Abstract
Sjögren's disease is often an underdiagnosed autoimmune condition that primarily affects the exocrine glands, resulting in symptoms such as dry eyes and dry mouth. Diagnostic challenges stem from nonspecific symptoms, the absence of definitive biomarkers, and the invasiveness and cost associated with current clinical tests. This study investigates the application of attenuated total reflectance Fourier-transform infrared spectroscopy (ATR-FTIR), in conjunction with a neural network and Monte Carlo dropout-based uncertainty estimation to distinguish patients with Sjögren's disease from healthy individuals. Uncertainty estimates were used to identify ambiguous spectra and refine training data, improving predictive performance and enabling confidence-aware interpretation. To minimize bias related to sex-based prevalence, model training was limited to female participants, among whom the disease is more common. To guarantee independent validation, a stratified cross-validation strategy was used. In each fold, the network is initialized with random weights, trained exclusively on the training subset, and evaluated on the corresponding unseen data. The refined (uncertainty-aware) model achieved 75% accuracy and 73% sensitivity at the individual spectrum level, and 79% accuracy and 76% sensitivity at the patient level. These findings highlight the potential of an uncertainty-aware, noninvasive, machine learning-based method as a reliable tool for the early diagnosis of Sjögren's Disease.