Abstract

Background: Sjögren's Syndrome (SjS) is a complex autoimmune disease predominantly affecting the salivary and lacrimal glands, leading to hallmark symptoms such as dry eyes and mouth. Its distinctive clinical manifestations, coupled with symptom overlap with other disorders, make its diagnosis challenging. Traditional diagnostic criteria, as outlined by the American College of Rheumatology (ACR), encompass clinical assessments, including Anti-Ro/SSA antibody levels, focus score, and quantification of dry eye and mouth. Despite their utility, these criteria often suffer from issues of reproducibility and sensitivity. This underscores the pressing need for more advanced and reliable diagnostic techniques in the field. Objectives: In this work we developed a deep learning algorithm that combines clinical data with histopathological images of minor salivary glands in order to predict for Sjögren's syndrome. By analyzing H&E stained labial gland biopsies using a Convolutional Neural Network (CNN) encoder in conjunction with a Multi-Layer Perceptron (MLP) and the ACR criteria. This strategy aims to improve diagnostic precision above and beyond the capabilities of traditional techniques, providing a notable improvement in the diagnosis and treatment planning of Sjögren's Syndrome patients. Methods: Data for this study was sourced from the DIApSS (Diagnostic Suspicion of Primitive Sjögren's Syndrome - Brest Cohort, NCT03681964) observational study. Our dataset comprised 167 patients, 95 confirmed Sjögren's patients and 72 non-Sjögren's participants. Patients were split into three groups training: 99 for training, 33 for validation, and 33 for testing. Our pipeline includes a Multi-Layer Perceptron (MLP) and a Convolutional Neural Network (CNN) encoder, implemented using PyTorch. The MLP processes clinical data: gender, xerophthalmia symptoms, Schirmer's test, and Anti-Ro/SSA (UA/mL) and the CNN encoder, which processes H&E stained labial gland biopsy images. The combined output of both models is then fed into a classification head for prediction. Results: Evaluating the diagnostic precision using the clinical data alongside H&E Whole Slide Images (WSI), we obtained an Area Under the Curve (AUC) of 0.98, Accuracy of 0.86, kappa score 0.69, and a recall of 0.86. Cohen's kappa score of 0.69 indicates that the model's predictions show a moderate degree of agreement that goes beyond chance. Conclusion: This study successfully demonstrates the potential of integrating clinical data and histopathological images to obtain significant diagnostic accuracy. This approach paves the way for more reliable and reproducible diagnoses of Sjögren's Syndrome.In future studies we will focus on refining the trained models and exploring further the analysis of the histopathological images. REFERENCES: NIL Acknowledgements: NIL. Disclosure of Interests: None declared.