Contents I. Introduction
Electrocardiogram (ECG) is a record of the electrical activity of the heart captured using different techniques. These waveforms are used in diagnosing different cardiopathies. Moreover, several applications have been proposed and implemented where ECG signals were used in emotion recognition and biometric identification. As it is impractical for a cardiologist to interpret very long sequence of data obtained over a long period of time, many systems were developed to automate the process of translating the waveforms to extract useful information. Several ECG signal acquisition techniques such as the standard 12 electrode ECG, Smart Chair using capacitively coupled electrodes [1], wearable devices, and RF based ones were used to capture the electrical activity of the heart. To track and monitor abnormal heart rhythms and cardiac symptoms over a longer period of time many wearable devices were developed. However, as the ECG signal captured from wrist tend to be weak, we need a powerful processing blocks to eliminate or suppress the noise originating from artifacts. For wearable devices, power efficient processors are also required to extract features that can be used for classification stage. In the process of ECG signal acquisition, different types of noise and artifact contaminate captured ECG signals. Various techniques are applied to reduce the effect of noise in the pre-processing stage of the system. Common types of noise are power line interference (a signal in the frequency of 50 or 60 Hz), baseline wonder noise (low-frequency 0.15 up to 0.3 Hz), electrode contact noise, electrode motion artifacts, muscle contractions, electrosurgical noise, and instrumentation noise. Noise and artifact suppression is performed with a series of sophisticated filtering blocks [2]. Once the signal’s SNR is improved with the pre-processing stage a feature extraction technique is applied to obtain the distinctive ECG features such as the P-QRS-T complex features, statistical features, morphological features, and wavelet features. Among the most powerful algorithms for ECG QRS complex detection are Pan-Tompkins algorithm [3] [4] derivative [3] [5], digital filters [3], wavelet transform [5], Hilbert-Huang transform [7], neural networks [12]. The features extracted are then used in the classification stage of the system. The major categories of these classifiers are artificial neural networks (ANNs), LDA, k nearest neighbor (kNN), support vector machine (SVM), decision tree (DT), and Bayesian classifiers. To eliminate the feature extraction stage, a new End-to-End ECG Classification with Raw Signal Extraction and Deep Neural Networks (DNN) was proposed in [8], where DNN is used for both feature extraction and classification based on aligned heartbeats. That is raw ECG signal is fed to the proposed system and beat by beat classification decision is obtained. Though (ANN) is not a new concept, it has only started recently to drive the rapid development of many applications with emerging advancements in Graphical Processing Unit (GPU), availability of massive data and advanced chip manufacturing technologies. However, ANNs highly demanding computational complexity and large area/power requirement of its hardware implementation has made it impractical for low power real time applications [9]. Moreover, the extremely slow training phase is another major challenge with many applications of ANN especially in edge devices and mobile systems. Application of deep learning neural network for ECG feature extraction started growing recently in terms of accuracy and complexity with the increased availability of neural network hardware implementations [10] [11] . Two-level convolutional neural network (CNN) based QRS complex detection was proposed in [10]. ECG signal variations, avoids the need for hand-crafted features, improves accuracy in detecting features, and reduces computational cost. With the advancement of deep learning neural networks, subject independent ECG feature extraction and classification can be achieved with accuracy equivalent to state-of-the-art-techniques [12].
