[Development of machine learning models based on multidimensional preoperative variables to aid in differentiating benign from malignant parathyroid tumors: a clinical study]

Objective: To develop and validate machine learning models based on multidimensional preoperative variables to aid in differentiating benign from malignant parathyroid tumors. Methods: This is a retrospective case series study. Data were collected from 2 773 patients with pathologically confirmed parathyroid tumors who underwent surgery in the Department of General Surgery, Peking Union Medical College Hospital between December 2010 and December 2025. There were 717 male and 2 056 female patients, with an age (M(IQR)) of 54 (18) years (range: 10 to 87 years). The cohort comprised 2 633 patients with benign tumors and 140 patients with malignant tumors. The cohort was divided by stratified random sampling in an 8∶2 ratio into a training set of 2 218 patients, including 112 malignant cases, and a test set of 555 patients, including 28 malignant cases. The test set was only used for final performance evaluation. Preoperative variables were collected. Models were developed using logistic regression, support vector machine, random forest, gradient boosting machine, extreme gradient boosting, and categorical boosting, and their diagnostic performance was compared. Class imbalance in the random forest model was addressed using balanced subsample class weights. Predicted probabilities from the random forest model were calibrated using sigmoid (Platt) calibration based on five-fold stratified cross-validation within the training set, and clinical net benefit was evaluated using decision-curve analysis. Results: In the test set, the random forest model achieved an area under the receiver operating characteristic curve of 0.903, an area under the precision-recall curve of 0.563, a sensitivity of 0.857, and a specificity of 0.799, showing favorable overall discriminative performance. After calibration, the Brier score was 0.035. Decision-curve analysis showed potential clinical net benefit across threshold probabilities of 2% to 20%. The final random forest model included 11 variables: age, history of renal calculi, tumor volume, tumor aspect ratio, tumor cystic change, serum calcium, plasma parathyroid hormone, alkaline phosphatase, serum phosphorus, absolute lymphocyte count, and neutrophil-to-lymphocyte ratio. Model interpretation showed that plasma parathyroid hormone, serum calcium, and tumor volume were the most important variables in this model. Conclusion: The random forest model based on multidimensional preoperative variables shows good performance in distinguishing benign from malignant parathyroid tumors and may be used as an adjunct to clinical decision-making.