Multivariate Sequential Analytics for Cardiovascular Disease Event Prediction.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Cardiovascular Disease Research

Background:

• Automated clinical decision support systems (CDSS) are crucial for cardiovascular disease (CVD) risk assessment and early intervention.
• Current CVD risk models lack explicit temporal patient history modeling, potentially limiting prediction accuracy.

Purpose of the Study:

• To investigate the impact of explicitly modeling temporal patient history on CVD event prediction accuracy.
• To compare the performance of sequential modeling techniques, including LSTM, against traditional methods.

Main Methods:

• Utilized multivariate sequential modeling, with a focus on long short-term memory (LSTM) recurrent neural networks.
• Integrated CVD decision support data with national health datasets (pharmaceuticals, hospitalizations, lab results, deaths).
• Employed a 2-year observation window and a 5-year prediction window for CVD death or hospitalization events in patients with lipid-lowering therapy history.

Main Results:

• Temporal models demonstrated value in 5-year CVD event prediction, with LSTM achieving the highest performance (AUROC 0.801, average precision 0.425).
• Non-temporal models, like ridge classifiers, were competitive, indicating the importance of temporal data even when aggregated.
• LSTM significantly outperformed common regression models like logistic regression and Cox proportional hazards.

Conclusions:

• Deep temporal models, especially LSTM, offer significant advantages for clinical decision support in chronic disease management.
• Explicitly incorporating patient history as a time series improves CVD risk prediction accuracy.
• LSTM-based models show promise for enhancing targeted early interventions in cardiovascular disease.