topic
Unsupervised ECG Analysis—A Review
Summary
An electrocardiogram (ECG) is the gold standard technique for detecting abnormal heart conditions. Automatic detection of ECG abnormalities helps clinicians analyze the vast amounts of data that heart monitors generate every day. Due to the limited number of cardiologist-labeled abnormal ECG samples for training supervised machine learning models, there is an increasing need for unsupervised learning methods for ECG analysis. Unsupervised learning aims to classify ECG samples into different abnormal categories without the need for labels provided by cardiologists, a process known as ECG clustering. In addition to anomaly detection, ECG clustering has recently uncovered valuable information reflecting inter-intra-individual patterns about the whole body and mind, such as emotions, psychological disturbances, and metabolic levels. ECG clustering can also address specific challenges faced by supervised learning systems, such as the imbalanced data problem, and can enhance biometric systems. Although there have been several review articles on supervised ECG systems, a comprehensive review on unsupervised ECG analysis techniques is still lacking. This study mainly reviews the major developments in ECG clustering techniques over the past decade. Emphasis is placed on recent machine learning and deep learning algorithms and their practical applications. We critically review and compare these techniques, discuss their applications and limitations, and provide directions for future research. This review provides in-depth insights into ECG clustering and provides the necessary information needed to adopt an appropriate algorithm for a specific application.
introduction
An electrocardiogram (ECG) shows the electrical activity of the heart. It is routinely recorded in intensive care units as well as in routine monitoring and wearable monitors, generating large amounts of data on a daily basis. A number of systems based on supervised machine learning have been developed to classify heartbeats into normal and several abnormal categories using ECG datasets and labels provided by cardiologists [1]–[5]. However, cardiologists can only analyze and label a small subset of the vast amount of ECG data to indicate common cardiac abnormalities. Furthermore, most labeled ECG datasets are obtained in controlled settings, such as hospitals and clinics, and contain very limited samples compared to the diverse ECG patterns that may arise in different physiological and pathological conditions. For example, the patterns of ECG data acquired from individuals under stress or with diabetes have been shown to differ from those acquired under normal conditions [6], [7]. Therefore, several unsupervised learning methods have recently been proposed for analyzing ECG data without labels provided by cardiologists, a process known as ECG clustering.
However, the need for unsupervised ECG analysis does not stem solely from a shortage of labels provided by cardiologists. There are actually inter-patient and intra-patient ECG patterns and structures that, if discovered, can further reveal valuable information about the cardiovascular system as well as the body and mind as a whole. Discovering these relationships can reveal complex mechanisms and important biomarkers of various health conditions, as well as states of mind and body, and ultimately guide physicians in making fine-grained treatment decisions. Notably, visual identification of these patterns was not possible due to their complexity and large amount of data. However, these patterns can be automatically identified through clustering techniques. For example, ECG clustering has helped researchers, especially in the field of psychophysiology, discover hidden ECG patterns associated with different emotional states (such as sadness and emotional stress), brain disorders (such as epilepsy), and states of lethargy [6] , [8], [9]. ECG clustering also allows researchers to identify significant cardiac abnormalities and metabolic differences in patients with different health conditions, including diabetes [7], nocturnal hypoglycemia [10], embolic stroke [11] and Atherosclerosis [12].
In addition to being applied as part of broader knowledge discovery systems, clustering techniques, especially deep learning-based unsupervised methods such as autoencoders [13], [14] and generative adversarial networks [15], have also been used to To overcome some of the challenges faced by ECG supervised learning systems, such as addressing the imbalanced data problem [16] and low-level automation of individualized ECG classifiers [17]–[19]. Furthermore, ECG clustering has been applied in biometric authentication [20]–[23], ECG segmentation [24], [25] and extraction of fetal ECG from abdominal ECG [26].
To date, several studies have reviewed supervised learning techniques for ECG analysis [2]–[4], [27]–[30]. However, to the authors' knowledge, this work is the first comprehensive and critical review of an unsupervised ECG analysis system. In this work, we review unsupervised ECG analysis systems in clinical/medical applications and the associated machine learning methods employed by these systems - from traditional models to recent deep learning models. For a comprehensive review, we searched multiple platforms, including IEEEEXplore, ScienceDirect, Google Scholar, Scopus, and PubMed databases, and selected most of the studies published within the past decade in prestigious journals and conferences ranked by well-known citation indices . We discuss these recent studies, make comparisons, outline their limitations, and provide future directions. This will allow researchers to easily obtain the information they need and choose the appropriate algorithm for their specific application.
Electrocardiogram (ECG) clustering
A typical electrocardiogram (ECG) clustering pipeline includes several data preparation and preprocessing steps before applying a clustering algorithm. In this section, we briefly review data preparation techniques for efficient clustering of ECG data, including denoising, segmentation, and feature engineering. We then extensively review traditional and state-of-the-art clustering algorithms, including deep learning approaches, and make critical comparisons.
Data Preparation for ECG Clustering
The methods used for denoising, segmentation, and feature engineering largely overlap with those used in supervised learning systems. Here, we briefly introduce these techniques, especially those tailored for ECG clustering. Interested readers can refer to [2] to [4] for more detailed information.
1) Denoising and Artifact Removal: This step aims to reduce the distorting effects of patient breathing, skin stretching, power line interference, and muscle contraction. ECG denoising systems are usually based on moving average filters, frequency selective filters, Wiener filters, adaptive filters, and discrete wavelet transform [31]. More information on ECG denoising methods can be found in [31].
2) Segmentation: The denoised ECG signal is usually segmented into quasi-periodic units by automatically recognizing heartbeats. A heartbeat consists of several electrical waves, called P, QRS, and T waves, which represent the depolarization (systole) and repolarization (dilation) of the heart's atria (atria and ventricles) [32].
Most ECG clustering studies detect the peak of the QRS complex, the R-peak, and consider the interval between two consecutive R-peaks along the signal, the entire cardiac cycle, as the unit of segmentation. Few studies considered other feature points of ECG signal for segmentation [33]–[35]. Given that the abnormal morphology of the PR, ST, and TP segments of ECG cycles can indicate common cardiac diseases such as myocardial ischemia, hypokalemia, and atrial fibrillation [32], adding other ECG feature points in segmentation can improve the clustering results. Instead of identifying cardiac cycles in ECG signals for segmentation, some studies divide the signal into fixed time intervals without identifying any feature points [36], [37]. There are also ECG clustering methods that do not perform segmentation [38], [39]. These methods directly extract features from ECG signals without identifying any physiological feature points.
3) Feature Engineering: This step aims to obtain the most informative features of ECG segments to facilitate downstream unsupervised learning tasks. Here, we briefly introduce traditional ECG feature engineering methods. More recent deep learning methods are presented in Section II-C.
Professional physicians typically examine the timing and amplitude characteristics of the P, QRS, and T waves to diagnose heart disease. However, cardiac abnormalities are not always visible in the temporal domain [40]. The frequency information of ECG signal obtained by power spectrum analysis and time-frequency analysis (such as wavelet transform) can fill this gap. In wavelet transform, the correlation between the input ECG and a set of finite persistent functions called wavelets is regarded as ECG features [36], [41]. Interested readers can refer to the following sources [3], [4], [29], [30] for more details on feature engineering methods in time domain, voltage domain, and frequency domain.
Recently, swarm-based optimization methods inspired by nature, such as the firefly algorithm and particle swarm optimization, have also been used for feature engineering [42], [43]. These methods search for those features that perform best in terms of classification or clustering performance among a large population of possible features. For example, Kora [42] considers each point on an ECG signal as a possible feature. Using the firefly algorithm, she looks for points that maximize the accuracy of the neural network used to classify ECG segments into normal and myocardial infarction categories. For more information on nature-inspired feature engineering methods, we refer readers to the following sources [44]–[46].
Finally, the segmented units of the ECG are clustered, with each group containing those segments whose corresponding eigenvectors are more similar to each other than to eigenvectors in other groups according to a predefined similarity measure.
Clustering Algorithm
In clustering, a measure of similarity (or dissimilarity) that measures the distance between two ECG segments is crucial. Among various existing similarity measures, Euclidean distance, cosine coefficient, and dynamic time warping distance [47] are three widely used metrics in ECG clustering. The former two are usually used to measure the similarity between ECG units represented by temporal and morphological features or wavelet coefficients. Dynamic time warping is a method of measuring the similarity between two time series, which may be of different lengths. Here, we briefly introduce traditional clustering algorithms for ECG clustering. Recent deep learning based clustering algorithms are described in Section II-C. The strengths and limitations of the presented algorithms for ECG analysis are discussed in Section II-D and summarized in Table I.
Clustering based on medoids
The medoid-based clustering technique divides ECG segments into different groups based on their similarity to the medoids of these groups. Center points are considered as representative segments of their corresponding groups. K-means is the best known medoid-based clustering algorithm, which considers the cluster medoid as the average of the ECG segments (or their eigenvectors) at the center of that cluster. In unsupervised ECG analysis, variants of K-means include fuzzy C-means [48], affinity propagation [49] and max-min clustering [50]. These algorithms differ in obtaining the center point. For example, affinity propagation acquires central points by exchanging messages carrying similarities between ECG segments. Clustering algorithms based on medoids are usually easy to implement and have low computational cost. However, they are generally not suitable for dealing with noise, outliers, and high-dimensional feature spaces.
hierarchical clustering
Hierarchical clustering treats each ECG segment as an independent cluster and merges the most similar clusters until only one cluster remains (including the entire dataset). Hierarchical clustering algorithms are generally more computationally expensive than medoid-based clustering algorithms; however, these methods are unique in the resulting dendrograms, which visualize the hierarchical relationships between clusters and help Dr. Yu's explanation [51].
distribution-based clustering
The goal of this type of clustering algorithm is to find the probability that an ECG segment belongs to each cluster. Gaussian Mixture Model (GMM) is a well-known distribution-based clustering algorithm that has been widely adopted in reviewed studies [36], [37]. GMMs assume that multiple Gaussian distributions generate ECG segments; that is, each cluster is defined by a Gaussian distribution with a mean and a standard deviation around the mean. Dirichlet process GMM is a variant of GMM that does not require an initial number of clusters (i.e. Gaussian components in the data space) [52]. DPGMM automatically learns the number of clusters through variational Bayesian inference, an iterative algorithm that estimates the prior distribution of clusters. Distribution-based clustering algorithms are suitable for dealing with noise and outliers, but usually incur a high computational cost. It is worth noting that DPGMM is more computationally expensive than GMM.
Density-based clustering
Density is usually defined as the number of data points within some predefined radius. Density-based clustering views clusters as regions of higher density in the data space. DBSCAN [53] and self-organizing map (SOM) [54] are well-known density-based clustering algorithms that have been used for ECG clustering. DBSCAN considers a region as a cluster if its density exceeds a predefined threshold. It handles noise and outliers effectively; however, the resulting clusters depend heavily on the choice of radius and threshold. A self-organizing map (SOM) is a type of neural network that maps input segments into a two-dimensional grid, assuming a specific topology between ECG segments. The resulting mesh bends and distorts in high-density regions [54]. SOM provides an interpretable organization of clusters in a two-dimensional grid; however, it incurs a high computational cost.
spectral clustering
Spectral clustering transforms the clustering problem into a graph segmentation problem [55], where the goal is to partition a graph into subgraphs such that the sum of the weights of edges connecting the subgraphs is minimized. In ECG clustering, ECG segments are considered as nodes and the similarity between them is expressed as the weight of the edges connecting the nodes [56]. Spectral clustering can effectively handle high-dimensional feature spaces, but incurs high computational and space costs [47].
Clustering based on swarm intelligence
The swarm intelligence model treats the clustering problem as an optimization task whose goal is to maximize the overall similarity between electrogram segments within a cluster. For example, in the ant colony clustering algorithm [57], a group of ants randomly moves from one ECG segment to another and assigns a value (i.e., pheromone) to the segments based on their similarity. Clusters are then identified as segments whose similarity values exceed a predefined threshold. Particle swarm optimization [58] and artificial bee colony [59] are other swarm intelligence based algorithms [41], [60], [61] for ECG clustering. The clustering algorithm based on swarm intelligence can avoid local optimal solutions when searching for the best cluster solution, and has high-quality clusters. However, due to the stochastic nature of these algorithms, they often incur high computational costs, especially in large-scale datasets.
Maximum Margin Clustering
Maximum Margin Clustering (MMC) [62] utilizes Support Vector Machines (SVM) to perform clustering on unlabeled data. Specifically, it finds labels for a set of ECG segments that maximize the separation obtained by running an SVM on the labeled segments. The main disadvantage of MMC is its computationally expensive step for solving non-convex integer problems [34].
Ensemble clustering
In ensemble clustering, the results from multiple runs of one or more clustering algorithms are combined to arrive at better consistent clusters of data than would be obtained by individual clustering algorithms. For example, Abawajy et al. [38] integrated the results of K-means and GMM for ECG clustering, while Aidos et al. [63] built an ensemble of 200 K-means runs, each with a different value of K .
permutation distribution clustering
This algorithm is specifically designed to cluster time series by analyzing differences in permutation distributions to find similarities between time series. This is achieved by counting the frequency of different sequential patterns in the time series embeddings [64].
Clustering based on deep learning
Clustering algorithms based on deep learning have received a lot of attention recently and have achieved superior performance than traditional machine learning algorithms in many tasks [65]. The main advantage of these algorithms over traditional clustering algorithms is that they skip the traditional feature engineering step and can automatically learn the optimal feature set for clustering. Deep learning-based clustering methods are classified into three categories according to their architecture [65]: (1) autoencoders, (2) feed-forward networks, and (3) deep generative models. For each category, we present recent advances in ECG clustering and introduce state-of-the-art algorithms that promise to further enhance the performance of ECG clustering.
Deep Autoencoder
An autoencoder consists of an encoder, which is a neural network that converts input data into a low-dimensional feature vector, followed by a decoder, which is a neural network that reconstructs the original input from this low-dimensional feature vector network. The encoder and decoder are trained simultaneously to minimize reconstruction loss: the difference between the original input and the decoded output. Deep Clustering Networks [13] and Deep Embedding Networks [14] are two common autoencoder-based clustering algorithms that have been used for ECG analysis [66]–[68]. The idea behind these algorithms is to introduce a clustering loss in addition to the reconstruction loss when training the network. In deep clustering networks, the K-means loss is introduced, while in deep embedding networks, two constraints, preserving locality and group sparsity, are introduced to preserve the local structure of the data and to check the relatedness of representations. Keratosis. Some applications of these algorithms are further reviewed in Section III-A.
Deep Feedforward Networks
This group of algorithms only introduces clustering loss when training deep networks. The network architecture can be fully connected, convolutional, or a combination of both. The weights of the network can be initialized randomly or fine-tuned using a Restricted Boltzmann Machine on a pretrained network [69]. Deep Adaptive Clustering (DAC) [70] is a popular deep feedforward clustering network. It is a single-layer convolutional neural network (CNN) trained with a binary pairwise classification method. Briefly, a CNN is first used to map input samples to a one-hot encoded vector. Then calculate the cosine distance between all pairs of samples. Since the ground truth similarity is unknown, an adaptive learning algorithm, i.e. adaptive latent variable learning [71], is used to train the weights of the CNN based on the estimated similarity. DAC was originally proposed for image clustering and achieved superior performance on several challenging image datasets. With some modifications to its CNN architecture, DAC can also be used for ECG analysis as a possible future direction.
deep generative model
Variational Autoencoders (VAEs) [72] and Generative Adversarial Networks (GANs) [15] are the most popular deep generative models in recent years. VAEs enforce that the latent representations learned by an autoencoder follow a predefined distribution, usually a mixture of Gaussians. Variational Deep Embedding (VDE), a VAE-based clustering algorithm, has been used for ECG analysis [74]. This algorithm can be viewed as a deep learning version of the traditional GMM clustering algorithm, where the feature space is learned automatically.
Generative Adversarial Networks (GANs) aim to generate a set of fake data based on real data, so that the distribution of fake data is similar to real data. In short, GAN consists of two sub-modules: (1) Generator, which is used to generate fake data, and (2) Discriminator, which is used to distinguish fake data generated by the generator from real data. Learn a set of parameters in G and D such that the minimax game between the generator and the discriminator reaches a Nash equilibrium.
CatGAN [75] is a common clustering algorithm based on GAN. It forces the discriminator to classify the training data into a predefined number of classes (rather than just fake and real data), while being less confident in classifying the samples generated by the generator. ClusterGAN, a variant of CatGAN, has recently demonstrated superior performance among many other deep learning-based clustering algorithms, performing well in different clustering tasks [76]. Since these algorithms have not been applied in ECG analysis, their application in ECG analysis may be a future research direction.
In addition to clustering, deep generation-based algorithms can also learn to generate new samples from the obtained clusters. Recently, several studies have employed GANs to generate new heartbeat data to address one of the persistent data imbalance issues in supervised ECG abnormality classification [18], [19]. In Section III-E, we discuss applications of these methods.
Comparison of Clustering Algorithms in Electrocardiogram Analysis
Table I provides a comparison of different ECG clustering algorithms. The ability of clustering algorithms to deal with noise and outliers is an important factor in ECG clustering, since outliers are present in most publicly available ECG datasets, which may adversely affect the structure of the clustering results.
The time complexity of the clustering algorithm is another important factor. For applications that require real-time analysis of ECGs (for example, in an intensive care unit setting), computationally expensive clustering algorithms may not be a viable option, although the probability of producing high-quality clusters may be high. K-means, fuzzy C-means, and max-min algorithms are less computationally expensive; however, their ability to handle outliers is lower than more computationally expensive algorithms such as DBSCAN, GMM, hierarchical clustering, spectral clustering, and deep learning-based clustering. Among these algorithms, DBSCAN has the lowest computational cost, although it is sensitive to the choice of its hyperparameters (neighborhood radius and minimum number of points in the neighborhood). Furthermore, DBSCAN is not applicable in cases where the inherent density of the data space is non-uniform.
Finding clusters of arbitrary shape (i.e., non-convex shape) and handling high-dimensional feature spaces are other important factors when choosing an efficient ECG clustering algorithm. While distance-based clustering algorithms such as K-means are mostly able to find clusters of convex shape, clusters in ECG datasets may assume arbitrary shapes. Density-based and deep learning-based clustering algorithms are able to efficiently find non-convex shaped clusters, but at a higher computational cost than distance-based clustering algorithms.
The feature space in ECG analysis is usually high-dimensional because many features are usually extracted from ECG, and the number of training ECG samples is usually limited. Graph-based and deep learning-based clustering algorithms usually incur high computational costs. In particular, deep learning-based algorithms are more effective in dealing with high-dimensional data than traditional algorithms; however, deep learning methods require a large amount of data for training.
In addition to the above considerations, some clustering algorithms provide unique capabilities for visualization. For example, hierarchical clustering provides a dendrogram that visualizes hierarchical relationships between clusters, and self-organizing maps provide a two-dimensional grid that visualizes some specific topology in a dataset. For more information on clustering algorithms and their strengths and weaknesses, we refer the reader to [47], [65], [77].
Application field
To date, we have categorized unsupervised ECG analysis research into six application domains, as shown in Figure 1. The most mature application is heartbeat clustering, which provides a concise and understandable organization of heartbeats in large volumes of ECG data. Recent and innovative research directions seek to explore the relationship between the cardiovascular system and the whole body and mind. Unsupervised ECG analysis has also been used to improve the performance of supervised anomaly detection and ECG-based authentication systems.
heartbeat clustering
The effectiveness of heartbeat clustering is typically measured on ECG datasets with cardiologist labels. Each resulting cluster is expected to contain only heartbeats belonging to one label. Classification metrics (such as accuracy and sensitivity) and similarity metrics (such as Jaccard coefficient and normalized mutual information) are widely used as success metrics. The values of the Jaccard coefficient and the normalized mutual information range from 0 to 1, with higher values indicating that the identified clusters match the true labels well. Silhouette score is another widely used measure that does not require true clusters. The silhouette score measures how similar a sample is to other clusters and ranges from -1 to 1, with higher values indicating a sample that matches its cluster well and separates it from other clusters.
Datasets used to evaluate ECG clustering methods include MIT-BIH arrhythmia dataset, Physikalisch-Technische Bundesanstalt (PTB) dataset, St.-Petersburg Institute of Cardiological Technics 12-lead arrhythmia (CTAD) dataset, UCR Arrhythmia dataset and BIDMC congestive heart failure dataset, among which the MIT-BIH arrhythmia dataset is one of the most commonly used datasets.
Among many studies, Lagerholm et al. designed an efficient heartbeat clustering system to classify QRS complexes represented by wavelet coefficients into 25 clusters and obtained results with high accuracy (98.5%) using self-organizing maps . By using self-organizing maps, they provide a neighborhood map (such as a 2D grid) that preserves certain topological information within the dataset, ultimately aiding interpretation by cardiologists.
A large number of studies have focused on improving the accuracy of ECG clustering by employing various clustering and optimization techniques, such as ant colony clustering, bee swarm clustering, maximum margin clustering, Gaussian mixture models, hierarchical clustering, K-means, parent and Propagation and Deep Autoencoder Networks. Among these algorithms, the clustering system proposed by Balouchestani and Krishnan achieved the highest accuracy (99.98%) on the MIT-BIH dataset. They used methods based on K-means, compressive sensing theory, and K-singular value decomposition to classify heartbeats into five groups: normal, upper ventricular ectopy, ventricular ectopy, fusion, and unclassifiable.
Despite the encouraging results of these deep learning studies, there is still a need for more advanced clustering algorithms that can automatically handle the imbalanced data problem without the need for preprocessing algorithms and expert analysis. In this regard, algorithms based on deep generative models, such as ClusterGAN, can be further investigated, which can learn to generate new samples from a small number of clusters. Future research should also focus on applying deep learning-based clustering algorithms on large-scale public ECG datasets, such as the dataset collected by Zheng et al. and the dataset by Wagner et al.
From a clinical perspective, several studies have performed innovative ECG clustering using symbolization. They converted the ECG signal into a string of symbols by clustering the heartbeats using a max-min clustering algorithm and assigning symbols to each identified cluster. Under the symbolic representation, they search for entropy-increasing subsequences that represent irregular activity. Their method successfully detected an atrial ectopic rhythm sequence that had been overlooked by cardiologists. Similarly, they extended their work to risk stratification, successfully identifying patients who were at increased risk of death despite similar treatment within 90 days of treatment for an acute coronary syndrome. From a clinical monitoring point of view of the heart, a 12-lead ECG is a standard clinical protocol that records electrical activity by attaching electrodes at 10 different locations on the patient's body. In order to obtain features from a 12-lead ECG, the features extracted from each lead are usually concatenated. However, this representation typically fails to preserve the relative positions of the 12 signals. To address this issue, He et al. used tensor decomposition techniques. They expressed the ECG recorded through each lead as its wavelet coefficient: W ∈ RV × L, where V and L denote the number of leads and wavelet coefficients, respectively. Then, by tensorization, W is decomposed into W′ ∈ RI1×I2×I3, where I1, I2 and I3 represent the recorded signal, sampling time and wavelet frequency subbands, respectively. They used a Gaussian mixture model to separate the tensor representation of 12-lead ECGs into two clusters, corresponding to normal and abnormal ECGs. Their system achieved a high Jaccard coefficient of 0.93 on a subset of the CTAD dataset consisting of two 30-minute 12-lead ECG recordings.
ECG and state of mind
State of mind, mood, and mental disorders are often associated with the autonomic nervous system (ANS). There is a bidirectional interaction between the autonomic nervous system and the heart via the sinoatrial node (SA node) of the heart [8], [115], [116]. The sinoatrial node, also known as the pacemaker of the heart, generates electrical impulses that stimulate the heart muscle to contract and pump blood [32]. In recent years, ECG clustering has helped researchers in fields such as psychophysiology to discover hidden ECG patterns associated with different mental states.
The datasets used in these studies typically consist of electrocardiogram (ECG) and impedance cardiography (ICG) signals obtained from healthy individuals in different emotional states or from patients with mental disorders. The feature engineering stage relies on the combination of a series of extracted ECG and impedance cardiogram features. Here, instead of assessing the quality of the resulting clusters, a statistical hypothesis test, such as a t-test, is often performed to measure the correlation of the ECG (and impedance cardiographic) features in each cluster with the state of mind of individuals in that cluster. Treat different mental states as ground truth clusters and identify them manually through questionnaires, or automatically by clustering electroencephalogram (EEG) signals. Therefore, metrics that measure the similarity between identified clusters and ground truth labels, such as NMI [105] and Jaccard coefficient [104], can be used to improve the reliability of these studies.
In this section, we review novel studies in this research direction. A summary of the studies reviewed is detailed in Table III. Since the reviewed studies used various private datasets for different psychological states, these clustering techniques were not compared with each other. However, we critically review existing research in terms of the clustering algorithms employed, feature engineering, and experimental protocols, and provide directions for future research.
Recently, Hoemann et al. [98] clustered daily electrocardiogram (ECG) and impedance cardiography (ICG) signals acquired from 67 participants to investigate the association between cardiorespiratory activity and emotional finesse. Emotional granularity describes an individual's ability to accurately distinguish emotions. Lower emotional granularity has been associated with psychiatric disorders, including schizophrenia, autism, and depression [98]. Hoemann et al. employed a Dirichlet process Gaussian mixture model to find the optimal number of clusters in the data. They found that ECG and ICG can be used to identify different levels of emotional nuance.
Leal et al. [8] studied the relationship between the ECG time interval characteristics of epilepsy patients and the pre-epileptic interval (ie, the short time before the seizure) by ECG clustering to predict seizures and allow patients enough time to prepare for the seizure. Imminent epilepsy. They performed K-means, DBSCAN, and Gaussian mixture model clustering on intervals extracted from ECGs of patients with epilepsy to see if there was a cluster that was clearly separated from the other clusters, representing the pre-epileptic interval. They found that such a cluster was present in 41 percent of seizures, representing the interval between 2 and 9 minutes before the seizure. Babaeian and Mozumdar [9] proposed a system to detect driver drowsiness by clustering ECG collected by wearable devices. They performed density-based clustering of time interval features and found three clusters associated with wakefulness, drowsiness, and sleep states.
Carreiras et al. [86] aimed to detect individual declines in attention while solving challenging mathematical problems by ECG clustering. Detecting attention deficits is important in attention-demanding tasks, such as surgery and driving, where lack of attention can be disastrous. Their work was inspired by the fact that it is more convenient to obtain an ECG through a wearable device than an EEG through a headset. They performed consensus clustering of ECG and EEG signals acquired while solving math problems from 24 subjects, which included multiple runs of the hierarchical algorithm using different distance measures. Their results showed a strong correlation between the clusters found in the ECG dataset and those in the EEG dataset, suggesting that ECG can help detect different levels of attention. Another finding showed that the number of ECG clusters was larger than those found in EEG datasets, which could provide more accurate information for in-depth analysis. Similarly, Wang et al. [117] showed a strong correlation between the clusters obtained by analyzing ECG obtained from the driver's palm and the clusters obtained by analyzing EEG. They suggest that ECG collected from the driver's palm, rather than EEG, could be used to identify different levels of driver attention.
Another application of ECG clustering is emotion detection. Its goal is to automatically identify different emotional states, such as joy and sadness, through clustering. Wan-Hui et al. [118] found that frequency-domain features of ECG signals were more able to distinguish joy from sadness than time-domain features. Zheng et al. [95] used the fuzzy C-means clustering algorithm to classify ECGs into two clusters of emotional stress and non-emotional stress, and showed the important role of time interval features in distinguishing these two clusters. Medina [119] performed ensemble clustering, including K-means and spectral clustering, on ECGs obtained from 25 subjects while solving a math problem. Their system successfully grouped together subjects with similar levels of emotion.
In an innovative study, Kupper et al. [6] investigated the relationship between emotional stress and cardiorespiratory activity in 744 young adults during stressful activities, including solving math problems and giving a presentation in front of two audiences. Relationship. They obtained pre-task and intra-task ECG and ICG signals from participants. Using distributed clustering, they found five clusters of participants that differed in autonomic balance and resting systolic blood pressure levels. The results also showed that smoking, regular physical activity and body mass index (BMI) were not associated with these clusters. In addition, men were more likely to be in the cluster of elevated systolic blood pressure and increased cardiac output when performing stressful tasks.
Gonzalez-Vel´azquez et al. [120] studied the relationship between emotional eating behavior and ECG by clustering. They performed K-means clustering (K=2) on the ECGs of 52 young adults, dividing individuals into two groups with and without emotional eating behavior. They found that emotional eating behavior was more prevalent in overweight individuals (BMI > 85th percentile). In addition, the high-frequency components were significantly larger in the RR time series of obese individuals.
Inspired by the success of deep learning techniques, Oskooei et al. [99] recently trained a convolutional autoencoder on the RR time series of 100 firefighter trainees to identify groups under significant stress while conducting drills. They applied DBSCAN to the latent representations learned by the autoencoder and found two clusters, a smaller one corresponding to firefighters who exhibited significantly more stress. They further show that, for this task, applying K-means fails to find groups of trainees under stress.
Various statistical hypothesis tests can be used to infer significant relationships between ECG characteristics and different psychological states. Most of the reviewed studies used t-tests, which assume a normal population distribution. However, this assumption requires further validation as the collected dataset is small and may not fit this distribution. Nonparametric tests, such as Friedman's test [121] and Spearman's rank correlation [122], can be further investigated in this application because they do not rely on the assumption of normality.
Furthermore, the ECG feature engineering phase in most studies is limited to RR interval and heart rate. Therefore, there is an unmet need to investigate the relationship between other ECG features, such as PR and QT intervals and voltage and frequency domain features, and different psychological states. Furthermore, the clustering algorithms used are limited to traditional K-means, hierarchical clustering, Gaussian mixture models and DBSCAN. In particular, deep learning-based clustering algorithms have not been used for this purpose. Larger datasets are needed to develop and train reliable deep learning algorithms to discover the relationship between ECG and different psychological states.
Finally, as mentioned before, most studies treat different mental states, such as different emotions or different degrees of psychological disturbance, as clusters of ground truth. The number of these clusters is required prior knowledge for most clustering algorithms used. However, a few studies have used algorithms that do not require such prior knowledge, such as DP-GMM [52] and DBSCAN [53]. They obtain a larger number of ECG clusters than the ground truth would suggest. An increase in the number of clusters generally results in better separation between different ECG modalities, but can make it difficult for experts to interpret the clustering results. Future work needs to focus on finding the optimal number of ECG clusters.
ECG and physical status
The focus of this research direction is to discover different clinical phenotypes through clustering, including ECG abnormalities, blood pressure conditions, metabolic indicators, and demographics, etc., to study the differences between patients with different diseases. Several studies have aimed to reveal underlying mechanisms and salient biomarkers in subgroups of patients with similar ECG patterns, which are nearly impossible to do by visual means.
The datasets used in these studies often include ECG and blood pressure signals from individuals with different disease conditions (e.g. diabetes, arteriosclerosis, embolic stroke) or chronic habits (e.g. smoking). After clustering the ECGs, determine the dominant ECG pattern within each cluster. Hypothesis testing is then performed to see if this pattern is associated with the disease (or level of disease severity) represented by the cluster. Since the ground truth clusters are often unknown in such applications, metrics that measure the similarity between within and without clusters, such as silhouette coefficients, can be used to improve the reliability of these analyses. Here, we review some innovative studies in this field of research. A summary of the studies reviewed is shown in Table IV.
Wang et al studied the effect of heart rate and blood pressure on the prediction of orthostatic cardiovascular dysregulation in patients with spinal cord injury. They performed hierarchical clustering of ECG and blood pressure signals obtained from 207 subjects (48 controls) while lying flat and passively turned into a sitting position. The clustering results with the best silhouette coefficients divided subjects into eight groups. They found that heart rate, systolic and diastolic blood pressure were effective in identifying the prevalence of cardiovascular dysregulation in the SCI population.
Tseng et al studied the relationship between ECG, diabetes, obesity, hypertension and smoking habits. Using ECGs from 268 subjects in the PTB dataset, they performed K-means clustering on time-interval features, classifying patients into eight groups. The results showed that almost all diabetic patients were assigned to the same group, suggesting a strong association between diabetes and EKG. However, smoking, hypertension, and obese patients were distributed in all clusters, indicating a weaker correlation between these conditions and the time-interval characteristics of the ECG.
Hernandez et al studied the relationship between electrocardiogram and physical activity capacity. They obtained electrocardiograms from 67 male subjects during rest, cycling and recovery states. During the electrocardiogram acquisition, the wearable body composition analyzer recorded the amount of fat stored in the abdominal cavity. They applied hierarchical clustering to time intervals and wavelet-extracted features, and found a four-cluster solution that properly separated the data space by analyzing dendrograms. Next, they applied K-means (K=4) and found the following four groups: (1) individuals with high physical work capacity, (2) younger individuals with lower physical work capacity, (3) older , individuals with low physical work capacity and low to moderate abdominal fat, and (4) older individuals with low physical work capacity and high abdominal fat.
Lattanzi et al studied the associations between ECG abnormalities, demographics, metabolic markers, and smoking habits in 127 patients with embolic stroke of unknown origin. Cardiac abnormalities identified by cardiologists, such as atrial fibrillation and hypertension, were used for clustering. They performed hierarchical clustering and found three subgroups of patients: (1) young men with patent foramen ovale and posterior circulation infarction, (2) those with hypertension, major stroke, left atrial heart disease, diabetes and multiple vascular field lesions, and (3) smoking patients with dyslipidemia, ipsilateral vulnerable carotid stenotic plaques, and anterior circulation field infarcts.
Hyun et al studied the association between ECG and blood pressure and atherosclerotic disease. They applied a consensus clustering method to cluster Holter and blood pressure signals obtained from 989 patients. They found 16 clusters, two of which contained a significant proportion of patients at high risk for atherosclerosis. Metabolic indicators, including diabetes, body mass index, and total cholesterol, were significantly higher in both clusters. Notably, age is generally associated with all clusters.
Porumb et al. trained a convolutional autoencoder on ECGs obtained from subjects with nocturnal hypoglycemia (lower blood sugar levels during sleep) to predict drops in blood sugar levels. They clustered and visualized the learned latent representations using a t-distributed stochastic nearest neighbor embedding method and showed that the autoencoder effectively separated ECGs recorded during low blood glucose levels from those recorded during normal blood glucose levels. They took the latent representation as input to a convolutional neural network and trained the network on labels provided by experts to classify ECGs as normal and low blood sugar levels. Their study achieved 90 percent accuracy in eight subjects who experienced nocturnal hypoglycemia.
In this research field, hierarchical clustering algorithms are widely used. The algorithm does not require the number of initial clusters and provides a hierarchical visualization of the resulting clusters. This visualization can greatly assist researchers in identifying underlying mechanisms and biomarkers within each subpopulation. Self-organizing maps [54] and t-distribution stochastic neighbor embedding [125] are other well-known algorithms that provide a 2D and 3D map for visualizing topology in the data space. Since these two algorithms preserve the local and global structure of the data, they are suitable candidates for this application.
Furthermore, the clustering algorithms used by the reviewed studies were limited to K-means and hierarchical clustering. Since the datasets collected by these studies are significantly larger than publicly available datasets (such as MIT-BIH), applying deep learning-based clustering algorithms can further improve the results.
Finally, similar to studies exploring the relationship between ECG and mental state, the extracted ECG features were also limited to RR interval and heart rate. Therefore, it is necessary to study and apply other time-domain, voltage-domain and frequency-domain features.
ECG-based biometric authentication and identification
Biometric authentication is the process of identity verification based on an individual's physiological characteristics, mainly including fingerprints and faces. Fingerprints and facial patterns are vulnerable to external attacks because their physical characteristics are easily exposed. However, ECG-based authentication systems are difficult to spoof because the basic biometrics of the heart's electrical activity are hidden.
One of the challenges of ECG-based authentication systems is intra-individual variability, which is due to differences in the physical and mental states of different individuals, which can lead to authentication failures. Several studies have aimed to improve the robustness of ECG-based authentication systems to within-individual variability through cluster analysis. The idea is to divide an individual's ECG (or heartbeat) into clusters under different conditions, such as individuals in different mental or physiological states, and use information about the clusters, such as the centers of the clusters, as a supervised learning method for performing authentication additional features. Most of the studies in this app used their ECG datasets recorded at different levels of emotion or stress. Instead of assessing the quality of the resulting clusters, classification metrics, such as sensitivity, specificity, and F1-score, are used to evaluate performance on downstream supervised certification tasks. Since the ground truth clusters are mostly unavailable in this application, metrics evaluating inter- and intra-cluster similarity, such as silhouette scores, can be further used to improve the reliability of these studies.
As an example of modern research, Zhou et al. [23] used Gaussian mixture model clustering to improve the robustness of authentication systems under stress. In particular, they divided the subjects' ECGs into several groups with different levels of stress. The centers of these clusters, combined with the latent representation of the ECG learned by a convolutional autoencoder, are provided as input feature vectors to a support vector machine that performs authentication. They tested their system on 23 healthy subjects under different stress conditions and achieved an average recognition rate of 95% and an average F1 score of 0.97.
Similar to authentication, biometric identification is the process of identifying an individual based on their biometric characteristics against a database of previously identified templates. ECG-based recognition systems typically incur high computational costs due to the need to cross-match a given ECG with all template ECGs stored in a database to find a match. Clustering helps reduce the computational cost of such systems by clustering template ECGs. During identification, only the centers of clusters most similar to a given ECG signal are searched. Neehal et al. [126] used K-means to divide the templates in the 50000 ECG database into five clusters. Searching only the most similar clusters during recognition, they reduced the recognition time by 79.26%. Under a similar approach, Sufi et al. [127] proposed a recognition system based on compressed ECG data. Compressed ECG data is often required in wireless cardiovascular monitoring. However, decompressing millions of compressed ECG signals is time-consuming. To address this issue, Sufi et al. designed a system based on Gaussian mixture models to directly cluster compressed ECG signals from a template ECG database.
The clustering algorithms used in these studies are mainly limited by K-means and GMM. Therefore, further research is needed on other clustering algorithms for ECG clustering, such as DBSCAN [53] and deep learning based methods. The dataset for an ECG-based identification system is very small (n < 30), and a larger dataset of ECGs recorded at different levels of emotion or stress is needed to further increase robustness to intra-individual variation. Research is also needed to develop clustering algorithms that can detect the synchrony of the ECG with other physiological signals [128].
Improving Supervised Anomaly Classification
In addition to knowledge discovery, clustering and deep learning-based unsupervised techniques can also be used to improve the performance of ECG classification systems and overcome challenges. In this application, the quality of the identified clusters is rarely assessed. Instead, classification metrics, such as sensitivity, specificity, and F1-score, are often used to evaluate performance on downstream classification tasks. However, since the ground truth clusters are known, the aforementioned classification metrics as well as similarity metrics, such as NMI and Jaccard coefficient, can be used to evaluate the performance of the clustering stage and improve the reliability of these studies. The MIT-BIH dataset is widely used in this research area. Since this dataset is relatively small, future work should also focus on using larger datasets, such as those collected by Zheng et al. [112] and Wagner et al. [113].
A persistent challenge in accurately classifying ECG abnormalities is the problem of severe data imbalance, since the labels provided by cardiologists in public ECG datasets are overwhelmingly normal heartbeats. For example, in the MIT-BIH dataset, more than 75% of labeled heartbeats belong to the normal category, while less than 1% of the labeled heartbeats belong to four abnormal categories, which are ventricular fibrillation, nodal escape, atrial premature beat, and ventricular escape. This unbalanced data leads to poor performance of classifiers in detecting minority classes. A well-known technique to overcome this problem is undersampling, where samples are randomly removed from the majority class in order to balance the training set [129]. However, this technique may lose relevant information that is crucial for classification tasks. To reduce information loss during undersampling, Carrillo-Alarcón et al. [16] clustered heartbeats in each majority class in the MIT-BIH dataset using self-organizing maps.
Unsupervised techniques based on deep learning have also been used to enhance the automation and performance of anomaly classification systems. [131] improved the performance of their deep neural network classifiers by initializing the weights of each layer using a greedy unsupervised algorithm. Each hidden layer is treated as a restricted Boltzmann machine [69] and optimized using the contrastive divergence algorithm [132] - a well-known unsupervised algorithm for training energy-based latent models. The entire network is then fine-tuned by minimizing the cross-entropy loss between ground truth labels and predictions. They evaluate their method using three patient-specific and one patient-independent experiments on the MIT-BIH dataset. Their system achieved 93.1%, 94.7%, and 99.9% accuracy on the three individuals in MIT-BIH, respectively. In patient-independent experiments, their system also showed good generalization to unseen patients, but with a lower accuracy of 91.8%.
Patient-specific ECG classifiers - classifiers trained and fine-tuned on a given patient's ECG - have shown superior performance compared to classifiers trained on a generic pool of ECGs. Zahi et al. [17] showed that retuning the classifier for patient-specific normal heartbeats can improve classification performance on MIT-BIH. Although they perform better, patient-specific classifiers have a lower level of automation because they require manual labeling of parts of the ECG for fine-tuning. To address this issue, Zahi et al. [17] proposed an unsupervised method to automatically identify normal ECG heartbeats. They clustered heartbeats based on their similarity to neighboring heartbeats and identified the heartbeat that exhibited the highest average similarity in the cluster as normal. Then, they fine-tuned their deep classifiers on identified normal heartbeats. Their system excelled at detecting two abnormality categories, ventricular and supraventricular ectopic heartbeats, achieving high accuracy rates of 97.4 percent and 98.6 percent, respectively.
In recent years, generative adversarial networks (GANs) [15] have also been used to improve the automation of patient-specific classifiers. The idea is to utilize the generator in a GAN to generate new patient-specific normal heartbeats. Zhou et al. [18] augmented the MIT-BIH dataset with GAN-generated normal heartbeats for more accurate training and classification of ventricular and supraventricular ectopic heartbeats, achieving an overall accuracy of 97%. Similarly, Golany et al. [19] trained a GAN on the first few minutes of unlabeled ECG data from each patient to generate normal heartbeats. In contrast to Zhou et al. [18] who used convolutional neural networks for arrhythmia classification, they used long short-term memory neural networks [133] and achieved similar good performance.
Deep learning-based unsupervised feature extraction techniques also improve the performance of supervised classification systems compared to classifiers using hand-crafted features. For example, Nurmaini et al. [68] combined CNN-based deep autoencoders as an unsupervised feature extraction technique with deep neural networks for arrhythmia classification. Their system achieves a high F1 score of 0.92 on the entire MIT-BIH dataset.
Another advantage of combining unsupervised learning with supervised ECG classification is the use of transfer learning [134]. The core idea is to transfer the parameters of a model trained on a large dataset to another model for classification on a smaller dataset, which may have inaccurate labels or some labels are missing. Weimann et al. [135] trained a deep residual network classifier [136] on the Icential11K dataset [137] (the largest publicly available ECG dataset containing 11,000 patients), and then presented in Physionet/CinC 2017 fine-tuned their network on the dataset [138] for atrial fibrillation detection. Jang et al. [139] pretrained a convolutional autoencoder on over two million ECG samples. They then fine-tuned their network on another dataset containing 10,000 12-lead ECGs to detect 11 arrhythmia categories and achieved an F1 score of 0.857.
Other applications
Apart from the applications discussed, ECG clustering is used in some other interesting applications. Xia et al. [24] used ECG clustering to improve the accuracy of the QRS detection system. The core idea is that in an ECG, the absolute slope of a line between mutually truncated pairs of points is significantly higher than the absolute slope of a line truncated between any other pair of points. By K-means, they divide all point pairs into two clusters based on the absolute slope between point pairs. They found that a cluster mainly contained point pairs belonging to the QRS region. Among these points, the point with the largest amplitude was determined as the R-peak. They achieved a sensitivity of 99.72% and a positive predictivity of 99.80% on R-peak detection across eight records at MIT-BIH. Along similar lines, in a recent study, Chen et al. [25] applied hierarchical clustering to the mean amplitudes of each pair of points as well as their slopes, dividing the points into two clusters: the R-wave cluster and the non-R-wave cluster. They achieved 99.89% sensitivity and 99.97% positive predictivity in R-peak detection on MIT-BIH.
Zhou et al. [26] extracted fetal QRS complexes from abdominal ECGs by clustering. Notably, fetal ECGs obtained from the maternal abdomen are contaminated by maternal heart activity, fetal brain activity, and various noises such as uterine contractions. Zhou et al suggested that the amplitude of the RS peak could serve as a unique feature to differentiate the maternal QRS complex from the fetal ECG, as the amplitudes of the R and S peaks in the maternal ECG were significantly larger than those in the fetal ECG. They applied K-means to adjacent local maximum-minimum pairs in the ECG and found three clusters. One of the clusters contains maternal RS peaks, another contains fetal RS peaks, and the last cluster contains non-RS peaks.
Salman et al. [140] attempted to reduce the average waiting time by clustering remote patients into groups of varying urgency. They applied fuzzy c-means to features extracted from ECG and blood pressure signals and divided patients into five groups. The identified clusters correspond to patients in normal, cold, sick, emergency, and high-risk states. They treated each cluster as a queue and proposed an algorithm that minimizes the average waiting time while prioritizing urgent patients.
Discussion and future directions
Clustering and Contrastive Learning Based on Deep Learning
Although promising ECG clustering techniques have been researched, there is still an urgent need for more advanced algorithms that can do so automatically when processing large amounts of data without the need for preprocessing steps and expert analysis. The main advantage of deep learning techniques over traditional machine learning methods lies in the automatic feature extraction and selection process. Deep learning techniques have shown to outperform traditional machine learning methods in complex tasks such as speech recognition and image classification. However, few studies have focused on using deep learning for unsupervised ECG analysis. Therefore, a new generation of deep learning algorithms, such as deep adaptive clustering [70] and ClusterGAN [76] (as described in Section II-C) are expected to be applied to ECG clustering systems.
A disadvantage of deep learning techniques is their lack of interpretability, since features are extracted in a black box. This is an increasingly important issue in ECG analysis as the interest in how results were obtained is as important as the results themselves. Future research in this area should focus on explaining the interpretability of deep learning techniques in ECG analysis. Algorithms similar to DeepLIFT [141] can be investigated. Given an input, DeepLIFT assigns a contribution score to each neuron in the neural network by backpropagating the activations of each neuron from the predicted output to the input's feature.
In addition to deep learning, the application of contrastive learning can be further investigated for ECG clustering. Contrastive learning aims to learn an embedding space where similar data points are closer to each other without labeled data. It has recently shown superior performance in solving vision and language processing tasks. In ECG analysis, it can be used for unsupervised or semi-supervised analysis, where cardiologists annotate a small subset of the dataset. Interested readers can refer to [142], [143] for more information.
Unsupervised analysis of electrocardiogram (ECG) data recorded by a wearable device
When electrocardiogram (ECG) data is analyzed within a wearable device, the time and space complexity of the clustering method must be within the computational capabilities of the device. And when ECG data is transmitted to a remote server, the reliability and latency of the transmission channel become important. Therefore, efficient compression and encryption algorithms are required for optimal and secure transmission of ECG data.
In addition, robustness to noise is also very important because the quality of electrocardiograms (ECGs) recorded by wearable devices is usually lower than that of standard clinical devices. Even after denoising, there is no guarantee that the signal will be noise-free, as practical implementations of any denoising system are imperfect [148]. In this case, it is very important to use an artifact rejection algorithm.
Most of the reviewed methods were developed and validated on the resting state ECG, when the heart rate is usually below 120 beats per minute. However, there may be significant changes in heart rate during prolonged monitoring of a subject's exercise state. Therefore, the robustness of the method to changes in heart rate is also another important consideration. Analysis of ECGs recorded via wearable devices is an emerging field with little research focus and further research is needed. Interested readers are referred to the following sources for more information: [144]–[147], [150]–[152].
Flow ECG clustering
The amount of electrocardiogram data generated every day is huge, and due to limited hardware resources, it cannot be actually stored. In addition, real-time monitoring of high-risk patients and immediate detection of abnormal events are critical. Therefore, future clustering systems need to process continuously arriving ECG data, the so-called data stream.
Streaming ECG clustering poses several key challenges to traditional clustering systems. First, the ECG should be analyzed in only one pass, since it is impractical to store all arriving signals. Second, the clustering may change as new ECGs arrive. Third, cardiac events must be identified in real time. It is worth noting that all the methods reviewed in this review dealt with non-flow ECG, leaving room for possible future research. Interested readers can refer to the following sources for more information on dataflow clustering analysis [153]–[157].
A public database of electrocardiograms recorded in mental and physiological states
So far, considerable efforts have been invested in developing open ECG datasets representing different cardiac abnormalities [1], [112], [113]. However, as described in Sections III-B to III-D, studies have found relationships between ECGs and different psychological and physiological states, and the development of ECG-based authentication systems uses private datasets. This makes it impossible to compare methods and reproduce their results. In order to further develop these innovative research fields, an open ECG database containing ECGs recorded under different psychological states (such as stress or psychological disorders) and different health conditions (such as diabetes) is needed. Such a database needs to be sufficiently large and balanced with respect to the sex and age of individuals.
Feature engineering based on P wave, QRS wave and T wave
Most retrospective studies have not aimed to identify P and T waves for feature engineering, however, abnormal morphology of these waves can indicate important cardiac disorders such as myocardial ischemia, hypokalemia, or atrial fibrillation [32]. Furthermore, R waves are generally assumed to be present in all recorded heartbeats. However, it is worth noting that in some abnormal cases, such as those with right heart disease, R waves may be absent [32].
To address these issues, researchers can employ state-of-the-art ECG segmentation systems such as those developed by Martinez et al. [158] and those developed by Bote et al. [159]. These systems are capable of efficiently identifying P, Q, R, S, and T waves on the ECG, allowing extraction of temporal and morphological features of all major waveforms.
in conclusion
In this paper, we provide a comprehensive and critical review of unsupervised machine learning methods for electrocardiogram (ECG) analysis. We review traditional and state-of-the-art ECG clustering algorithms and discuss their strengths and weaknesses in detail. We also extensively review various applications of unsupervised ECG analysis, describe recent research in each application area, outline their limitations, and suggest future directions.
We believe that the clustering methods reviewed in this paper will continue to play an important role in future ECG monitoring in the context of unsupervised biomedical signal processing.