Background: Bleeding is a common complication of percutaneous coronary intervention (PCI), leading to significant morbidity, mortality, and cost. While several risk models exist to predict post-PCI bleeding risk, however these existing models produce a single estimate of bleeding risk anchored at a single point in time. These models do not update the risk estimates as new clinical information emerges, despite the dynamic nature of risk.
Objective: We sought to develop models that update estimates of patient risk of bleeding over time, enabling a dynamic estimate of risk that incorporates evolving clinical information, and to demonstrate updated predictive performance by incorporating this information.
Methods: Using data available from the National Cardiovascular Data Registry (NCDR) CathPCI, we trained 6 different XGBoost tree-based machine learning models to estimate the risk of bleeding at key decision points: 1) choice of access site, 2) prescription of medication prior to PCI, and 3) the choice of closure device.
Results: We included 2,868,808 PCIs; 2,314,446 (80.7%) prior to 2014 for training and 554,362 (19.3%) remaining for validation. Discrimination improved from an AUROC of 0.812 (95% Confidence Interval: 0.812-0.812) using only presentation variables to 0.845 (0.845-0.845) using all variables. Among 123,712 patients classified as low risk by the initial model, 14,441 were reclassified as moderate risk (1.4% experienced bleeds), while 723 patients were reclassified as high risk (12.5% experienced bleeds). Among 160,165 patients classified as high risk by the initial model, there were 40 patients reclassified to low risk (0% experienced bleeds), and 43,265 patients reclassified to moderate risk (2.5% experienced bleeds).
Conclusion: Accounting for the time-varying nature of data and capturing the association between treatment decisions and changes in risk provide up-to-date information that may guide individualized care throughout a hospitalization.