Cardiovascular Disease Risk Screening

Inspiration

Cardiovascular disease remains one of the leading causes of preventable death, and many severe outcomes occur simply because risk is detected too late. Traditional screening tools rely on linear scoring systems that fail to capture complex physiological interactions and often miss high-risk individuals. This project was inspired by the need for an early-warning, sensitivity-first screening system that prioritizes identifying at-risk patients before symptoms appear.

What it does

The system is an AI-powered cardiovascular risk screening model designed for early detection. Instead of maximizing accuracy, it prioritizes recall (sensitivity) to minimize false negatives. By combining engineered clinical features with a hybrid neural network architecture, the model identifies individuals who are likely at risk and flags them for further medical evaluation.

How we built it

We built the system using a Hybrid Wide-and-Deep Neural Network trained on structured cardiovascular health data. Key steps included:

Extensive feature engineering (blood pressure ratios, BMI, metabolic interactions, vascular aging indicators)

Standardization and robust preprocessing of clinical variables

A sensitivity-weighted loss function to penalize missed risk cases

Threshold optimization to align predictions with real-world screening goals

Model explainability using SHAP to ensure clinical interpretability The pipeline was implemented using PyTorch, Scikit-Learn, and SHAP.

Challenges we ran into

One major challenge was balancing sensitivity and precision—maximizing recall inevitably increases false positives. Another challenge was working with imperfect and partially aligned biomedical data, especially when ECG information was not available for every patient. Designing a model that remains reliable under these constraints required careful feature design and validation.

Accomplishments that we're proud of

Achieved an ROC-AUC of ~0.74 while maintaining ~88% recall for high-risk patients

Successfully designed a screening-oriented evaluation strategy rather than a generic classifier

Built an explainable AI pipeline suitable for clinical decision support

Demonstrated that cohort-level physiological priors can improve model robustness even without individual ECGs

What we learned

We learned that in healthcare AI, accuracy is not the right metric for every problem. Screening systems must be optimized for clinical impact, not leaderboard scores. We also gained practical experience in handling real-world medical data, applying explainable AI, and aligning machine learning objectives with medical decision-making.

What's next for Cardiovascular Disease Risk Screening

Next, we plan to:

Integrate patient-specific ECG embeddings for fully multimodal prediction

Validate the model on external clinical datasets

Introduce uncertainty estimation for safer deployment

Develop a clinician-friendly dashboard for risk interpretation

Explore longitudinal modeling to track cardiovascular risk over time

Built With

• matplotlib
• numpy
• pandas
• python
• pytorch
• scikit-learn
• seaborn