Prediction of Venous Thromboembolism in Diverse Populations Using Machine Learning and Structured Electronic Health...

Area of Science:

• Computational biology
• Medical informatics
• Health services research

Background:

• Venous thromboembolism (VTE) poses a significant global health burden.
• Existing risk assessment tools for VTE have limitations in accuracy and applicability.
• There is a need for improved methods to predict VTE diagnosis and long-term risk.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting VTE diagnosis.
• To create ML models for predicting the 1-year risk of VTE.
• To utilize structured electronic health record (EHR) data for enhanced VTE risk prediction.

Main Methods:

• Trained and validated ML models on a large dataset (159,005 participants) from the Mount Sinai Data Warehouse.
• Externally validated models on diverse populations from the UK Biobank (401,723 participants) and All of Us (123,039 participants).
• Developed models of varying complexity (small, medium, large) to balance portability and performance.

Main Results:

• Models achieved high performance in predicting VTE diagnosis (AUC 0.72–0.83) and 1-year risk (AUC 0.64–0.78) on external test sets.
• ML models significantly outperformed the Padua score in predictive accuracy.
• Performance remained robust across diverse patient subsets, including various ethnicities and clinical statuses.

Conclusions:

• ML models leveraging EHR data offer a significant improvement for VTE diagnosis and 1-year risk prediction in diverse populations.
• These models can identify key risk factors, providing insights into VTE pathophysiology.
• Integration into EHR systems can enhance real-time risk assessment, leading to better VTE prevention and reduced morbidity/mortality.