Machine-learning prediction of postoperative delirium and characterization of clinical heterogeneity after cardiac surgery with cardiopulmonary bypass: a prospective observational cohort study

To develop and internally validate machine-learning models for predicting postoperative delirium (POD) after cardiac surgery with cardiopulmonary bypass (CPB), and to characterize clinical heterogeneity across data-driven patient subgroups within the same cohort. In this prospective single-center observational cohort study, 729 adult patients undergoing cardiac surgery with CPB were included. Multidimensional perioperative variables were analyzed, with postoperative predictors restricted to measurements obtained before the first documented delirium episode. Seven supervised machine-learning models were developed and internally validated for POD prediction, and unsupervised hierarchical clustering was performed in the full cohort to characterize clinical heterogeneity across data-driven patient subgroups. During the predefined postoperative observation period, 262 of 729 patients (35.9%) developed POD. In the independent test set, discrimination ranged from 0.718 to 0.861 across the seven models. In cross-validated analyses, random forest showed the highest discriminatory performance (AUC = 0.841) and the lowest overall prediction error by Brier score. In the clustering analysis, three data-driven subgroups with different observed POD burdens were identified, showing distinct perioperative profiles across inflammatory, functional, comorbidity, and procedural domains. POD was additionally associated with prolonged total hospital stay. Machine-learning models integrating preoperative, intraoperative, and early postoperative variables showed potential for predicting POD after cardiac surgery with CPB. In addition, unsupervised clustering provided a subgroup-level perspective on perioperative clinical heterogeneity within the cohort.