaining consistency between the dataset, feature preprocessing, trained models, and backend prediction requests. Integrating multiple models into a single application also introduced model serialization and compatibility issues.

Another challenge was connecting machine learning outputs with an intuitive 3D visualization while ensuring that probability estimates and feature explanations were presented responsibly. We also needed to consider input validation, API reliability, dependency compatibility, and the limitations of working with a relatively small dataset.

These challenges highlighted the importance of systematic testing, reproducible model training, and reliable integration between software components.

Accomplishments that we're proud of

Designed a unified platform combining machine learning and interactive 3D visualization.
Established a multi-target prediction workflow covering CAD, LAD, LCX, and RCA.
Integrated a structured clinical dataset with a modular backend architecture.
Explored explainable AI to improve the interpretability of model-generated predictions.
Built a foundation for future experiments in cardiac data visualization and machine learning evaluation.

Our key accomplishment is bringing these different technologies together into a single research-oriented application.

What we learned

Through CardioScope3D, we explored full-stack development, REST API integration, machine learning inference, dataset preprocessing, model persistence, and interactive 3D visualization.

We also learned about the practical challenges of deploying multiple trained models, preserving feature consistency, investigating serialization errors, and evaluating model reliability. The project reinforced the importance of distinguishing predictive performance from clinical validity and communicating the limitations of AI-based healthcare research.