An accurate predictive model for depressive symptoms in patients with primary Sjögren's disease based on machine...

To develop and validate a machine-learning (ML) model that flags primary Sjögren's disease (pSjD) patients at high risk of depressive symptoms for earlier clinical attention.

We retrospectively studied 147 pSjD patients (Nanjing First Hospital, 2019-2022). Depressive symptoms were screened with Patient Health Questionnaire-9 (PHQ-9); PHQ-9 ≥5 was the primary endpoint. Missing data were handled by multiple imputation. Data were split 70/30 for training/testing. After univariate screening and LASSO selection, eight ML algorithms (e.g., logistic regression, support vector machine (SVM), tree/boosting methods) were trained with stratified 10-fold cross-validation. Performance was summarised by AUROC/AUPRC, accuracy, precision/recall, Brier score, and calibration; SHAP provided model explainability.

Four routinely available predictors were retained: fatigue frequency, sleep duration, lymphocyte count, and anti-Ro52 status. Across repeated cross-validation, SVM showed the best overall discrimination (mean AUROC≈0.90) with strong precision and accuracy. In the held-out test set, SVM maintained high performance (AUROC=0.929; AUPRC=0.959; Brier=0.106). SHAP confirmed predictor importance, indicating higher risk with shorter sleep, lower lymphocyte counts, greater fatigue frequency, and anti-Ro52 positivity.

This study presents the first ML-based model for predicting depressive symptoms in pSjD patients, highlighting the significance of immuno-inflammatory and clinical factors in depression pathogenesis. The SVM model offers a robust, non-invasive tool for early identification of high-risk individuals, enabling timely and personalised interventions. However, this single-centre, retrospective design with a modest sample limits generalisability; therefore, independent multi-centre validation is required before clinical use.