Machine Learning for Cardiovascular Prevention Prescriptions: Real-World vs. Synthetic Data

Area of Science:

• Cardiovascular Medicine
• Artificial Intelligence in Healthcare
• Data Science

Background:

• Cardiovascular diseases are a leading global cause of mortality.
• Atherosclerosis is a primary driver, managed by lipid-lowering agents like statins and berberine.

Purpose of the Study:

• To investigate machine learning (ML) for predicting expert prescriptions of lipid-lowering therapies.
• To evaluate the performance of ML models using real-world data (RD) and synthetic datasets (Avatar-based and Statistical Data Vault).

Main Methods:

• Feature selection (automatic and manual) was performed on clinical data.
• Multiple ML and neural network models were trained and cross-validated using the F1-score.
• Performance was assessed on held-out datasets, comparing RD, SD_avatar, and SDV-based data.

Main Results:

• Real-world data achieved the highest predictive performance (F1 = 0.67).
• Avatar-based synthetic data (alone or combined with RD) showed comparable performance (F1 = 0.67).
• Statistical Data Vault synthetic data yielded significantly lower performance (F1 = 0.39).

Conclusions:

• Machine learning can effectively model prescription behavior for lipid-lowering therapies.
• Avatar-generated synthetic data shows promise for augmenting or replacing real-world data in certain ML applications.
• A gap persists in ML models' ability to fully replicate the nuanced clinical reasoning of expert prescribers.