AI-ENABLED APPROACHFORTHEEARLYDETECTIONOF GESTATIONAL DIABETES ANDRETINALVESSEL SEGMENTATION

Abstract

Gestational diabetes mellitus (GDM) is one of the most prevalent disease for pregnant woman in GCC. In the present work, we employed machine learning (ML) based approach for the early detection of GDM using only first trimester data from Qatar Biobank (QBB). The proposed stack ensemble based model was able to detect GDM with 98.6\% sensitivity, 80.6\% specificity, and 91.3\% accuracy in the cohort. While compared to the other existing model for GDM, we showed that our model performed superior to other existing model. Based on our analysis we have identified History of high glucose level/diabetes, NT pro-BNP, FT3, HOMA-IR- (22.5 Scale), Urea, Insulin, Magnesium, Prothrombin Time, Basophil Auto \%, MPV, Cholesterol, LDL-Calc, MCHC, Calcium, Fibriogen, Sodium, Homocysteine Plasma LC-MSMS, Total Protein, Alk Phos, Lymphocyte Auto \%, Basophil Auto #, Eosiophil Auto #, FT4 as potential biomarkers for the early detection of GDM. Moreover, our findings align to the findings from other GDM cohort. In the second part of this thesis, we focused on understanding the status of retinal vessel structure in diabetes and related co-morbidities. The structure of blood vessels in the retina is a crucial factor in identifying and forecasting various eye diseases like cardiovascular diseases, diabetes, and other diseases. Therefore, detecting the structure of blood vessels from retinal fundus images is a critical field of research in healthcare. This study employed a novel deep learning model to segment vessels for different diseases, including Glaucoma, Diabetic Retinopathy (DR), and Age-related Macular Degeneration (AMD). We considered multiple transfer learning-based models and discovered that the ResNet-based U-Net architecture was the most effective for vessel segmentation, achieving the highest Dice Score above 84\% for disease-agnostic, and 82\%-84\% for disease-specific conditions. We believe the proposed methodology will help to advance retinal vessel segmentation process and enhance the screening process of diseases based on retinal fundus images in clinical settings.

Date of Award	2024
Original language	American English
Awarding Institution	HBKU College of Science and Engineering