An Explainable Federated Deep Learning Framework for HRV-Based Early Detection of Diabetic Autonomic Dysfunction: Experimental Evaluation Using ROC–AUC Optimization and Computational Efficiency Analysis

Abstract

Diabetic autonomic dysfunction is a progressive and frequently underdiagnosed complication of diabetes mellitus that significantly increases cardiovascular morbidity and mortality. Early identification through non-invasive physiological biomarkers such as heart rate variability (HRV) provides a critical opportunity for preventive intervention. This study proposes a structured federated analytical framework for HRV-based early detection of diabetic autonomic dysfunction, emphasizing predictive discrimination and computational stability. Experimental findings demonstrate progressive improvement in classification performance across training epochs, reflected by increasing accuracy and area under the curve (AUC) metrics. Receiver operating characteristic (ROC) analysis confirms reliable discriminative capability between affected and non-affected individuals. Loss trends indicate stable optimization without significant overfitting, while time complexity analysis shows consistent computational efficiency per epoch. The framework further enables interpretation of HRV biomarker contributions, supporting clinical transparency and translational relevance. Collectively, the results validate the feasibility of HRV-driven computational modelling as a scalable and reliable approach for early cardiovascular risk stratification in diabetic populations. The proposed system demonstrates potential integration into modern digital health infrastructures to enhance preventive cardiometabolic care.