MACHINE LEARNING TO PREDICT DELIRIUM AND LONG-TERM COGNITIVE DECLINE FOLLOWING SURGERY

Prevention of postoperative delirium and long-term cognitive decline (LTCD) requires risk quantification that is valid in the general clinical setting. Existing prediction rules were primarily developed using stepwise regression, which tends to be overly influenced by specific data sets. We explored multiple different machine learning (ML) algorithms to predict delirium and LTCD (2–36 months). ML better protects against overfitting. We used model development and validation samples from the SAGES cohort (N = 560 adults ≥70 years). A cross-validated logistic regression approach generated the optimal predictive model for delirium (sensitivity 73%), while a random forest optimally prediction of magnitude of LTCD (R2 .20). Pre-operative cognitive level and comorbidity were important in predicting both outcomes, but the final models were otherwise distinct. These results demonstrate the utility of ML for generalizable prediction in this setting and underscores the contribution of baseline vulnerability to risk of delirium and cognitive decline.