Abstract 15728: Machine Learning Helps Predict Short and Intermediate-term Risk for All Cause Mortality in Patients Undergoing Percutaneous Coronary Intervention

Abstract

Background: Identifying patients at a high risk of mortality post percutaneous coronary intervention (PCI) is of vital clinical importance. We investigated the utility of machine learning algorithms to predict short and intermediate-term risk of all-cause mortality in patients undergoing PCI. Methods: Patient-level demographics, clinical, electrocardiographic ,echocardiographic and angiographic data from January 2006 to December 2017 were extracted from the Mayo Clinic CathPCI registry and clinical records. For patients with multiple PCI events, data collected at the time of the index PCI was used for analysis. Patients who underwent bailout coronary artery bypass graft surgery (CABG) prior to discharge were excluded. 306 variables were incorporated into random forest machine learning model (RF) to predict all-cause mortality at 6 months and 1 year after PCI. Ten-fold cross-validation repeated five times was used to optimize the hyperparameters and estimate its external performance. The National Cardiovascular Data Registry (NCDR) based logistic regression model was used for comparison. The area under receiver operator characteristic curves (AUC) was calculated to assess the ability of the models to predict all-cause mortality. Results: A total of 17356 unique patients were included for the final analysis after excluding 165 patients who underwent CABG surgery during the index hospitalization. The mean age was 66.9 ± 12.5 years;71% were male. Indications for PCI were ST-elevation myocardial infarction (9.4%), non-ST elevation myocardial infarction (12.9%), unstable angina (17.7%), and stable angina (52.8%) in the cohort. In-hospital, 6-month & 1 year mortality rates were 1.9%,4.2% & 5.8% respectively. The RF model was superior to the NCDR model in predicting inhospital, 6-month, 1 year mortality (p<0.0001) ( Figure 1 ). Conclusion: Machine learning is superior to NCDR model in predicting short and intermediate risk of all-cause mortality post PCI.