A deformable lead-attention fusion network for multi-label ECG classification integrating clinical metadata.

Cardiovascular diseases remain a major global health burden, making accurate electrocardiogram (ECG) analysis essential for timely diagnosis. While deep learning-based methods have made significant progress in automating ECG classification, they are often limitedbythe variability of ECG signals, the frequent presence of multiple cardiac abnormalities, and the inadequate integration of diverse clinical data. This study proposes Deformable Lead-Path and Multi-Modal Fusion Network (DLM-Net): a robust deep learning framework for multi-label ECG classification, designed to address these challenges. DLM-Net incorporates several key innovations. First, the Lead-Specific Deformable Path applies deformable convolutions independently to each lead. This enables adaptive modeling of non-rigid and complex waveform variations. It is also enhanced by a high-frequency module, which captures fine-grained temporal details. Secondly, a spatial path based on a ResNet-like backbone and feature pyramid network fuses multi-scale representations in order to effectively capture both global and local ECG patterns. In parallel, the lead path uses a lead attention Transformer to explicitly model inter-lead dependencies. Finally, ECG features from the spatial and lead paths are integrated with clinical tabular data through a multimodal fusion module, in which compact patient-level tabular embeddings provide complementary contextual information for ECG representation. Extensive experiments on three public multi-label ECG datasets- Physikalisch-Technische Bundesanstalt-XL, CPSC2018 and Chapman-demonstrate that DLM-Net achieves competitive and stable performance across datasets, with label-wise accuracies of 89.11%, 96.05% and 97.82%, respectively. Overall, DLM-Net provides an effective and flexible framework for multi-label ECG classification, offering improved accuracy and robustness across diverse datasets and demonstrating the effectiveness of the proposed architecture in handling complex multi-label diagnostic tasks in retrospective settings.