Inspiration
Cardiovascular disease is a serious global health problem, but clinical data can often be difficult to understand when it is presented only as tables, measurements, and percentages. We were inspired by a simple question: What if an AI prediction could be connected to the part of the heart it is actually talking about? Instead of showing a patient only a number such as 78% CAD risk, we wanted to create an experience where the prediction, the factors behind it, and the corresponding coronary vessel could be understood together. That idea led us to build CardioVision 3D — an AI-powered cardiovascular risk visualization system combining machine learning, explainable AI, and interactive 3D anatomy.
What it does
CardioVision 3D takes clinical, physiological, ECG, laboratory, and echocardiographic information and generates predictions for:
Overall Coronary Artery Disease (CAD) LAD stenosis LCX stenosis RCA stenosis
The predicted vessel-specific probabilities are visualized on an interactive 3D heart model, allowing users to rotate, zoom, and inspect the coronary vessels. The system also provides an explainability layer using SHAP, showing which patient features contributed most to the model's prediction. Our workflow can be summarized as: [ \text{Clinical Data} \rightarrow \text{ML Prediction} \rightarrow \text{SHAP Explanation} \rightarrow \text{3D Visualization} ] The goal is to make AI predictions easier to interpret and connect them with an intuitive anatomical representation. Important: CardioVision 3D is a prototype for educational and decision-support purposes. It is not a medical diagnosis and does not replace professional clinical evaluation or diagnostic imaging.
How I built it
We started with the Extension of the Z-Alizadeh Sani dataset, containing clinical and cardiovascular information from patients. We built the machine-learning pipeline in Python using tools such as Pandas, NumPy, Scikit-learn, and SHAP. Separate classification pipelines were created for overall CAD and the three target vessels — LAD, LCX, and RCA. We carefully removed LAD, LCX, RCA, and Cath from the model inputs when appropriate to prevent target leakage. We evaluated the models using: Accuracy Precision Recall F1-score ROC-AUC For the application layer, we connected the ML pipeline to a backend API and built an interactive frontend using React and Three.js / React Three Fiber. The frontend allows users to enter patient information, request predictions, view vessel-specific probabilities, inspect SHAP explanations, and interact with the 3D heart.
Challenges I ran into
One of our biggest challenges was the relatively small size of the dataset. With only a few hundred records, we had to be careful about overfitting and avoid treating a complex deep-learning architecture as automatically better. Another challenge was preventing data leakage. Because the dataset already contains vessel-level labels, we had to make sure those target variables were not accidentally used as model features. The 3D visualization also required careful thinking. The dataset provides vessel-level stenosis labels, but it does not contain patient-specific 3D anatomical lesion coordinates. Therefore, we could not honestly claim that the model had identified the exact physical location of a blockage. Instead, we designed the system to highlight the corresponding coronary vessel based on its predicted probability. Finally, integrating machine learning, SHAP explanations, an API, and a real-time 3D interface into one application required us to think about the entire system rather than just the ML model.
Accomplishments that I'm proud of
We are proud that CardioVision 3D goes beyond simply training a classification model. We created a complete workflow connecting: Patient Data → Prediction → Explanation → Anatomy The project combines several technologies that are normally demonstrated separately: Machine learning Explainable AI Clinical data processing REST API integration Interactive 3D visualization Modern web development We are especially proud of the vessel-level visualization because it gives the prediction a visual context instead of leaving it as another number on a dashboard
What I learned
This project taught us that building AI for healthcare requires much more than achieving a good metric. We learned about: Data preprocessing and leakage prevention Classification and model evaluation Probability interpretation Explainable AI with SHAP Connecting ML models to web applications Interactive 3D visualization with Three.js Designing interfaces for complex technical information The importance of communicating the limitations of medical AI Most importantly, we learned that visualization should never imply information that the underlying dataset does not contain. For example, clinical features such as RWMA can contribute to a prediction, but they should not be treated as direct 3D coordinates of a coronary lesion. That distinction helped us design CardioVision 3D as an interpretable visualization and decision-support prototype rather than presenting it as an automated diagnostic tool.
What's next for CardioVision 3D
CardioVision 3D is only the beginning. The next version could incorporate larger and more diverse clinical datasets, better probability calibration, additional cardiovascular risk factors, and more advanced anatomical models. A particularly exciting direction would be integrating patient-specific medical imaging, such as CT angiography, so that the visualization could eventually move from population-level vessel risk prediction toward patient-specific anatomical visualization. We also want to explore longitudinal patient data, allowing users to see how cardiovascular risk changes over time. Our ultimate goal is to continue exploring how AI + Explainability + 3D Visualization can make complex medical information easier to understand while keeping clinicians, diagnostic imaging, and established medical workflows at the center.
Built With
- fastapi
- fiber
- python
- react
- scikit-learn
- shap
- three
- three.js/react