MACHINE LEARNING-BASED PREDICTION OF POSTOPERATIVE DELIRIUM IN ELDERLY ANESTHESIA PATIENTS USING MULTIMODAL PERIOPERATIVE BIOMARKERS AND ELECTRONIC HEALTH RECORDS

Abstract

The postoperative delirium (POD) is one of the most common complications after anesthetic in elderly patients and may cause morbidity, prolonged hospital stay, poor long-term cognitive function and cost of health care. The authors created and validated an internal machine-learning based early POD (E-POD) risk stratification model that incorporates multimodal periop markers and the longitudinal EHR data (L-EHR). A retrospective study of 1248 patients of elderly age that received non-cardiac surgery was performed to assess demographic, comorbidities, pre-operative medication history, time-series signals (inhalation, heart rate, respiratory rate) during surgery, and biochemical inflammatory markers. The missing data were imputed multiple times and the most informative (and non-redundant) predictors were selected using feature selection algorithms Boruta and LASSO. Logistic regression, random forest, XGBoost, a temporal transformer and an ensemble model of all of the above were the candidate models. The ensemble model proposed showed to be the most effective model in the validation phase with an AUC of 0.902, average precision of 0.847, sensitivity of 0.858, specificity of 0.819 and F1 score of 0.838. Furthermore, calibration analysis was used to support the clinical reliability, with a Brier score of 0.091 and Spiegelhalter z-test P-value of 0.214. The baseline cognition, mean arterial pressure during surgery, CRP, age, anesthetic depth variability, IL-6, comorbidity burden, and surgery duration were identified as key risk drivers via explainability analysis. Considering these outcomes, a multimodal machine learning system could potentially predict POD in geriatric surgical patients with accurate and meaningful predictions that are easily understood and can be acted upon. It may help to identify risk at an earlier stage, provide personalised prevention and support improved peri-OP decision making, but it needs further external and prospective validation before it can be implemented in clinical practice.