Predicting Hungry Bone Syndrome with Interpretable Machine Learning: A Single-Center Cohort of Dialysis Patients Undergoing Parathyroidectomy.

Background/

Objectives

Hungry bone syndrome (HBS) is a frequent and potentially life-threatening complication following parathyroidectomy (PTX) for secondary hyperparathyroidism (SHPT) in dialysis patients, yet existing prediction tools offer limited discriminative accuracy. This study aimed to develop and internally validate an interpretable machine learning (ML) framework for preoperative HBS prediction and to derive a pragmatic bedside risk score from ML-derived feature importance.

Methods

Ninety end-stage renal disease patients who underwent PTX for drug-refractory SHPT at a single center (2019-2023) were analyzed. Eight supervised ML classifiers were trained on 24 preoperative features (19 raw variables plus 5 engineered features) and evaluated under 5-fold stratified cross-validation repeated 10 times. SHapley Additive exPlanations (SHAP) analysis was applied for model interpretability, and a composite bedside risk score was constructed from SHAP-derived feature rankings.

Results

HBS occurred in 41 patients (45.6%). Random forest achieved the numerically highest discrimination among multi-feature models (AUC = 0.933 ± 0.065), outperforming previously published models, though univariate alkaline phosphatase (ALP) alone achieved a comparable cross-validated AUC of 0.958. ALP overwhelmingly dominated all predictors (mean |SHAP| = 3.37, exceeding the next-ranked feature by approximately 6.5-fold). Partial dependence analysis revealed a sigmoid-shaped ALP-HBS relationship with a critical inflection zone between 250-350 U/L, and SHAP dependence plots demonstrated that total parathyroidectomy amplifies ALP-mediated risk. A SHAP-guided composite bedside risk score (range 0-9) achieved an AUC of 0.883, with observed HBS rates rising monotonically from 0% (score 0) to 100% (score ≥ 6). Decision-curve analysis showed that univariate ALP and the multi-feature pipeline yielded comparable net benefit, with ALP preferable in the high-sensitivity regime and the multi-feature model preferable at high-specificity thresholds; net reclassification improvement was negative for the multi-feature model vs. univariate ALP, supporting the framework's role as an interpretive rather than discriminative advance.

Conclusions

An interpretable ML framework substantially improves HBS prediction over conventional models, confirms ALP as the overwhelmingly dominant predictor through a nonlinear dose-response relationship, and yields a clinically interpretable bedside risk score that, pending external validation, may support preoperative risk stratification.