Comprehensive Geriatric Assessment–Augmented Risk Stratification for Postoperative Delirium After Spine Surgery: A Cross-Surgery Benchmarking and Model-Updating Study

Abstract Background Postoperative delirium (POD) after spine surgery in older adults reflects both baseline geriatric vulnerability and perioperative stress. Although externally derived electronic health record models may provide useful benchmarks, their direct transportability across surgical domains is uncertain. We evaluated whether comprehensive geriatric assessment (CGA)–augmented local model updating improves POD risk stratification in older adults undergoing spine surgery. Methods This retrospective benchmarking and model-updating study used an in-house spine-surgery cohort of 162 older adults, including 16 POD events, and a publicly available external hip-fracture cohort of 245 patients, including 30 POD events. A harmonized predictor set included demographics, comorbidities, nutritional and laboratory reserve, operative burden, and postoperative context. Dataset shift was assessed using standardized mean differences and a cohort-membership model, followed by overlap weighting. Predictor transportability was examined using cohort-specific univariable log odds ratios. Prediction models were evaluated in staged fashion: external-to-spine transport, recalibration, local core updating, CGA augmentation, and a simplified bedside score. Discrimination, calibration, Brier score, and decision-curve analysis were assessed. Results POD occurred in 30 of 245 patients in the external hip-fracture cohort and 16 of 162 patients in the in-house spine cohort. Among POD cases, the spine cohort was younger but had longer anesthesia or surgery duration. Before weighting, the cohorts showed substantial case-mix imbalance, with a maximum absolute standardized mean difference of 1.72; overlap weighting reduced the maximum residual imbalance to 0.16. The cohort-membership model showed near-separation between cohorts (AUROC 0.98). The transported external model showed modest discrimination in the spine cohort (AUROC 0.62), and recalibration did not improve discrimination (AUROC 0.62). Local updating improved AUROC to 0.88, and CGA augmentation further improved AUROC to 0.94. CGA vulnerability showed a graded association with POD incidence across quartiles: 0/41, 1/40, 6/40, and 9/41. A simplified frozen bedside score retained good discrimination (AUROC 0.86). At a score threshold of ≥ 3, sensitivity was 94%, specificity 70%, positive predictive value 25%, and negative predictive value 99%. Conclusions Risk domains from an external hip-fracture cohort showed partial conceptual transportability to older spine-surgery patients, but local updating with measured CGA variables substantially improved risk stratification. A CGA-augmented bedside score may provide a practical rule-out tool for targeting delirium-prevention resources, although prospective multicenter validation is needed.