A body shape index-based machine learning model for predicting postoperative delirium and three-year mortality: a prospective cohort study

Background Postoperative delirium (POD) is a serious complication linked to adverse outcomes. This study examined the association of A Body Shape Index (ABSI) with POD and three-year mortality in patients with POD and developed a machine-learning (ML) model for POD prediction. Methods We included 855 patients undergoing knee/hip arthroplasty. POD was assessed using the Confusion Assessment Method. Multivariable logistic regression, restricted cubic splines, and subgroup analyses were used to examine the association between ABSI and POD. Generalized structural equation modeling assessed potential mediation by A β 42, T-tau, and P-tau. Ten ML models were developed using eight predictors selected through training-set statistical assessment and consideration of clinical relevance. The cohort was stratified into training and held-out test sets (8:2), with model development performed exclusively in the training set. Kaplan–Meier analysis compared three-year mortality between low- and high-ABSI POD patients. Results POD occurred in 84 patients (9.82%). ABSI was significantly associated with POD in the crude model and after adjustment for traditional clinical risk factors; however, this association was attenuated and became non-significant after further adjustment for CSF biomarkers. The multivariable logistic model incorporating ABSI and 10 covariates showed an apparent AUC of 0.801, with an optimism-corrected AUC of 0.779 after bootstrap internal validation. Bootstrap analyses showed no statistically significant indirect effect through A β 42, T-tau, or P-tau. Among the ten ML models, the Gradient Boosting Machine (GBM) showed the most favorable overall performance, with a mean repeated cross-validation AUC of 0.805 ± 0.066 and a held-out test-set AUC of 0.848 (95% CI, 0.761–0.915). The test-set Brier score was 0.073. SHAP analysis ranked ABSI as the fourth most influential predictor after A β 42, age, and hypertension. In POD patients, high ABSI was associated with significantly higher three-year mortality ( P = 0.0415). Conclusion Elevated ABSI was associated with an increased risk of POD following knee/hip arthroplasty; however, no statistically significant indirect effect through A β 42, T-tau, or P-tau was identified. Among patients who developed POD, higher ABSI was associated with increased three-year mortality. The GBM model incorporating ABSI showed favorable internally validated discrimination and calibration, although external validation is required before clinical implementation. Clinical trial registration https://www.chictr.org.cn/showproj.html?proj=54473 , Chinese Clinical Trial Registry (ChiCTR), ChiCTR2000033439. Registered 1 June 2020.