**5. Arrhythmia Detection**

Computer-aided ECG and PPG analysis has vastly improved the detection and classification of arrhythmias. The most current process for classification of arrhythmias consists of: (1) pre-processing, where baseline wander and unwanted noises and frequencies are filtered out; (2) feature extraction, where the most important features of a wave are identified, and (3) classification, where the most important features are input into a model to predict the class of arrhythmia of a given signal [73]. For filtering ECG signals, the P- and T-waves are typically found between 0.5 Hz and 10 Hz, while the QRS complex is found between 4 Hz and 20 Hz. Discrete wavelet transforms (DWTs) can be used in combination with low- and high-pass filters to remove unwanted frequencies [74,75]. For PPG signals, wavelet decomposition has also been investigated. However, filtering of motion artifacts for PPG signals has only been proven for weak noise, and very noisy PPG signals need to be discarded [76].

Current arrhythmia-detection models are often based on machine learning, as their accuracy is easily increased with large training datasets. These machine learning methods require feature extraction in either the time domain, frequency domain, time-frequency domain, or nonlinear domain. They are often based on physiological incidents. For example, dimensionality-reduction techniques, such as principal component analysis (PCA), have been used in the time domain, and Fourier transforms have been used in the frequency domain [77,78]. Machine learning models for both ECG and PPG signals have included support vector machine (SVM), multilayer perceptron (MLP), and decision tree (DT) models [78–81]. The cutting-edge machine learning area is deep learning. Deep learning methods, such as convolutional neural network (CNN), deep belief network (DBN), recurrent neural network (RNN), long short-term memory (LSTM), and gated recurrent unit (GRU), have been applied to arrhythmia detection [82–86]. Table 2 shows these machine learning methods for arrhythmia detection and classification.

Kaisti et al. were able to perfectly distinguish between 13 sinus-rhythm subjects and seven subjects with atrial fibrillation using a k-means clustering-based approach [87]. The input to the algorithm was time-frequency data derived from a soft, band-based MEMS pressure-sensor array, showing the feasibility of combining flexible devices and current arrhythmia detection algorithms to provide high detection accuracy. Improving on this concept, Dong et al. used an arrhythmia-detection system consisting DWT and SVM algorithms, with a novel acetylene carbon black/PDMS ECG recording patch as the input, which achieved a high online classification accuracy of 98.7% [88].


**Table 2.** Comparison of arrhythmia-detection methodologies using wearable devices.
