Effectiveness of artificial intelligence method in diagnosis of primary sjögren disease

Objective: The use of artificial intelligence (AI) models in the field of rheumatology is increasingly supported by evidences. This study aims to investigate the diagnostic performance of an AI model based on deep learning using parotid ultrasonography (PUSG) images in the diagnosis of primary Sjögren's syndrome (pSS). Methods: The study included 123 pSS patients meeting the 2016 ACR-EULAR classification criteria and a control group of 123 age- and sex-matched healthy volunteers. PUSG scans were performed on all the participants. For the AI deep learning model, a training set consisting of 392 images from 196 volunteers (98 pSS patients and 98 controls), a validation set comprising 48 images from 24 volunteers (12 pSS patients and 12 controls), and 52 PUSG images from 26 volunteers (13 pSS patients and 13 controls) for deep learning analysis were used. The study examined the performance of the deep learning system in diagnosing pSS based on PUSG images. Results: The sex distribution (p=1.00) and age averages(pSS group: 57.3 ± 11.8 years; control group: 55.1 ± 11.2 years) were similar between the pSS and control groups (p=0.15). The mean age at diagnosis for the 123 pSS patients was 50.6 ± 11.5 years. When evaluating PUSG images of pSS patients and healthy individuals according to OMERACT criteria, the grades of the right and left parotid glands were consistent (Kappa: 0.656 and Kappa: 0.777, respectively). When it is analyzed with a total OMERACT score ≥2, the diagnostic accuracy, sensitivity, and specificity of PUSG for pSS were 87%, 82%, and 92%, respectively. The AI model's diagnostic accuracy, sensitivity, and specificity for pSS were 87%, 77%, and 96%, respectively. Conclusion: The AI model programmed using PUSG images demonstrated high diagnostic accuracy for pSS. It is conceivable that AI models could be integrated into ultrasound devices to serve as image-based diagnostic tools. Parotid ultrasonography, in conjunction with these models, could become a viable diagnostic method for pSS