CardioVisionAI — Project Story

Inspiration

Cardiovascular diseases are among the major causes of death worldwide, and identifying potential risks early can help people seek timely medical evaluation. However, cardiovascular data can be complex, making it difficult to understand the relationship between clinical measurements and potential coronary artery disease.

We created CardioVisionAI to explore how artificial intelligence, machine learning, and interactive 3D visualization can make cardiovascular risk information easier to understand. Our goal is to transform clinical data into meaningful predictions and visual insights that can support medical education, research, and clinical decision-support workflows.

What it does

CardioVisionAI is an AI-powered cardiovascular risk prediction and visualization project focused on coronary artery disease (CAD).

The system is designed to:

• Predict the likelihood of overall coronary artery disease using relevant clinical and demographic features.
• Estimate vessel-specific risk for the left anterior descending (LAD), left circumflex (LCX), and right coronary artery (RCA).
• Use machine learning models to analyze structured clinical data.
• Present prediction results through a user-friendly dashboard.
• Visualize the heart and major coronary vessels in an interactive 3D environment.
• Explore explainable AI techniques, including SHAP, to help users understand which input features influence model predictions.

CardioVisionAI is intended for educational and research purposes. Its predictions are experimental and are not a substitute for professional medical advice, diagnostic imaging, or evaluation by a qualified healthcare professional.

How we built it

We developed CardioVisionAI using a modular approach that combines machine learning, data processing, backend development, and frontend visualization.

Machine learning and data processing: We used the Z-Alizadeh Sani cardiovascular dataset as the basis for experimentation. Python and pandas were used to inspect and preprocess the data, while scikit-learn provided the preprocessing pipelines, model training, and evaluation tools.

Model development: We experimented with Logistic Regression, Random Forest, and HistGradientBoosting models. The workflow includes separate prediction tasks for overall CAD and the LAD, LCX, and RCA targets. We also considered data leakage prevention by excluding target-related columns from the input features.

Model evaluation: We evaluated model performance using accuracy, precision, recall, F1-score, and ROC-AUC. These metrics help us understand different aspects of prediction quality rather than relying on accuracy alone.

Backend and frontend: We organized the project into a Python backend and a web-based frontend to support model integration and user interaction.

3D visualization and explainability: We prepared heart and coronary-vessel 3D assets for visualization and explored SHAP-based explanations to make model behavior more interpretable.

Challenges we ran into

One of our main challenges was working with a relatively small clinical dataset. Limited data makes it difficult to ensure that model performance generalizes to patients outside the dataset.

We also had to pay close attention to preprocessing and feature selection. Clinical datasets may contain categorical values, missing information, and columns that could introduce target leakage if used incorrectly.

Another challenge was comparing multiple models across different prediction targets. A model that performs well for overall CAD may not perform equally well when predicting risk associated with an individual coronary artery.

Integrating machine learning with an interactive web interface and 3D anatomical assets introduced additional development and dependency-management challenges. We also explored explainability tools while resolving software-version compatibility issues.

These challenges helped us recognize that building a useful AI application requires more than training a model: data quality, reproducibility, interpretability, responsible presentation, and integration all matter.

Accomplishments that we're proud of

• Built an organized end-to-end machine learning workflow for cardiovascular risk prediction.
• Experimented with multiple classification algorithms and compared their performance using several evaluation metrics.
• Developed separate prediction pipelines for overall CAD and the LAD, LCX, and RCA targets.
• Saved trained models and supporting label mappings for future integration.
• Prepared a modular project structure spanning data processing, model development, backend code, frontend assets, and documentation.
• Explored explainable AI and 3D coronary anatomy visualization to make the project more understandable and interactive.

Most importantly, we transformed a healthcare-focused idea into a practical AI/ML engineering project and identified the steps needed to improve its reliability and usability.

What we learned

Through CardioVisionAI, we learned that responsible healthcare AI requires careful data handling, appropriate evaluation, and transparent communication of uncertainty.

We gained practical experience with data preprocessing, scikit-learn pipelines, classification algorithms, model comparison, evaluation metrics, model serialization, and explainability workflows.

We also learned about the challenges of connecting a machine learning pipeline to a web application and combining predictive outputs with anatomical visualization.

A key lesson was that good performance on a small dataset does not establish clinical reliability. Models need stronger validation, careful bias assessment, and testing on independent data before they can be considered for real-world clinical use.

What's next for CardioVisionAI

Our next goal is to make CardioVisionAI more interactive, interpretable, and technically robust.

Planned improvements include:

• Completing the integration of trained models with the web dashboard.
• Developing interactive 3D heart and coronary-vessel visualization with vessel-specific risk indicators.
• Integrating SHAP explanations into the user interface.
• Improving data validation, error handling, and reproducibility.
• Performing stronger validation using suitable independent datasets and more rigorous evaluation methods.
• Assessing model calibration, potential biases, and performance across relevant patient groups.
• Improving documentation, testing, deployment, and the project demonstration.

Our long-term vision is to build an educational and research-oriented platform that connects clinical data, machine learning, explainable AI, and interactive anatomy visualization in one place.

CardioVisionAI represents our step toward making complex cardiovascular data more understandable through responsible and interpretable AI.

Built With

• 3d
• ai
• artificial
• cardiovascular
• css
• data
• explainable
• github
• healthcare
• html
• intelligence
• javascript
• learning
• machine
• numpy
• pandas
• prediction
• python
• risk
• scikit-learn
• shap
• visualization