Advances from Traditional Clinical Decision Support Tools to Machine Learning Applications in Primary Prevention of Thromboembolic Disorders

Abstract Primary thromboprophylaxis for arterial and venous thromboembolism (VTE) still presents many challenges, as patient- and disease-specific risk factors complicate risk assessment and necessitate tailored strategies. To tackle this, multiple thromboembolism risk assessment models have been developed and validated, with machine learning (ML) models now emerging. Some ML models have been integrated into electronic health record (EHR)-based clinical decision support systems (CDSS) to improve guideline adherence and optimize appropriate anticoagulant use. We conducted a narrative review exploring the intersection of ML and clinical decision support (CDS) in primary thromboprophylaxis for VTE and stroke prevention in atrial fibrillation. Traditional risk assessment models integrated into EHR-based CDS have improved thromboprophylaxis rates and, in some studies, reduced thrombotic events. More comprehensive CDS strategies, such as the deployment of an EHR-agnostic CDS incorporating the IMPROVE-DD score for venous thromboprophylaxis, have demonstrated broader clinical impact, with reductions in both venous and arterial thromboembolism. With advances in artificial intelligence, ML-based risk prediction models for both venous and arterial thromboembolism have demonstrated superior discriminative performance compared with traditional models, predominantly in internal validation studies. Furthermore, their integration into active CDSS represents a promising strategy to improve thromboprophylaxis. However, most ML models lack external validation, remaining susceptible to overfitting and limiting generalizability. ML models integrated into active CDS hold much promise given their enhanced predictive performance, ability to integrate high-dimensional data and capture complex nonlinear interactions. Nevertheless, further work is needed to establish clinical utility through prospective outcome studies, ensure generalizability via external validation, and achieve integration into real clinical workflows.