AI-ENABLED ELUCIDATION OF DUODENAL PATHOPHYSIOLOGY IN TYPE 2 DIABETES: PAVING THE WAY FOR NOVEL THERAPEUTIC APPROACHES

Background: The duodenum is integral as a metabolic signaling center, making it a promising therapeutic target for Type 2 Diabetes (T2D). However, the specific pathophysiological alterations in the duodenum under metabolic stress, especially in human, are not well-understood. This study aims to elucidate these changes, taking advantage of the duodenum's accessibility for sampling via esophagogastroduodenoscopy (EGD) and the sophisticated analytical capabilities of artificial intelligence (AI).
Methods: We analyzed duodenal biopsies from T2D patients and non-diabetic controls (non-DM) (2010-2017), with grossly normal esophagogastroduodenoscopy (EGD) exam. Biopsies were selected based on normal histology, confirmed by a blinded GI pathologist. These samples were digitized into whole slide images (WSIs). Using a deep Gaussian process (DGP)-based model in a weakly supervised learning setup, we segmented each WSI into 256x256 pixel tiles. The DGP model evaluated each tile for T2D markers, with a geometric pooling model aggregating these probabilities for slide-level T2D likelihood after performing 5X cross-validation. Additionally, we applied the Getis-Ord-Gi* statistic to identify 'lymphocyte hotspots' in the slides, employing deep learning for insights into stroma and lymphocyte distribution.
Results: Our study encompassed 231 patients, comprising 111 with Type 2 Diabetes (T2D) and 120 non-diabetic (non-DM) individuals. Detailed baseline demographics for both groups are presented in Table 1. A total of 320 duodenal biopsy slides were analyzed, including 172 from T2D patients and 148 from non-DM patients. The Deep Gaussian Process (DGP) model's predictive efficacy for T2D was evaluated, with its overall performance indicated by an Area Under the Curve (AUC) of 0.7591 (p=0.0507). Furthermore, the model achieved an F1 score of 0.7471 (p=0.0378), reflecting its accuracy in binary classification. Utilizing the Getis-Ord-Gi* statistic, we identified a significantly higher prevalence of lymphocyte clustering, termed 'lymphocyte hotspots,' accompanied by stromal expansion in the lamina propria of T2D patient biopsies compared to controls (p < 0.001), as illustrated in Figure 1. This finding underscores a distinct immunological pattern in the duodenal mucosa of T2D patients.
Conclusion: Our investigation identified significant lamina propria and submucosal inflammation in the duodenum of Type 2 Diabetes (T2D) patients, suggesting a new mechanistic pathway for T2D pathology and potential therapeutic interventions. These findings highlight the duodenum as a novel target for T2D treatment, warranting further research into strategies for modulating duodenal inflammation.