For Reviewers

Cardiovascular Disease (CVD) is the leading cause of human death and was responsible for 31% of deaths worldwide in 2016 [1]. An electrocardiogram (ECG) is widely used in the medical field to diagnose heart diseases. Usually, ECG is obtained by electrodes placed on the skin of the patient, which record the electrical changes during the cardiac cycles, from cardiac muscle depolarization to repolarization. A typical ECG signal, as illustrated in Fig. 1, consists of a P wave, QRS complex, T wave, and U wave; the size and shape of this signal contains useful information about the nature of the disease or abnormality afflicting the heart.

Fig. 1. 
The typical ECG waveform.

In most cases, the existing examination methods are inefficient owing to a considerable amount of heterogeneous data, which must be rigorously analyzed to obtain high accuracy in diagnosis. Deep learning (DL) refers to the study of knowledge extraction, predictions, intelligent decision making, or recognizing patterns with the help of a set of training data. Compared with the conventional learning techniques, deep neural networks (DNNs) are more scalable because higher accuracy is achieved by increasing the size of the network or the training dataset. In particular, DNNs are extensively used for classification purposes in different domains. The majority of DL based algorithms have developed for image (2-D) classification.

In [2], a convolutional neural network (CNN) was proposed, which is the most popular DNN architecture usually trained with the gradient-based optimization algorithm. In general, a CNN consists of multiple back-to-back layers connected in a feed-forward manner. The main layers include the convolutional, normalization, pooling, and fully-connected layers. The layers at the beginning are responsible for extracting features, whereas the fully-connected layers at the end are in charge of classification. A CNN architecture has exhibited 98.1% accuracy for classifying five arrhythmias [2]. In [3], the authors proposed an ECG classification system based on long short term memory and a CNN; their classifier achieved 88.33% accuracy. In [4], recurrent neural network based arrhythmia classification was conducted, and an accuracy of 88.1% was achieved. Also, Ref. [5,6,7] showed the recent works based on Deep CNN and RNN for accurate arrhythmia detection. These algorithms automatically extract the features from the given data and hence showed high accuracy. These works demonstrated high accuracy; however, the DL architecture was deep, which resulted in a high computational cost. This poses a challenge for low power devices. In order to reduce the computational cost in a DL architecture, spiking neural networks (SNNs) were introduced. In this type of network, the information is processed based on the propagation of spikes through the network as well as the timing of the spikes. It provides the opportunity for low-power operation [8,9].

In Ref. [10], a wearable device for ECG classification system was developed, which achieved an accuracy of 97.9%. The paper employed spike-timing-dependent plasticity (STDP), in which the model weights are trained according to spike signal timing and reward or punishment signals. In Ref. [11], an energy-efficient ECG classification method was designed with an SNN, and 77.49% accuracy was achieved. The conventional SNN has low accuracy or high computational cost. Therefore, this paper proposes a new binarized SNN which modifies the synaptic weights of SNN constraining it to be binary (+1 and -1), compares the traditional DL models, the conventional SNN, and the proposed Binarized SNN in terms of accuracy and power.

Binarized SNN converts the high-precision weights into binary weights, i.e., -1 and 1. The mathematical formulation is given by

where x is the real value of the weight and x^b is the binarized output weight.

Usually, in full-precision networks, a large number of calculations is spent on calculating dot products of matrices as needed by fully connected and convolutional layers. The method was motivated by the binarized neural network in which all the network parameters are constrained to be binary. This conversion technique can reduce the overall training time and cost of the SNN architecture.

This section discusses the parameters of different DL architectures and evaluates them based on resource utilization, accuracy, and power.

The training dataset in the case of CNN are 15,000 samples per Normal/Abnormal ECG. The network parameters shown in Table 1 are used from Ref. [8] for ECG classification. The CNN model is tested only on the software layer because of more number of layers, which will consume more resources as compared to other networks.

The probabilistic encoding method converts the pixel value of the input image to spike trains and feeds it into the SNN. In addition, the same encoding method is adopted in conventional and binarized SNN in Table 2, 3 respectively for ECG classification.

The number of training samples for both the SNN architectures is 2000, where 1000 samples are of Normal ECG and 1000 are of abnormal ECG taken from Physionet’s PTB Diagnostic ECG Database [20].

All simulations were performed on PyCharm IDE. Then the accuracy is obtained and deployed the algorithms on PYNQ-Z2 FPGA synthesize the resource utilization by the respective algorithms. Table 4 compares various evaluation factors of the algorithms.

As shown in Table 4, the CNN utilized the highest amount of memory during the synthesis process; however, the power for the CNN was not measured. In addition, the CNN gained a higher accuracy; however, it is inferior in resource utilization. In contrast, from a hardware perspective, the SNN performed well. The performance is tested on static 2D ECG images. Because of the term binarization, the SNN network has to compromise with the accuracy and only an improvement of 3% was achieved. As a whole, from Table 4, the proposed binarized SNN proves to be the optimal architecture for normal and abnormal ECG classification.

This study compared DL models to the energy-efficient SNN model in terms of accuracy, power, and memory. Among the architectures described here, binarized SNN consumed the least power (approximately 0.176 W). The CNN was more accurate than the other two models. However, we conclude that to constrain the power of embedded devices, the accuracy needs to be compromised. The algorithm successfully classified ECG images as Normal and Abnormal with a 2× performance improvement in a binarized SNN.

Future studies will aim to extend the number of arrhythmias to be classified and make the network hardware friendly. There is a growing belief that an SNN can potentially overcome the limitations of DNNs. Moreover, it can potentially enable low-power hardware evaluation.