Development and validation of machine learning-derived frailty index in predicting outcomes of patients undergoing percutaneous coronary intervention.

Introduction: Frailty is associated with increased mortality in patients with percutaneous coronary intervention (PCI). Existing operationalized frailty measurement tools are limited and require resource intensive process. We developed and validated a tool to identify and stratify frailty using collected data for patients who underwent PCI and explored its predictive power to predict adverse clinical outcomes post PCI. Methods: Between 2014 and 2015, 1,732 patients who underwent semi-urgent or elective PCI in a tertiary centre were included. Variables including demographics, co-morbidities, investigations and clinical outcomes to 33 ± 37 months were analysed. Logistic regression model and Extreme Gradient Boosting (XGBoost) machine learning model were constructed to identify predictors of adverse clinical outcomes post PCI. The final models' predicted probabilities were assessed with area under receiver operating characteristic curve (AUC). Results: With model analysis, frailty index (FI), age and gender were the 3 most important features for adverse clinical outcomes prediction, with FI contributing the most. After adjustment, the odds of FI to predict cardiac death and in-hospital death post PCI remained significant [1.94 (95 %CI1.79-2.10); p < 0.001, 2.04(95 %CI 1.87-2.23); p < 0.001 respectively]. The XGBoost machine learning models improved predictive power for cardiac death [AUC 0.83(95 %CI 0.80-0.86)] and in hospital death [AUC 0.83(95 %CI 0.80-0.86)] post PCI compared to logistic regression models. Conclusion: The resultant model developed using novel machine learning methodologies had good predictive power for significant clinical outcomes post PCI with potential to be automated within hospital information systems.