Inferring preoperative cognitive function from intraoperative electroencephalography in elderly patients using machine learning

To develop and evaluate machine learning (ML) models that infer preoperative cognitive function from intraoperative electroencephalography (EEG). This was a retrospective ML study that used a training dataset derived from the MINDDS study (306 patients, USA), and an external testing dataset from the Electroencephalographic Biomarker to Predict Acute Post-Operatory Cognitive Dysfunction study (92 patients, Chile). Both contained patients older than 60 years undergoing either cardiac (training dataset) or non-cardiac (testing dataset) surgery under general anesthesia. Preoperative cognitive function was assessed using the Montreal Cognitive Assessment (MoCA) in both cohorts. Four types of ML models were used: logistic regression with L2 penalty, random forest, gradient boosting tree, and extreme gradient boosting. Models were evaluated in terms of weighted root mean square error (WRMSE) and monotonic correlations towards actual MoCA scores (Spearman's rho). A logistic regression model with L2 regularization performed best in the training dataset (WRMSE 2.82 2.60 - 3.03 95%CI, Spearman's rho 0.18 0.06 - 0.29, p 0.0015). This performance mostly generalized to the test dataset (WRMSE 2.72 2.51 - 2.94, Spearman's rho 0.14 -0.05 - 0.31, p 0.18). This study shows that ML models trained on intraoperative EEG can effectively infer preoperative cognitive function in older patients, with generalizability across distinct populations and relatively low error (<3 MoCA points). However, the correlations were weak, indicating limited ability to capture consistent monotonic relationships. Incorporating this approach into perioperative care could enable early detection and mitigation of neurocognitive disorders, improving surgical outcomes through tailored interventions. Further refinement and validation are required before clinical implementation.