Diabetic Cardiomyopathy Staging using a Tri-branch Deep Learning Framework with Optimized Adaptive Loss Mechanism

Abstract

Diabetic cardiomyopathy (DCM) is a progressive cardiac disorder characterized by stage-dependent structural, functional, and metabolic alterations. Most existing approaches treat DCM as a single risk prediction problem, ignoring the ordered and asymmetric nature of disease progression and resulting in poor discrimination of intermediate stages. To address these limitations, this study aims to propose a stage-aware tri-branch deep learning framework for DCM stage classification. A phenotype-informed tabular dataset is first constructed using a Conditional Tabular Generative Adversarial Network (CTGAN) to mitigate the scarcity of labeled stage-specific clinical data. Following preprocessing, the input data is transformed through a shared Neutral Attentive Feature Encoder (NAFE) to produce an unbiased latent representation. This representation is processed by tri-branch Bi-directional Long Short-Term Memory (BiLSTM) branches, each specialized in learning the characteristic temporal patterns of a specific disease stage. Final stage prediction is achieved using an ArgMax-based Winner Takes All (WTA) mechanism. The model is trained through a novel stage-adaptive competitive loss and a Stage-Aware Adam optimizer (SA-Adam). Experimental results show that the proposed framework performed better than the existing works with an accuracy of 99.2%, precision of 98.61%, and recall of 97.89%. This ensures the proposed framework enables DCM stage classification, improved interpretability, and supports stage-specific treatment planning.