Dual antiplatelet therapy (DAPT) with aspirin and a P2Y12 inhibitor is the standard of care for secondary prevention after percutaneous coronary intervention (PCI). DAPT has demonstrated its efficacy in reducing ischemic complications (including stent thrombosis) after PCI although at the cost of an increased risk of bleeding. As both event types have been independently linked with excess morbidity and mortality, international guidelines emphasize the need to tailor DAPT duration and intensity according to the individual ischemic and bleeding risk profile of each patient. In this context, several predictive risk models for bleeding and thrombosis have been developed with the aim of guiding clinical decisions on DAPT management post-PCI. However, many of these risk models have shown only modest performance and limited applicability in real-world clinical practice. Such limitations can be attributed, at least in part, to the analytical approaches used for their development, mostly based on linear models unable to capture the complex interplay between different clinical covariates. Machine learning methods offer the potential to overcome these limitations by leveraging computer algorithms to large datasets that capture high-dimensional, non-linear relationships among variables. However, the feasibility and usefulness of machine learning-based prognostic risk models in PCI patients remain relatively unexplored.
The present study will analyze data from 11 clinical trials encompassing approximately 19,000 patients undergoing percutaneous coronary intervention (PCI) with an everolimus-eluting stent to develop a machine learning-based algorithm. Institutional review board approval or informed patient consent was not required as this study is an analysis of previously published clinical trials and all individual patient data were deidentified. The goal is to predict the probability of an ischemic or bleeding event up to 1 year after PCI. Patient-level data from the eligible clinical trials listed per the XIENCE Machine Learning Data Acquisition Protocol (90961902) will be pooled and randomly split into a training cohort (\~75%) and a validation cohort (\~25%). These include both Abbott- sponsored and investigator-initiated XIENCE studies (i.e., XIENCE V, XIENCE 28 USA, XIENCE 28 GLOBAL, XIENCE 90, ABSORB III, ABSORB IV, Compare ABSORB, Compare Acute, EXAMINATION, SIERRA-75 and ITALIC). The performance of different trials of machine learning classifiers will be compared with traditional statistical approaches for the prediction of ischemic and bleeding outcomes. The best-performing machine learning model will then be selected and tested against a pre-defined performance goal to assess its clinical usefulness. Based on existing literature on established risk scores that currently inform clinical practice, the performance goal for the model is set at C-index value equal or greater than 0.65 at the 97.5% lower confidence interval of the bootstrap C-index distribution. This is to ensure that the true value of the C-index is still within a clinically relevant range and to validate the clinical usefulness of the risk prediction model.
