An interpretable machine learning tool for predicting perioperative cardiac events in patients scheduled for hip fracture surgery: insights from the multicenter LUSHIP study

Abstract

Background: Elderly patients undergoing surgery for hip fractures are at high risk for perioperative Major Adverse Cardiac Events (MACE), which can markedly compromise postoperative outcomes. This study aims to develop a machine learning (ML) based, interpretable tool to predict MACE using clinical and ultrasound-based variables in this population. Methods: We analyzed data from 877 patients in the multicenter LUSHIP study, incorporating demographics, Revised Cardiac Risk Index (RCRI), functional status, and preoperative lung ultrasound (LUS) scores. Multiple ML models were trained and validated using bootstrap resampling. The final ensemble meta-model combined GBM (Gradient Boosting Machine) and GLMNET (Elastic-Net Regularized Generalized Linear Models). Results: The ensemble model achieved an AUROC of 0.86, with sensitivity and specificity of 0.72 and 0.83, respectively. These results significantly improve over traditional tools such as the Revised Cardiac Risk Index (RCRI), particularly when used alone. A significant contribution of this work is the integration of lung ultrasound (LUS) as a non-invasive, bedside biomarker, which notably improved risk prediction compared to the performance of the individual LUS marker alone (AUC = 0.78). Relevant predictors for the ML model are LUS score, RCRI score, and patient age. A web-based Shiny application was developed to enable real-time personalized risk estimation. Conclusion: This interpretable ML model improves perioperative cardiac risk stratification and profiling in elderly hip fracture patients and may guide targeted preventive strategies and resource allocation. Trial registration: CT04074876.