Longitudinal Analysis of Risk Factors Associated with Diabetes Using Multimodal Data and Machine Learning

Abstract

Diabetes mellitus is a chronic metabolic disorder characterized by persistent hyperglycemia due to impaired insulin function. It is a rapidly growing global health concern, with projections estimating 693 million affected adults by 2045. Its rising prevalence, along with complications such as nephropathy, retinopathy, and neuropathy, highlights the need for early prediction and intervention. In this context, machine learning and multimodal data fusion provide promising approaches for identifying at-risk individuals and enabling targeted prevention. This research develops a longitudinal, multimodal machine learning framework for predicting diabetes onset using Qatar Biobank (QBB) data. The study is structured into three phases: prediction, longitudinal risk analysis, and complication assessment. Multiple data modalities including clinical data, DXA imaging, proteomics, retinal images, and spirometry were analyzed using conventional machine learning, deep learning, and transformer-based models. Both unimodal and multimodal approaches were explored, with interpretability achieved through statistical methods and SHAP-based analysis. The results indicate that diabetes is associated with multi-system physiological changes. DXA-derived measures showed elevated bone mineral density and content in diabetic participants, while respiratory indicators, FVC and FEV1 suggested pulmonary involvement. Multimodal analysis identified visceral adipose tissue, lipid markers (HDL, LDL), and key proteomic biomarkers such as ADIPOQ, PRKACA, and WISP1 as dominant predictors, emphasizing the role of internal fat distribution and metabolic signaling pathways. Distinct patterns in metabolic markers, including cholesterol, triglycerides, and C-peptide, further differentiate healthy and diabetic groups. Demographic analysis showed improved prediction performance with increasing age, with stronger results in older individuals, while males exhibited higher susceptibility to diabetes onset. Transformer-based models outperformed conventional approaches, demonstrating superior capability in modeling complex multimodal relationships. Overall, this work highlights the potential of integrating multimodal biomedical data with interpretable machine learning for early diabetes prediction, enabling timely interventions and improved health outcomes.

Date of Award	2026
Original language	American English
Awarding Institution	HBKU College of Science and Engineering