Immune-inflammatory and metabolic signatures for osteoporosis risk stratification in primary Sjögren's syndrome: development and internal validation of an interpretable machine-learning model.

Background: Primary Sjögren's syndrome (pSS) is a systemic autoimmune disease. Osteoporosis (OP) is a common complication in patients with pSS (pSS-OP), which significantly affects their quality of life and prognosis. Currently, there is a lack of efficient and objective clinical tools for the early identification of pSS patients at high risk of osteoporosis, limiting the implementation of precise interventions. Objective: This study aims to integrate clinical indicators and immunological characteristics to develop and validate a machine learning model for predicting the risk of osteoporosis in patients with pSS, thereby facilitating early clinical identification and decision-making. Methods: Clinical data were collected from 384 patients with pSS. Lymphocyte subsets and serum cytokine levels of IL-2, IL-4, and IL-6 were measured in all participants. Missing data were handled using multiple imputation, followed by intergroup comparisons. Feature selection was performed using Lasso regression, random forest, and stepwise regression, and the intersection of these methods was used to identify the final predictors. Based on the selected features, nine machine learning models were constructed and compared, including decision tree, k-nearest neighbor, logistic regression, elastic net, random forest, support vector machine, multilayer perceptron, LightGBM, and XGBoost. Model performance was evaluated using ROC curves, calibration curves, decision curve, accuracy, F1 score, sensitivity, specificity, and other metrics. The SHAP method was applied to interpret the optimal model. Results: Patients were divided into a pSS with OP and a pSS without OP group. Six core predictors were identified: Age, ESR, Urea, MON, Ca, and Fibrinogen. Among the nine models, the XGBoost model demonstrated the best performance. In the training set, the AUC was 0.890, and the F1 score was 0.90. In the test set, the AUC was 0.808, and the F1 score was 0.81. SHAP analysis revealed the following order of feature importance: Age, MON, ESR, Ca, Urea, and Fibrinogen. Among these, Ca contributed negatively to the model prediction, while the remaining features contributed positively. Conclusion: This study developed and internally validated an XGBoost machine learning model based on clinical and laboratory indicators for predicting osteoporosis risk in patients with pSS. The model demonstrated good discrimination and may assist early risk assessment.