Machine-learning models for preoperative prediction of new-onset hypocalcemia within 48 h after intertrochanteric fracture surgery in older adults

Background Older adults undergoing surgery for intertrochanteric fractures are vulnerable to early perioperative metabolic disturbances. We evaluated preoperative machine-learning models and compared them with single-predictor and conventional regression benchmarks for new-onset hypocalcemia within 48 h after surgery. Methods This single-center retrospective study included 332 patients aged ≥65 years with preoperative corrected calcium concentrations ≥2.15 mmol/L. Hypocalcemia was defined as an albumin-corrected calcium concentration <2.15 mmol/L. The primary analysis used Boruta feature selection and stratified five-fold outer cross-validation repeated five times, with four-fold inner cross-validation for hyperparameter tuning by AUROC. Imputation, preprocessing, feature selection, and tuning were repeated within each outer training fold. Models were compared using patient-level repeated out-of-fold predictions. Results New-onset hypocalcemia occurred in 92 patients (27.7%). The out-of-fold AUROCs were 0.639 for corrected calcium alone, 0.615 for albumin alone, and 0.442 for measured total calcium alone, compared with 0.670 for conventional logistic regression, 0.671 for a generalized additive model, and 0.6829 (95% CI: 0.6177–0.7407) for Stable Kernel Naive Bayes. The AUROC difference between Stable Kernel Naive Bayes and corrected calcium alone was 0.044 (95% CI: −0.022 to 0.105). Stable Kernel Naive Bayes had a PR-AUC of 0.4276 (95% CI: 0.3401–0.5337) and a Brier score of 0.1853 (95% CI: 0.1631–0.2078). Corrected calcium was selected in all 25 outer training folds. Chronological evaluation using 2015–2022 for development and 2023–2025 for evaluation yielded an AUROC of 0.7083 (95% CI: 0.5957–0.8210). Conclusion Preoperative corrected calcium captured much of the available predictive information. Multivariable models showed modest discrimination, without a clear incremental advantage over corrected calcium alone. The findings support further evaluation of preoperative risk stratification; effects on monitoring and patient outcomes require prospective assessment.