Multi-task learning for multi-label electrocardiogram disease detection via point-level delineation-guided fusion

Despite substantial research over the past decade on analyzing electrocardiograms (ECGs) using deep learning models, disease detection in complex scenarios remains challenging, particularly for co-occurring diseases. Additionally, existing methods lack sufficient interpretability for multi-label disease classification. To address these issues, we propose the Point-level Delineation Guided Multi-Task Learning network (PDG-MTL), integrating the main task of ECG classification with an auxiliary task of ECG delineation. Firstly, during training, fiducial points are generated using the ECGDeli tool and converted into labels for the auxiliary task, including three segments and three waves of ECG delineation. Secondly, we proposed a novel decoder-focused multi-task architecture comprising initial task predictions and an attention-guided fusion module. The initial prediction heads provide deep supervision, obtaining multi-modal predictions that include both regions of interest for ECG delineation and meaningful diagnostic features. The attention-guided module fuses the feature maps of waveforms and delineation information, which bridge the dimensional gap between heterogeneous features. In the results, PDG-MTL outperforms prior models across a diverse patient cohort on a large-scale dataset, particularly in multi-label classification involving over 20 classes. The average Area under the Receiver Operating Characteristic (AUROC) increases by up to 3.4 % for morphological cardiac diseases. It also significantly improves performance for low-sample diseases, with F1 scores increasing by nearly 60 %. The model demonstrates exceptional robustness, with minimal confidence intervals across six experiments. The attention map reveals alignment between the model's focus on delineation regions and expert expectations across disease groups, highlighting the potential of interpretability.