Abstract
Brain morphometry derived from clinical imaging has an underexplored potential for the multimodal prediction of postoperative delirium (POD), an acute encephalopathy that can lead to long-term adverse outcomes or death. This study conducted a comprehensive analysis of patient trajectories, integrating magnetic resonance imaging (MRI) data and electronic health records (EHRs) across two general surgical cohorts. We applied univariate test methods and linear mixed-effects models correcting for confounding. Non-linear multi-layer perceptrons (MLPs), boosted decision trees, and logistic regressions were trained on EHR data, brain morphometry measures, and their multimodal fusion to predict POD. Age-adjusted correlations identified cortical thickness of temporal gyri, as well as thalamic and brainstem volumes to be POD-relevant neuroanatomical features. MLP models demonstrated robust predictive capability, achieving notably high performances up to 86% AUROC (area under the receiver operating characteristic). Multimodal fusion yielded pronounced benefits in less critically ill patients. MLP model weights showed high predictive potential for cerebral atrophy in higher-order cortical regions, including the temporal pole, superior frontal gyrus, and the insula. These findings reveal the previously unrecognized potential of clinically derived brain morphometry in enhancing early multimodal predictions of POD. A better understanding of brain vulnerability in POD may translate into improved clinical decision making based on multimodal health care data.