Lead Recovery Guide Residual Network for Myocardial Infarction Detection by Restoring 12-Lead Spatial Information

To address the conflict between the need to use 12-lead in detecting myocardial infarction (MI) and inadequate diagnostic data due to an insufficient number of leads, this study proposes a novel network called Lead Recovery Guide Residual Network (LRGRN), which mitigates the effect of the restricted number of leads. We constructed a lead recovery guide to restore the spatial information of all 12-lead, given only leads I, II, and V2. Limb leads were reconstructed through a linear model, while precordial leads were reconstructed using a convolutional bidirectional long short-term memory network to capture high-level abstract features and temporal characteristics of electrocardiogram (ECG) signals. The restored 12-lead ECG can overcome the limitations of the original 3-lead ECG and provide a comprehensive reflection of MI. In the overall architecture of LRGRN, patient data strictly follow the inter-patient principle. ResNet maintains a stable flow of frequency information based on the reconstructed 12-lead ECG data, while the multi-lead single-channel structure enables the model to better capture the overall ECG information. The average detection accuracy of the LRGRN model for MI is 96.33%. The correlation coefficient (CC) of the recovered limb leads was 100%, The CC for recovery of precordial leads for feedback was 95.62%. Compared with the models presented in other studies, the LRGRN model overcomes inter-individual variability and excels in MI detection with limited lead information.