Risk prediction modelling of 30-day all-cause mortality following percutaneous coronary intervention in an Australian...

Preprocedural risk prediction of 30-day all-cause mortality after percutaneous coronary intervention (PCI) aids in clinical decision-making and benchmarking hospital performance. This study aimed to identify preprocedural factors to predict the risk of 30-day all-cause mortality post-PCI using machine learning (ML) approaches.

The study analysed 93 055 consecutive PCI procedures recorded in the Victorian Cardiac Outcomes Registry in Australia. The Boruta feature selection method was used to identify key predictive variables. Seven ML algorithms were employed for models' development and validation. Models' performance was assessed using standard metrics for validation data set. SHapley Additive exPlanations method was used to explain leading predictive variables.

Among the seven ML algorithms, the Extreme Gradient Boosting (XGB) model had the better performance across most metrics, such as accuracy (86.7%), root mean square error (36.5%), specificity (82.5%), precision (54.0%), F1 score (52.7%) and Brier score (13.3%). The XGB model also demonstrated strong discriminatory power, achieving a receiver operating characteristics-area under the curve of 85.5% (95% CI 83.5% to 87.4%). The XGB model identified left ventricular ejection fraction, acute coronary syndrome, estimated glomerular filtration rate, age and complex lesions as the five leading factors associated with 30-day mortality post-PCI. Other factors, in order, were cardiogenic shock, body mass index, intubated out-of-hospital cardiac arrest, lesion location, mechanical ventricular support, gender and peripheral vascular disease.

The XGB model demonstrated the best performance in predicting 30-day all-cause mortality post-PCI, identified most influential predictors such as severely reduced ejection fraction, ST-elevation myocardial infarction presentation, severe renal impairment, age 80 years and older and complex lesion. These factors from the XGB model could support individualised risk assessment, informed clinical decision-making, improved patient care or efficient resource utilisation for an Australian population. Further external validation is essential to confirm the model's generalisability across different populations.