Edge-BioFormer: Decentralized Self-Supervised Transformer Framework for Early Diabetes Detection

Abstract

Early detection of Type 2 Diabetes (T2D) is essential for preventing disease progression and reducing long-term health complications. However, conventional screening methods rely on periodic clinical examinations and centralized data processing, limiting their ability to provide continuous and privacy-preserving monitoring. This paper proposes Edge-BioFormer, a decentralized self-supervised transformer framework for early diabetes detection using wearable sensor data. The proposed framework applies Metabolic Pattern Modeling (MPM) to learn normal metabolic patterns from unlabeled physiological signals and detect abnormal changes associated with diabetes risk. It integrates key biomarkers, including the Advanced Glycation End Products (AGEs) Index, Body Mass Index (BMI), Age, Resting Heart Rate (RHR), Heart Rate Variability (HRV), and daily activity data. Federated learning enables collaborative model training while preserving user privacy by keeping personal health data on local wearable devices. In addition, Layer-wise Relevance Propagation (LRP) provides interpretable predictions by identifying the physiological features that contribute to each risk assessment. Experimental results demonstrate that the proposed stacking ensemble model achieves the best performance with an ROC-AUC of 0.86, outperforming individual ensemble models while providing high sensitivity for early diabetes detection. The findings also identify the AGEs Index, BMI, Age, and RHR as the most influential biomarkers, whereas higher HRV and better sleep quality act as protective factors. These results indicate that Edge-BioFormer provides an effective, privacy-preserving, and clinically interpretable framework for continuous diabetes monitoring. Future work will focus on improving signal denoising, reducing communication overhead in federated learning, and validating the framework using larger and more diverse longitudinal datasets.Early detection of Type 2 Diabetes (T2D) is essential for preventing disease progression and reducing long-term health complications. However, conventional screening methods rely on periodic clinical examinations and centralized data processing, limiting their ability to provide continuous and privacy-preserving monitoring. This paper proposes Edge-BioFormer, a decentralized self-supervised transformer framework for early diabetes detection using wearable sensor data. The proposed framework applies Metabolic Pattern Modeling (MPM) to learn normal metabolic patterns from unlabeled physiological signals and detect abnormal changes associated with diabetes risk. It integrates key biomarkers, including the Advanced Glycation End Products (AGEs) Index, Body Mass Index (BMI), Age, Resting Heart Rate (RHR), Heart Rate Variability (HRV), and daily activity data. Federated learning enables collaborative model training while preserving user privacy by keeping personal health data on local wearable devices. In addition, Layer-wise Relevance Propagation (LRP) provides interpretable predictions by identifying the physiological features that contribute to each risk assessment. Experimental results demonstrate that the proposed stacking ensemble model achieves the best performance with an ROC-AUC of 0.86, outperforming individual ensemble models while providing high sensitivity for early diabetes detection. The findings also identify the AGEs Index, BMI, Age, and RHR as the most influential biomarkers, whereas higher HRV and better sleep quality act as protective factors. These results indicate that Edge-BioFormer provides an effective, privacy-preserving, and clinically interpretable framework for continuous diabetes monitoring. Future work will focus on improving signal denoising, reducing communication overhead in federated learning, and validating the framework using larger and more diverse longitudinal datasets.