Development and internal validation of an interpretable machine learning-based risk stratification model for sarcopenic obesity in patients with type 2 diabetes mellitus

Objective To analyze the factors associated with sarcopenic obesity (SO) in patients with type 2 diabetes (T2DM), and to develop and internally validate an SO risk stratification model based on interpretable machine learning, thereby providing guidance for the clinical screening of high-risk patients and the implementation of personalized risk assessment. Methods A single-center cross-sectional study design was adopted, enrolling 378 patients with T2DM, who were divided into an SO group and a non-SO group based on SO diagnostic results. Demographic characteristics, disease-related data, anthropometric measurements, and laboratory test results were collected. The cohort was randomly divided into a training set ( n = 265) and a validation set ( n = 113) in a 7:3 ratio. Univariate analysis, LASSO regression, and multivariate logistic regression were used to identify factors associated with SO, and based on these results, five risk-stratification models were constructed: logistic regression, random forest, SVM, LightGBM, and XGBoost. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), area under the precision-recall curve (AUPRC), accuracy, sensitivity, recall, precision, F1 score, and Brier score. Model calibration performance and clinical applicability were assessed using calibration curves, calibration slope, calibration intercept, calibration-in-the-large (CITL), and decision curve analysis (DCA). Supplementary internal validation of the LightGBM model was performed using 1,000 bootstrap resamples, and interpretability analysis was conducted using the SHAP method. Results In this study, the detection rate of SO in patients with T2DM was 36.2%, age, ADL, waist circumference, HDL, and 25(OH)D were the primary factors influencing SO. The AUC values of the five models on the validation set ranged from 0.806 to 0.821. Among them, LightGBM achieved the numerically highest AUC in this sample, at 0.821 (95% CI: 0.744–0.897), with an AUPRC of 0.712; however, the DeLong test indicated that the differences in AUC among the models were not statistically significant (all P > 0.05). Overall, the models demonstrated good calibration performance and potential clinical net benefit; however, bootstrap internal validation indicated that LightGBM still exhibited a certain degree of optimism bias and under-calibration. SHAP analysis revealed that age, ADL, waist circumference, HDL, and 25(OH)D contributed significantly to model predictions, with age being the most important predictive feature. Conclusion The SO risk stratification models developed using machine learning methods demonstrated good performance and potential clinical utility, with LightGBM achieving the highest numerical performance. SHAP-based interpretability analysis enhanced model transparency and clinical credibility, providing a reference for assisted screening, risk stratification, and individualized assessment of SO in patients with T2DM.