Novel Transformation Deep Learning Model for Electrocardiogram Classification and Arrhythmia Detection using Edge Computing

Abstract

The diagnosis of the cardiovascular disease relies heavily on the automated classification of electrocardiograms (ECG) for arrhythmia monitoring, which is often performed using machine learning (ML) algorithms. However, current ML algorithms are typically deployed using cloud-based inferences, which may not meet the reliability and security requirements for ECG monitoring. A newer solution, edge inference, has been developed to address speed, security, connection, and reliability issues. This paper presents an edge-based algorithm that combines continuous wavelet transform (CWT), and short-time Fourier transform (STFT), in a hybrid convolutional neural network (CNN) and Long Short-Term Memory (LSTM) model techniques for real-time ECG classification and arrhythmia detection. The algorithm incorporates an STFT CWT-based 1D convolutional (Conv1D) layer as a Finite Impulse Response (FIR) filter to generate the spectrogram of the input ECG signal. The output feature maps from the Conv1D layer are then reshaped into a 2D heart map image and fed into a hybrid convolutional neural network (2D-CNN) and Long Short-Term Memory (LSTM) classification model. The MIT-BIH arrhythmia database is used to train and evaluate the model. Using a cloud platform, four model versions are learned, considered, and optimized for edge computing on a Raspberry Pi device. Techniques such as weight quantization and pruning enhance the algorithms created for edge inference. The proposed classifiers can operate with a total target size of 90 KB, an overall inference time of 9 ms, and higher memory use of 12 MB while achieving up to 99.6% classification accuracy and a 99.88% F1-score at the edge. Thanks to its results, the suggested classifier is highly versatile and can be used for arrhythmia monitoring on various edge devices.