Explainable artificial intelligence for predicting mortality in geriatric patients undergoing hip arthroplasty

We performed machine learning analysis of population data to determine the major determinants of 1-year mortality among geriatric patients undergoing hip arthroplasty. The data of 17,290 patients aged ≥ 65 years who underwent hip arthroplasty in 2019 were extracted from the Korea National Health Insurance Service claims database. With 1-year mortality as the dependent variable, random forest variable importance and Shapley Additive Explanations (SHAP) were used to identify the major predictors among 31 included predictors and their associations with mortality. The area under the curve of the random forest model was 74.6%. The top 10 predictors were age, red blood cell transfusion, male sex, dementia, low socioeconomic status, presence of solid tumors, history of congestive heart failure, history of chronic kidney disease, statin use, and history of peripheral vascular disease. Univariate analysis and SHAP analysis identified age (maximum SHAP value 0.17), red blood cell transfusion (0.10), male sex (0.09), dementia (0.04), low socioeconomic status (0.03), presence of solid tumors (0.06), history of congestive heart failure (0.04), history of chronic kidney disease (0.05), general anesthesia (0.02), and iron use (0.02) as positive contributors to model predictions of mortality. Higher SHAP values indicate greater contributions to model output relative to the baseline prediction. We constructed an effective prediction model for predicting mortality among patients undergoing hip arthroplasty. Appropriate interventions should be provided for high-risk patients (i.e., advanced age, multiple comorbidities, or anticipated transfusion needs).