Electrocardiogram signal classification for diagnosing heart abnormalities using a deep learning approach with lead-wise expert prior knowledge framework

Electrocardiogram (ECG) classification using deep learning (DL) techniques in heart abnormalities classification based on ECG signals has emerged as a promising approach for accurate and efficient diagnosis in cardiovascular medicine. This method enables early detection and intervention strategies at a population level, potentially reducing the prevalence and burden of heart diseases. However, the investigation of effectiveness of DL on multi-label ECG classification is limited. This study aimed to design a lead-wise prior knowledge framework (LPFK) integrated in DL based framework for multi-label heart abnormalities classification based on ECG signals and validate this proposed DL framework, using ECG data from PTB-XL dataset. This study involved six stages: preprocessing ECG signals, developing lead-wise prior knowledge attentions (LPKAs), designing a lead-wise prior knowledge module (LPKM), a lead-wise prior knowledge framework (LPKF) that utilizing five backbone DL models (Xresnet1d101, InceptionTime, ResNet, LSTM and LSTM-bidir) for feature extraction and predication, and assessing the enhanced model's effectiveness. The results indicated that the LPKF-enhanced InceptionTime model demonstrated enhancements of 0.017, 0.018 and 0.016 in Macro-AUC for the classification of superclass diagnostic, subclass diagnostic and diagnostic categories respectively. Furthermore, the LPKF-enhanced InceptionTime model exhibited an average improvement of 0.07 in macro-F1 score compared to the existing state-of-the-art (SOTA) methods. Based on ablation studies and model interpretability test, the proposed components within the LPKF were justified and the LPKF-enhanced InceptionTime model’s ability in capturing important diagnostic symptoms from ECG signals has been demonstrated. In conclusion, the LPKF-enhanced InceptionTime model holds significant potential for further enhancing the accuracy and robustness of multi-label ECG classification, which can provide secondary prevention in epidemiology and health policy decisions aimed at reducing the incidence of heart diseases.

Text FPSK (p) 2024 16 - Declaration Form.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (709kB) Text FPSK (p) 2024 16 - Full Text.pdf Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (10MB) Text FPSK (p) 2024 16.pdf Restricted to Repository staff only Available under License Creative Commons Attribution Non-commercial No Derivatives. Download (10MB)

Actions (login required)

View Item