Near real-time single-beat myocardial infarction detection from single-lead electrocardiogram using Long Short-Term Memory Neural Network

Abstract

This study proposes a novel Long Short-Term Memory Neural Network (LSTM) architecture for the diagnosis of myocardial infarctions from individual heartbeats of single-lead electrocardiograms (ECGs). The proposed model is trained using an unbiased patient split approach and validated using 10-fold cross-validation over 148 myocardial infarction and 52 Healthy Control patients from the Physikalisch-Technische Bundesanstalt diagnostic ECG Database to generate an inter-patient classifier. We further demonstrate why special care must be taken when generating the training and testing datasets by exploring the effects of various data-split techniques that could mask the occurrence of overfitting and produce misleadingly high testing metrics of the model's performance. A thorough assessment of these results is provided using several standard metrics for different data split methods to show their tendency to overfitting, data leakage, and bias introduced from previously seen heart beats during the training phase. The design achieves near real-time diagnosis of 40 ms while providing an accuracy of 89.56% (with a 95% Confidence Interval (CI) of ±2.79%), recall/sensitivity of 91.88% (±3.13% 95%CI), and a specificity of 80.81% (±9.62% 95%CI). The fast processing makes the model readily deployable on currently existing mobile devices and testing instruments. The achieved performance makes the proposed method a new research direction for attaining real-time and unbiased diagnosis. While, the modular architectural design of the LSTM network structure, which is amenable for the inclusion of other ECG leads, could serve as a platform for early detection of myocardial infarction and for the planning of early treatment(s).