Abstract 17973: Leveraging Machine Learning Techniques to Forecast Patient Prognosis After Percutaneous Coronary Intervention
Introduction: Contemporary risk models for event prediction after percutaneous coronary intervention (PCI) have limited predictive ability. Machine learning (ML) methods have the potential to identify complex non-linear patterns, improving predictive power.
Hypothesis: ML can be used to derive better discriminatory models to identify patients at risk for in-hospital death and congestive heart failure (CHF) rehospitalization after PCI.
Methods: We evaluated 11,709 distinct patients that underwent 14,349 PCIs during 14,024 admissions between 2004-2013 who were prospectively followed in the Mayo Clinic PCI registry. 53 demographic and clinical parameters known at the time of admission and 356 additional parameters available at discharge were examined to identify patients at risk for readmission due to CHF and in-hospital mortality. For each target event, we trained a random forest regression model to estimate the time to event. 8 fold cross-validation was used to estimate model performance. We used the predicted time to event as a score and generated a ROC curve. After choosing a desired operating point on the curve, sensitivity, specificity, and PPV were calculated. Bootstrap technique was used to estimate confidence intervals. Using input permutation and linear regression parameters, the contribution of each predictor was estimated.
Results: The predictive algorithm identified a high risk cohort consisting of 2% of all patients with a risk of in-hospital mortality of 45.5% (95% CI: 43.5-47.5%) compared to a risk of 2.1% for the general population (AUC 0.925; 95% CI: 0.92-0.93), using only parameters known at admission. Age, CHF, and shock on presentation were leading predictors for the outcome. A high risk group comprised of 1% of all patients was identified with a risk of 30 day CHF rehospitalization of 8.1% (95% CI: 6.3-10.2%) compared to 0.7% for the whole population (AUC 0.90; 95% CI: 0.89-0.91), using parameters known at discharge. CHF history, post-procedure hemoglobin, and discharge hematocrit were leading predictors.
Conclusions: Our cross-validated ML algorithm was more predictive and discriminative than standard regression methods at identifying patients at the highest risk for in-hospital mortality and 30 day CHF rehospitalization.
- Coronary interventions
- Coronary artery disease
- Percutaneous coronary intervention (PCI)
- Quality improvement
Author Disclosures: C. Senecal: None. C.J. Zack: None. Y. Kinar: None. Y. Bar-Sinai: None. Y. Metzger: None. G. Hilevitz: None. A. Lerman: Consultant/Advisory Board; Modest; Medial Research. R. Gulati: None.
- © 2016 by American Heart Association, Inc.