Mamba—GCN: Fusion Framework for Interpretable Arrhythmia Diagnosis Using 12-Lead ECG Signals

Abstract

The development of automated electrocardiogram (ECG) analysis systems that are both accurate and clinically interpretable is driven by the fact that cardiovascular diseases continue to be a major cause of death worldwide. Although many deep learning techniques function as black-box models and provide little insight into the spatial reasoning across ECG leads that supports clinical decision-making, recent methods have demonstrated strong diagnostic performance on large-scale ECG datasets. Moreover, Transformer-based architectures introduce significant computational overhead when modelling long-duration biosignals. With an emphasis on graph-based learning and Selective State Space Models (SSMs), this paper provides an organised overview and methodological synthesis of recent developments in ECG modelling. In addition to reviewing their extensions that include convolutional front-ends, spectro-temporal embeddings, bidirec-tional processing, and multi-branch tokenisation techniques, we examine the rise of Mamba-based architectures as effective substitutes for long-range temporal modelling. Simultaneously, we investigate graph neural network formulations that explicitly model inter-lead relationships, emphasising the growing interest in dynamic graph learning and the drawbacks of static adjacency assumptions. We describe a Mamba-first, graph-second pipeline that sepa-rates temporal feature extraction from spatial reasoning based on these discoveries. In this framework, Mamba-based encoders independently model lead-wise temporal dynamics, after which a dynamically learned graph captures pathology-dependent inter-lead interactions, yielding an interpretable reasoning matrix. We demonstrate that hybrid spatio-temporal architectures can achieve diagnostic performance comparable to strong convo-lutional baselines while offering significantly improved trans-parency through a quantitative comparison of representative models and visual analysis of performance—interpretability trade-offs. All things considered, this work summarises recent develop-ments in effective temporal modelling and comprehensible spatial reasoning for ECG analysis and suggests promising avenues for further investigation toward reliable, clinically useful ECG decision-support systems.