Explainable machine learning model for in-hospital hypoglycemia risk in patients with latent autoimmune diabetes in adults.

Background: Latent autoimmune diabetes in adults (LADA) is characterized by progressive β-cell impairment and severe glycemic lability, predisposing patients to in-hospital hypoglycemia. Few tailored risk-stratification models exist for this population. This study aimed to develop and validate an interpretable machine learning model using routine clinical data to predict in-hospital hypoglycemia in LADA inpatients. Methods: This multicenter retrospective study recruited participants from five Chinese tertiary hospitals between January 2019 and September 2025. Data from four centers formed the derivation cohort, and the remaining center served as the independent external validation cohort. The primary endpoint was in-hospital hypoglycemia (blood glucose < 3.9 mmol/L). Three machine learning models, including logistic regression, random forest, and XGBoost, were developed using routine clinical data and assessed for discrimination, calibration, and clinical utility. SHAP analysis was applied to improve model interpretability. Exploratory subgroup analyses in the internal validation cohort examined model performance across clinical subgroups. Results: A total of 752 LADA inpatients were enrolled. The incidence of in-hospital hypoglycemia was 44.8% in the derivation cohort and 54.4% in the external validation cohort. Six core predictive factors were identified: largest amplitude of glycemic excursion, fasting C-peptide, glycated hemoglobin, sex, insulin pump use, and previous hypoglycemia. The three models yielded numerically variable discriminative performance across cohorts. Pairwise DeLong tests indicated no statistically significant differences in the AUROC among the three algorithms during external validation. All models showed comparable calibration and threshold-dependent predictive performance in the external cohort. XGBoost was selected as the final model after comprehensive evaluation. Fasting C-peptide was identified as the most influential predictor. Exploratory subgroup analyses demonstrated generally stable model performance across clinical strata. These findings are limited by small subgroup sample sizes and wide confidence intervals, and thus cannot be generalized to external populations. Sensitivity analysis suggested that model performance was not predominantly dependent on the retained glucose-derived predictor. Conclusions: The interpretable XGBoost model showed acceptable discrimination, calibration, and potential clinical utility for in-hospital hypoglycemia risk stratification in patients with LADA. This pragmatic predictive tool has the potential to support individualized inpatient glycemic management and facilitate targeted clinical intervention for LADA populations.