Classification of Prehospital-Electrocardiograms taken in Ambulance According to Severity using Deep Learning Neural Network

Abstract

The prehospital-electrocardiogram (PH-ECG) is an electrocardiogram (ECG) measurement performed by paramedics on a patient suspected of having a myocardial infarction. For example, in an ambulance and the data are transmitted to the hospital. A physician at the hospital can diagnose the condition of a patient based on the transmitted ECG, thus making efficient use of the time before the patient arrives and enabling an early start to treatment. The PH-ECG is particularly useful for patients who require immediate medical attention, such as those with ST-elevation myocardial infarction (STEMI). Multiple studies have shown that PH-ECG improves door-to-balloon time and in-hospital mortality. However, it is necessary to understand the various patterns of abnormal waveforms when analyzing PH-ECG, and it is difficult to make an accurate diagnosis quickly without a cardiologist. In areas where there is a shortage of hospitals and physicians, diagnosis is performed by non-cardiologists, and there is a need for an automated diagnosis system with performance similar to that of cardiologists. Recent studies on automated ECG diagnosis have focused on diagnosing specific abnormal findings, especially the classification and discrimination of myocardial infarction and arrhythmias. On the other hand, there are few studies on the classification of disease severity, independent of the types of abnormal findings. In this work, we analyzed a 12-lead ECG measured in an ambulance by using deep learning neural network to classify and evaluate the abnormal waveforms according to degrees of severity. For 88 cases of 12-lead ECG image data measured in the ambulance, each 12-lead waveform was divided into three parts, and 36 one-lead ECGs were extracted. An expert cardiologist annotated each image. The images were labeled in three classes according to the degree of severity, “normal,” “mild or moderate,” and “severe.” Each image was thinned and binarized. Of 3,168 final images, 1,590 were normal waveforms, and 1,578 were abnormal waveforms. 80% of the images were used as training data and 20% of the images were used as test data. A total of 20% of the training data were used as validation data, five-fold cross validation was performed. EfficientnetB0 was used. The model was defined by using the network designer in MATLAB. The input image size was set to 224 × 224 pixels, and resizing was performed when no match was found. The optimization method was Adam, and the hyperparameters were set to <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\alpha=0.0001, \beta_{-}1=0.9$</tex> , and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\beta_{-}2= 0.999$</tex> . We set the mini-batch size to 64 and the epoch to 100. We achieved a kappa coefficient of 0.810 and maximum classification accuracy of 86.6% for the test data. The result indicates the feasibility of an automatic diagnosis system using noisy ECGs measured in ambulances and is expected to provide a new research direction.