Autoantibody Profiling for Accurate Differentiation of Type 1 and Type 2 Diabetes Mellitus in Omani Patients: A Retrospective Study

Abstract

Background/Objectives: Differentiating Type 1 from Type 2 diabetes mellitus (T1DM vs. T2DM) remains clinically challenging, especially in early-onset cases with overlapping features. This study assessed the diagnostic utility of diabetes-related autoantibodies in an Omani cohort and evaluated their predictive performance using machine learning. Methods: Clinical and laboratory data from 448 patients (aged ≥ 2 years) in Al Batinah North, Oman, were retrospectively analyzed. We assessed autoantibody positivity (anti-GAD, anti-islet, anti-TPO, anti-tissue), age, sex, and HbA1c. Receiver operating characteristic (ROC) curves and a neural network model were used to evaluate diagnostic accuracy. Results: Anti-GAD and anti-islet antibodies were significantly more prevalent in T1DM (69.0% and 64.1%) than T2DM (7.4% and 3.8%; p < 0.0001). HbA1c was elevated in both subtypes but lacked discriminatory specificity. Nearly half (48.5%) of T1DM patients showed multiple antibody positivity, especially in younger age groups. Anti-TPO and anti-tissue antibodies were more frequently detected in T1DM, suggesting broader autoimmunity. ROC analysis showed strong predictive value for anti-islet (AUC = 0.835) and anti-GAD (AUC = 0.827). Neural network modeling identified anti-GAD, anti-islet, and age as the most informative predictors, achieving over 92% classification accuracy. Importantly, antibody positivity in a subset of insulin-treated T2DM patients suggested potential latent autoimmune diabetes (LADA) misclassification. Conclusions: This is the first study in Oman to combine autoantibody screening with AI-based modeling to refine diabetes classification. Our findings highlight the value of immunological profiling in early diagnosis, uncover possible misclassification, and support AI integration to guide individualized management.