A Clinically Aligned Two‐Stage Machine Learning Framework for Predicting Hungry Bone Syndrome After Parathyroidectomy

BACKGROUND: Hungry bone syndrome (HBS) is a frequent and clinically significant complication following parathyroidectomy (PTX) in patients with secondary hyperparathyroidism (SHPT), often leading to prolonged hypocalcaemia and increased healthcare burden. Existing prediction models are limited by small sample sizes and inability to capture complex clinical interactions. This study aimed to develop and validate a clinically aligned, two-stage machine learning (ML) framework to predict HBS after PTX. MATERIALS AND METHODS: A retrospective cohort of patients undergoing PTX for SHPT between 2008 and 2025 at a tertiary centre was analysed. A two-stage ML framework was constructed: stage 1 used preoperative variables to generate a risk score, and stage 2 integrated this score with intraoperative features. Multiple ML models were evaluated using area under the receiver operating characteristic curve (AUROC), calibration metrics and resampling techniques. RESULTS: A total of 882 patients were included, with an HBS incidence of 69.9%. EasyEnsemble and logistic regression demonstrated the highest discrimination (AUROC 0.712), outperforming the k-nearest neighbours baseline. EasyEnsemble achieved the best overall performance (accuracy 0.707, F1 score 0.666) and calibration (Brier score 0.186). Key predictors included elevated preoperative alkaline phosphatase, higher intact parathyroid hormone levels and lower serum calcium. CONCLUSION: This two-stage ML framework demonstrated acceptable predictive performance and aligns with clinical decision-making processes. It enables early identification of high-risk patients and may support individualised perioperative management to mitigate HBS and its complications.