Convolutional Neural Networks for patient-specific ECG classification

Abstract

We propose a fast and accurate patient-specific electrocardiogram (ECG) classification and monitoring system using an adaptive implementation of 1D Convolutional Neural Networks (CNNs) that can fuse feature extraction and classification into a unified learner. In this way, a dedicated CNN will be trained for each patient by using relatively small common and patient-specific training data and thus it can also be used to classify long ECG records such as Holter registers in a fast and accurate manner. Alternatively, such a solution can conveniently be used for real-time ECG monitoring and early alert system on a light-weight wearable device. The experimental results demonstrate that the proposed system achieves a superior classification performance for the detection of ventricular ectopic beats (VEB) and supraventricular ectopic beats (SVEB).

1-D Convolutional Neural Networks for Signal Processing Applications

“…This is why they are often referred to as, "2D CNNs". As an alternative, a modified version of 2D CNNs called 1D Convolutional Neural Networks (1D CNNs) has recently been developed [47][48][49][50][51][52][53][54][55][56]. These studies have shown that for certain applications 1D CNNs are advantageous and thus preferable to their 2D counterparts in dealing with 1D signals due to the following reasons:…”

“…Overview of the arrhythmia detection and identification system proposed in [47,48]. Figure 10: The creation of the training dataset as proposed in [49] for an arbitrary user (Person-X) using the real Nbeats.…”

“…To incorporate such drawbacks, in 2015 Kiranyaz et al [47] proposed the first compact and adaptive 1D CNNs to operate directly on patient-specific ECG data. In a relatively short time, 1D CNNs have become popular with a stateof-the-art performance in various signal processing applications such as early arrhythmia detection in electrocardiogram (ECG) beats [47][48][49], structural health monitoring and structural damage detection [50][51][52][53][54], high power engine fault monitoring [55] and real-time monitoring of high-power circuitry [56]. Furthermore, two recent studies have utilized 1D CNNs for damage detection in bearings [57][58][59][60].…”

Fault Detection and Severity Identification of Ball Bearings by Online Condition Monitoring

“…The conventional 2D CNNs are a type of deep, biologically inspired feed-forward ANNs which are based on a core model for mammalian visual cortex. The proposed 1D CNNs [39] are their counterparts that work on 1D signals. Similar to 2D CNNs, two types of layers exist in the 1D CNNs: 1) the so-called "CNNlayers" where both 1D convolutions and sub-sampling occur, and 2) Fully-connected layers that are identical to the layers of a typical Multi-layer Perceptron (MLP) and hence called "MLP-layers".…”

“…Since the learning stage is tuned for bearing damage detection in this study, a 1D CNN structure can now blend the extraction of features and "damage learning" stages of the raw accelerometer data from wireless sensors. As such, the CNN topology will permit the variations in the input layer dimension since the sub-sampling factor of the output CNN layer is set adaptively [39]. Details regarding forward and back-propagation in CNN layers are covered in Appendix A.…”