Improved state refinement for LSTM determined 3D CAISR-LSTM model for automatic myocardial infarction detection

Objective. Electrocardiograms (ECGs) contain valuable information in the clinical diagnosis of myocardial infarction (MI). However, its interpretation process is dependent on cardiologists with extensive clinical experience and expertise. The issue not only causes a paucity of medical resources, but also restricts patients from receiving timely diagnoses. Thus, a novel approach for MI automatic detection is developed, based on 12-lead ECG and an improved state refinement for long short-term memory (LSTM) determined 3D convolution-attention (3D CAISR-LSTM) model. Approach. The proposed 3D CAISR-LSTM model is trained in an end-to-end fashion. The input 12-lead ECG signals are preprocessed to eliminate power line interference, high-frequency noise and baseline wander. Then, the ECG signals are transformed into time-frequency images using continuous wavelet transform (CWT) and bilinear interpolation. The obtained images are constructed into three-dimensional spatiotemporal features, serving as input to the 3D CAISR-LSTM model. In the 3D CAISR-LSTM model, there are three main components: a convolutional module, four identical convolutional attention modules, and an improved state refinement for LSTM. Performance of the 3D CAISR-LSTM model in automatic detection of MI versus healthy controls (HC) is evaluated through ten-fold cross validation on the publicly available PTB diagnostic ECG database. Main results. Experimental results demonstrate that the 3D CAISR-LSTM model achieves an accuracy of 98.45%, sensitivity of 98.69%, specificity of 97.50%, and F1 score of 99.03%, outperforming various advanced 2D and 3D deep neural network architectures. Significance. The proposed approach is expected to provide an early warning before obvious MI symptoms appear. It also has the potential to be developed into a lightweight embedded MI detection equipment.