Light Weight Residual Convolutional Neural Network for Atrial Fibrillation Detection in Single-lead ECG Recordings

Abstract

Electrocardiogram (ECG) analysis constitutes the most important approach able to classify heart infarction anomalies. These anomalies can be identified from the changes in various features of the ECG signal. In this paper, we are proposing a new features-based classification method of short-time single-lead ECG signals. The goal of this method is to classify these ECG signal into one of the following classes: normal, atrial fibrillation, other abnormalities, and too noisy as defined by the dataset. This is a challenging problem because of the severe imbalance between the classes, where the normal class makes up the majority of the samples in the dataset. The second challenge in this dataset is the fact that the sample ECG signals have a variable length (it varies between 3 to 60 seconds). The proposed method considers three main processes. The first process consists of detecting inverted ECG record by analyzing the signal range and mean in a sliding-window. The second process involves the extraction of many features effective in characterizing ECG signals and detecting abnormalities. These features include morphological, Heart Rate Variability, statistical, time/frequency amplitudes, and special Atrial Fibrillation (AF) features. The third process represents the main contribution by designing a lightweight residual Convolutional Neural Network (CNN) model for the classification of short-time single-lead ECG signals. This model is composed of five layers with two residual connections where advanced CNN concepts such as Batch Normalization, DropOut, and Leaky-ReLU are used. Compared to state-of-the-art solutions, the proposed method achieved the best performance with F1-score of 95.11% using inversion correction.