PSO Optimized Hybrid Deep Learning Model for Detection and Localization of Myocardial Infarction
Garima Sahu, Kailash Chandra Ray
Abstract
Myocardial infarction (MI) damages the heart muscles irreversibly and can result in death. Hence, it is essential to diagnose MI early to save lives. However, traditional detection algorithms using an electrocardiogram (ECG) rely mainly on hand-crafted characteristics and demand a high level of subject knowledge. Hence, two hybrid deep learning (DL) models, i.e., convolutional neural network (CNN) with long short-term memory (LSTM) and CNN with the gated recurrent unit (GRU) are proposed for classifying MI. Particle swarm optimization (PSO) is employed to optimize the hyperparameters of the hybrid models for optimal performance. The hybrid models’ simulations are performed using intra- and inter-patient data from the Physikalisch-Technische Bundesanstalt (PTB)-ECG database in Python for all standard 12 leads individually for the detection and localization of MI. Among MI, normal, and non-MI, the MI is detected by employing the optimized CNN-LSTM on intra-patient data from lead-V2 that provides an accuracy of 99.96%. In contrast, the optimized CNN-GRU method offers an accuracy of 90.15% on inter-patient data from lead-V5 for MI detection. For MI localization, the optimized CNN-LSTM model on intra-patient lead-V1 data provides an accuracy of 99.87% to classify 11 classes. On the other hand, CNN-GRU provides an accuracy of 66.32% to classify five types of MI on inter-patient lead-V1 data. This study and analysis conclude that this article’s two proposed hybrid models outperform state-of-the-art methods on chest leads.