0519: A Multi-criterion Feature Integration Framework for Accurate Diagnosis of Primary Sjögren's Syndrome Using Routine Laboratory Tests: A Multicentre, Retrospective Cohort Study

Background/Purpose: Primary Sjögren's syndrome (pSS) is a prevalent systemic autoimmune disorder. However, the early and accurate diagnosis of pSS remains challenging due to its non-specific clinical manifestations, the absence of definitive molecular markers, and the limitations of current diagnostic tests. The present study aims to evaluate the diagnostic utility of routine laboratory tests for predicting pSS and to develop a robust, cost-effective and generalizable integrated artificial intelligence (AI) model to assist in the identification of pSS.

Methods: This multicentre, retrospective cohort study included data from 99 laboratory items of 27,432 patients with and without pSS who were admitted to one regional centre hospital and two local hospitals in China between January 1, 2013, and January 31, 2023. The construction of Sjögren Multi-criterion Feature Integration Framework (SMFIF) model was based on 16 optimal features, using multi-algorithm feature selection, SHAP value analysis, model performance evaluation, and ensemble learning strategies. The model was evaluated using an internal validation set comprising 9,329 individuals, and an external validation set consisting of 545 individuals. The primary outcome was the prediction accuracy of the model in identifying pSS.

Results: The SMFIF model attained an area under the receiver-operating characteristic curve (AUC) of 0.929 (95% CI, 0.923-0.935) in the testing set, 0.934 (95% CI, 0.929-0.939) in the internal validation set, and 0.964 (95% CI, 0.943-0.986) in the external validation set. The model demonstrated higher AUC, ACC and sensitivity in comparison with anti-SSA/Ro, anti-SSB/La, antinuclear antibodies (ANA), Globulin (GLO), Rheumatoid factor (RF), immunoglobulin G (IgG), and Complement (C3 and C4) in the diagnosis of pSS. The superior performance of SMFIF in pSS prediction was confirmed through calibration curves and confusion matrices. The SMFIF was developed as a publicly available prediction tool that will provide an estimated probability of pSS based on input laboratory data.

Conclusion: The SMFIF model has been shown to consistently achieve satisfactory performance in the prediction of pSS when utilising routine laboratory tests from the testing set and two validation sets. This model offers a low-cost, easily accessible, and accurate diagnostic tool for pSS.