A machine-learning-assisted logistic regression model for predicting post-operative delirium in older adults undergoing total knee arthroplasty

Background Post-operative delirium (POD) is a common complication in older adults undergoing total knee arthroplasty (TKA). Procedure-specific predictive models remain limited. This study aimed to develop a preliminary, internally validated, machine-learning-assisted logistic regression model using routinely available perioperative variables. Methods We retrospectively included 451 patients aged ≥60 years undergoing TKA. The cohort was randomly split into a training set (70%) and a validation set (30%). A two-stage selection was applied: Random Forest ranked candidate predictors, and LASSO regression reduced dimensionality. Selected variables were entered into a multivariable logistic regression model to estimate odds ratios (OR) and 95% confidence intervals (CI). Model performance was assessed in the validation set using the area under the receiver operating characteristic curve (AUC). Results POD occurred in 137 patients (30.4%). Six variables were selected using Random Forest and LASSO. Logistic regression showed intraoperative hypoxia as the strongest predictor (OR 6.69, 95% CI 2.90–16.6), followed by history of surgery (OR 2.55, 95% CI 1.22–5.32) and higher ASA class (ASA 2: OR 2.31, 95% CI 1.20–4.54; ASA 3: OR 2.01, 95% CI 0.99–4.15). Creatinine (per umol/L) and airway management type also showed smaller associations, while time from admission to surgery was not statistically significant. The model achieved an AUC of 0.704 in the validation set, indicating moderate discrimination. Conclusion In older adults undergoing TKA, intraoperative hypoxia, prior surgery, and higher ASA class were associated with POD risk. This preliminary machine-learning-assisted logistic regression model showed moderate discrimination in internal validation and should be externally validated before routine clinical implementation.