ECG recordings are classified into two types based on the recording duration, i.e., resting and ambulatory ECG (10). A resting ECG contains only 5 to 10 min of heart function data recording, whereas an ambulatory ECG records 24–48 h of information (11). Detecting abnormal episodes from this enormous quantity of data is very difficult. Therefore, effective automated diagnosis tools are required to detect important episodes in cardiac patients. Initially, we used template-based and rule-based techniques that are usually utilized to detect the type of ECG beat (12–17). Rule-based approaches rely on predefined rules and thresholds to classify ECG beats. These rules are often based on expert knowledge or heuristics. However, these rules may not be able to handle the wide range of variations and complexities observed in real-world ECG signals. As a result, rule-based approaches may struggle to adapt to different types of beats or handle new patterns that were not considered during rule creation (18). Developing accurate and comprehensive rules for ECG beat classification can be challenging. It requires a deep understanding of ECG signal characteristics and considerable domain expertise. Designing rules that cover all possible scenarios and variations is complex and time-consuming. Rule-based approaches are typically designed to classify beats based on specific features or patterns (19). They may struggle to generalize well to new or unseen data that do not conform to the predefined rules. The rule-based classifier can produce incorrect or inconsistent results if the ECG data deviates from the expected patterns. A set of templates representing different types of beats is required in template-based approaches (13). Choosing appropriate templates that accurately represent the various beat morphologies in ECG signals can be challenging. There is a need to consider inter-subject and intra-subject variability and variations due to different conditions and diseases (12). The classification in template-based approaches relies on comparing the input ECG beat with a set of templates to find the best match. However, template matching can be sensitive to noise, baseline wander (BW), and other artifacts present in the signal (20). These issues can affect the accuracy of the match and lead to misclassification. Template-based approaches often struggle with scalability when dealing with large datasets or real-time applications (17). Comparing each beat with a set of templates can be computationally expensive, especially if the number of templates is high. The time complexity increases as the number of beats and templates grows, making it impractical for large-scale applications (21).

Machine learning (ML) algorithms are becoming popular for effective classification of ECG beats (18, 19, 22–28). Hierarchical representations of the latest ECG beat classification techniques based on machine and deep learning (DL) techniques have been reported in the literature, as shown in Figure 2. Handcrafted features are required to detect the type of ECG beat when using machine learning techniques (22–24). Most handcrafted features are extracted based on time, frequency, and time-frequency domains (29, 30). Time-domain features alone are not sufficient for effective ECG beat classification; together with this, the frequency-domain features increase beat detection performance (11, 29). In addition, features based on time and frequency are more effective than individual features in the time and frequency domain (31). Handcrafted approaches rely on manual feature engineering, where domain knowledge is used to design and extract features from ECG signals. This process can be time-consuming, involving the design of algorithms and signal processing techniques to extract meaningful features (5). Handcrafted approaches require selecting relevant features that capture the discriminative information from the ECG signals. Identifying the most informative and robust features is non-trivial and often requires domain expertise (32). Choosing inappropriate features or excluding important ones can lead to suboptimal classifier performance. Handcrafted feature sets may not generalize well to new or unseen data that significantly differ from the training data. ECG signals vary considerably due to age, sex, underlying conditions, and noise (33). The classification accuracy may be compromised if the handcrafted features do not represent the new data. ECG beat classification involves capturing complex relationships and patterns within signals. Handcrafted approaches may struggle to capture these intricate relationships, as they typically rely on pre-defined algorithms and feature engineering techniques. These problems have motivated researchers to develop automatic feature extraction approaches.

Deep learning models can automatically learn relevant features directly from the raw ECG signals, eliminating the need for manual feature engineering. This feature learning process enables the model to capture intricate patterns and complex relationships that may be difficult to capture using handcrafted features (18, 19, 25–28). Deep learning models can learn hierarchical representations of ECG signals, allowing them to extract meaningful features at different levels of abstraction. The different deep learning architectures and their advantages in the extraction of meaningful features are presented in Table 1. Deep learning models are known for generalizing unseen data well. They can learn from large amounts of labeled ECG data and capture the underlying patterns that are characteristic of different beat types. Deep learning models can handle large-scale ECG datasets efficiently (20). Once trained, the models can process ECG beats quickly, making them suitable for real-time applications. In addition, deep learning models can be deployed on parallel computing architectures, such as graphical processing units (GPUs), to improve computational performance, enabling rapid and scalable ECG beat classification (9, 18). Deep learning models can be updated and fine-tuned with new data to improve performance and adapt to changes in ECG signals (25). This ability for continual learning allows the model to incorporate new knowledge and adjust its classification capabilities as new data become available. It enables the model to stay up-to-date with the emerging beat types or changes in the data distribution (28). Deep learning models can adapt and generalize well to new beat types and variations not encountered during training. The models can discover and classify novel patterns and variations by learning from diverse examples. This adaptability makes deep learning models suitable for dynamic environments where ECG data may evolve.

The remainder of this article is organized as follows. Section 2 discusses the main objectives and methodology of the proposed study. Section 3 refers to the publicly available databases for experimentation on ECG beat classification. The background and significance of ECG beat classification are discussed in Section 4, with pre-processing, feature extraction techniques, and ECG beat classification algorithms, along with their performance, elaborated on in Sections 4.1–4.3. A general discussion of machine learning and deep learning for the detection of ECG beats is presented in Section 5. Finally, limitations, future directions, recommendations from the state-of-the-art review, and conclusions are presented in Sections 6, 8, and 9.

We used the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) criteria to find studies pertinent to the classification of ECG beats (41). Articles published up until April 2024 utilizing the terms “Electrocardiogram,” “ECG beat Classification,” “Artificial Intelligence,” “Machine Learning,” and “Deep Learning” in their respective Boolean combinations were searched for in the PubMed, Institute of Electrical and Electronics Engineers Association (IEEE), and Science Direct databases. The authors excluded non-English publications, duplicate titles, irrelevant works, review articles, pilot studies, non-accessible articles, and articles published before 2014 (42). Thus, this study consisted of 106 articles that focused on AI-based ECG beat classification. Figure 3 shows the detailed search method, which was compliant with the PRISMA criteria (42). In Figure 4, it can be observed that the yearly article publication trend (2014–2024) and yearly citation trend (2014–2024) for ECG beat classification indicate a strong and growing research interest in this domain. The publication trend shows a steady increase in the number of articles, which peaked in recent years (2022–2024), highlighting the expanding focus on deep learning, signal processing, and patient-specific classification approaches. This growth suggests an increasing number of researchers contributing to advancements in ECG analysis. A peak in citations often follows high publication periods, indicating that time is required for newer methods to be widely cited. These trends emphasize the need for continued innovation and the adoption of novel approaches to sustain research in ECG beat classification.

To ensure the inclusion of high-quality and peer-reviewed research, the study employed a quality-filtering criterion based on the Scimago Journal Rank (SJR) and Journal Citation Reports (JCR) quartile (Q) indexing systems. In this review, only articles published in journals indexed in SJR or those with JCR quartile rankings (Q1–Q4) were considered eligible for inclusion. Studies published in journals not indexed in either SJR or JCR, or lacking identifiable quality metrics, were classified as “Non-SJR/Q” and consequently excluded. The SJR metric reflects the scientific influence of scholarly journals by accounting for both the number of citations received and the prestige of the citing journals, while the JCR-Q system ranks journals from Q1 (highest impact) to Q4 (lowest) based on citation distributions. This quality screening ensured that the included literature represented peer-reviewed, credible, and widely recognized sources within the scientific community. Furthermore, conference papers, pilot studies, non-English articles, and inaccessible manuscripts were excluded to maintain methodological rigor and focus on reproducible, peer-reviewed work.

Figure 5 summarizes the most commonly reported performance indicators, namely, accuracy, precision, recall, and F-Score. It is acknowledged that these metrics can be misleading under severe class imbalance. Accuracy, in particular, tends to overestimate performance when normal beats dominate the dataset. Although several reviewed studies reported more robust and threshold-independent metrics, such as the Matthews correlation coefficient (MCC) and area under the precision-recall curve (AUPRC), these measures were not consistently available across all publications, preventing their inclusion in the aggregated summary plots. To maintain comparability across the 106 reviewed works, we therefore only visualized the universally reported metrics. Nevertheless, we recognize that MCC and AUPRC provide a more balanced and informative assessment of classifier performance, especially for minority arrhythmic classes such as supraventricular ectopic beat (SVEBs) and ventricular ectopic beat (VEB). This limitation highlights the need for future ECG classification studies to adopt standardized, imbalance-aware metrics to enable more equitable and clinically meaningful performance comparisons. Initially, machine learning methods relied heavily on manual feature extraction, where domain experts would identify relevant characteristics of ECG signals, such as QRS complexes, heart rate variability, and waveform shapes (29, 33, 43–46). These handcrafted features were then fed into classifiers, such as support vector machine (SVM), random forest (RF), and k-nearest neighbors (kNN), which, despite their effectiveness, were often limited by the quality and completeness of the extracted features. With the advent of deep learning, the process became more automated and data-driven, allowing models such as CNNs and RNNs to learn directly from raw ECG data (25–27). These networks captured complex temporal dependencies and subtle morphological variations without explicit feature engineering, resulting in more robust and accurate ECG beat classification systems. This shift improved classification performance and opened new possibilities for real-time and patient-specific ECG analysis.

Data imbalance is a significant challenge in ECG beat classification, particularly when using the MIT-BIH Arrhythmia (8), St. Petersburg INCART (8, 47), and MIT-BIH Supraventricular Arrhythmia (47) databases. These databases contain varied distributions of ECG beat types, often resulting in an overrepresentation of normal or common beats and an underrepresentation of rare arrhythmias, which may degrade the model performance and reduce generalizability. In the MIT-BIH Arrhythmia database, N beats vastly outnumber abnormal arrhythmias such as VEB or SVEB (8). As a result, classifiers trained on this database often achieve high accuracy by predominantly predicting the majority class (N beats), while failing to correctly identify rarer arrhythmias. This class imbalance can lead to a high false-negative rate for life-threatening conditions such as ventricular tachycardia, as the classifier may not learn sufficient patterns for these rare events. Techniques, such as the oversampling of minority classes and undersampling of the majority class, and applying synthetic data generation methods, such as the synthetic minority oversampling technique (SMOTE), can help mitigate this imbalance (54). In addition, cost-sensitive learning, where higher penalties are assigned to misclassifying rare classes, can also improve classification performance.

The INCART database, which includes recordings from 25 subjects with different types of arrhythmias, also suffers from data imbalance (8, 47). The majority of the beats in this dataset are NSR, while pathological beats such as ischaemic events are underrepresented. Similar to the MIT-BIH Arrhythmia database, this imbalance can bias classifiers towards predicting normal beats. Addressing this issue is crucial for the real-world application of ECG beat classifiers, as it ensures the model can detect critical events such as ischaemia. To address this imbalance, data augmentation techniques, such as adding noise or time-shifting the ECG signals of rare arrhythmias, can be used, as well as balancing the dataset through stratified sampling (55). The MIT-BIH Supraventricular Arrhythmia database also presents a unique challenge due to its smaller dataset size and the imbalance between normal and supraventricular arrhythmias (47). Given the relatively low number of supraventricular beats compared to normal beats, classifiers can become biased towards predicting normal rhythms, further complicating the detection of less frequent but clinically important supraventricular ectopic beats.

Bias in machine learning refers to systematic errors that cause a model to make consistently incorrect predictions, often favoring certain patterns while neglecting others (56). Bias can arise due to several factors, such as imbalanced training data, poor feature selection, or overly simplistic model assumptions that fail to capture the true complexity of the data (19). Bias in ECG beat classification is a critical issue, often stemming from imbalanced datasets where certain heartbeat types are overrepresented while others are underrepresented. This imbalance can lead to classifiers favoring the majority class, resulting in poor sensitivity for minority classes, such as supraventricular and fusion beats (57). One study demonstrated that standard classification models trained on imbalanced datasets tend to overfit the majority class while failing to recognize less frequent arrhythmias effectively (58). This issue is exacerbated by linear dependencies in ECG data, which further skew the model’s learning process and introduce classification bias (57). Moreover, the impact of inter-patient variability, where models perform well on seen patients but fail to generalize to unseen cases, further contributes to biased ECG classification (59). A detailed overview of certain issues due to bias is presented in Table 4.

To address bias in ECG beat classification, several state-of-the-art methods have been proposed (19, 58–60). One promising approach is the use of a dynamic minority-biased batch weighting loss function, which enhances the learning process for minority classes while maintaining the model’s ability to classify the majority classes accurately (59). In addition, feature fusion neural networks, which integrate multiple ECG representations, have been shown to improve classification fairness by extracting diverse feature sets that reduce bias (59). Another strategy involves differential beat accuracy (DBA), a metric that optimizes classifier performance by adjusting the learning process based on the statistical distribution of different beat types, ensuring a more balanced classification (58). Ensemble-based techniques, such as multiple-classifier architectures, have also been effective in reducing bias by leveraging diverse model outputs to correct misclassifications (60). Collectively, these approaches contribute to improving the fairness and generalization of ECG beat classification models, making them more suitable for real-world clinical applications. In addition to traditional methods for addressing imbalance, transfer learning can be a beneficial approach in this case (55). A model pre-trained on a larger, more balanced ECG dataset can be fine-tuned using the smaller MIT-BIH Supraventricular database, allowing the model to better generalize across beat types. In all these cases, proper evaluation metrics should be used to assess the classifier’s performance under imbalance conditions. Accuracy alone may not be a reliable metric, as it can be misleading in imbalanced datasets. Instead, metrics such as the F1-score, sensitivity, specificity, and AUPRC provide a more comprehensive understanding of model performance on both the majority and minority classes (61). By carefully considering these strategies and metrics, researchers can better handle the class imbalance problem in ECG beat classification, leading to more robust and clinically useful models.

Pre-processing is the initial and critical stage in ECG beat classification systems, aiming to enhance the quality of the raw ECG signals (Table 5). This stage is essential to mitigate the impact of noise and artifacts that can obscure the true physiological information in the ECG data. Pre-processing technique details and their categorization are illustrated in Figure 7. Common sources of noise include baseline wander, powerline interference (PLI), and muscle artifacts (MAs) (36). Techniques such as bandpass filtering, wavelet transforms, and normalization are employed to remove or reduce these interferences. In addition, this stage may involve the segmentation of the ECG signal into individual beats, setting the foundation for accurate feature extraction and classification in subsequent phases. First, the raw ECG signal is filtered to eliminate unwanted noise and artifacts in the preprocessing stage (35). The next step is to separate the signal into individual heartbeats, often identified by the R-peaks in the ECG. In the feature extraction phase, data from each heartbeat are parsed for various features based on its time domain, frequency domain, and morphology. Heartbeat rhythms and shapes are characterized using these properties, which can be utilized to identify certain arrhythmias. Classification involves applying machine learning methods to the retrieved attributes of individual heartbeats (90). The basic techniques for pre-processing ECG signals can be summarized as follows: classical filtering techniques, transform-based techniques, statistical and adaptive techniques, modern machine learning approaches, and advanced techniques.

Feature extraction techniques help to identify the different patterns and characteristics of the ECG signal, which can then be used to diagnose specific cardiac conditions. For example, feature extraction techniques can be used to identify abnormalities in the various segments of the acquired ECG signal, which can be indicative of different cardiac disorders. Therefore, accurate feature extraction and classification techniques are crucial for effective diagnosis and treatment. Section 4.2 reviews the existing techniques for feature extraction and classification of ECG signals in the literature.

In the ECG beat classification system, a classifier automatically classifies different heartbeats based on their ECG waveform features. Various machine and deep learning techniques have been reported for identifying different types of heartbeats (18–20, 30, 112, 114, 122–129). These classifiers utilized extracted features from the ECG signal to distinguish between different types of heartbeats. The classifier’s performance depends on the quality of the ECG signal, the feature extraction quality, and the choice of the classification algorithm. There are three different types of classification algorithms, namely, (i) unsupervised, (ii) semi-supervised, and (iii) supervised.

An overview of previous studies on the classification of ECG beats through the application of machine learning techniques is presented in Table 6. The table reviews various ECG beat classification techniques using machine learning models, highlighting their applications, databases, and performance metrics. Techniques such as radial basis function neural networks (RBFNNs), SVMs, and ensemble classifiers have been applied in datasets such as MITDB, PTBDB, and INCART, achieving accuracy rates from 84.60% to 99.99%. Each approach has specific advantages, such as resilience to noise and improved specificity, but also faces limitations, including increased system complexity or computational demands. Notably, classifiers such as SVM and RF consistently showed high performance, making them suitable for real-world applications; however, challenges in feature selection and model generalization were critical issues in different studies. After extracting the features from the ECG beats, artificial intelligence methods from machine learning can be used to build models from these data to classify arrhythmia heartbeats (43, 43–46). SVMs (126), ANNs (164), KNNs (24), and LDAs (165) are some of the most popular machine learning methods used for ECG beat classification. Despite noise and outliers in the ECG data, SVM continues to produce effective results because it finds the ideal hyperplane that maximizes the margin between distinct classes. Basic artificial neural network structures with fewer layers and parameters could improve interpretability, making it easier for clinicians to comprehend and have confidence in the model’s decisions when classifying ECG beats. The KNN algorithm is beneficial due to its simplicity and non-parametric nature, which makes it suited for small-to-medium-sized datasets and circumstances with extremely non-linear decision boundaries (24). However, when the classes in the ECG data are well-separated, LDA emerges as a linear classifier that strives to maximize the between-class variation while minimizing the within-class variance (165). Moreover, it is important to consider the dataset’s size, the task’s difficulty level, and the available processing resources before selecting an approach. Each algorithm has both advantages and disadvantages.

ECG beat classification using machine learning classification algorithms increases cardiac health assessment efficiency, accuracy, and automation. Initially, these methods offered rapid and automated ECG data analysis, saving healthcare workers time and effort. This may improve heart abnormality diagnosis and treatment outcomes and lead to lower healthcare expenditures. In addition, machine learning classifiers can efficiently process significant ECG data, making scalable and cost-effective analysis possible in clinical settings. Furthermore, these methods can understand intricate patterns and correlations within ECG readings, which enables the detection of minor anomalies that may not be evident to traditional human observers.

However, machine learning algorithms for ECG beat classification have been associated with notable challenges and limitations. These models depend on the quality and quantity of training data, which may be noisy, artifact-filled, and variable among populations. In addition, the sophisticated decision-making processes of machine learning models in ECG analysis may be difficult for clinicians to understand. This misinterpretability may limit the clinical acceptance of machine learning-based diagnostic technologies.

A comprehensive literature review on ECG beat classification using deep learning models is presented in Table 7. The review of deep learning models for ECG beat classification revealed various approaches, such as CNNs, LSTMs, and hybrid models, applied to datasets such as MITDB, INCART, and SVDB. CNNs show robust performance with accuracies up to 99.90%, while LSTM-based methods excel in capturing temporal features, improving sensitivity and specificity. The advantages of these models include the automation of feature extraction and their potential for real-time applications. However, challenges remain, such as computational intensity, model complexity, and limited generalization across diverse datasets. Ensuring robustness in noisy, real-world conditions and addressing resource-intensive training remains critical for their wider clinical adoption. Deep learning has become an essential tool in ECG beat classification, with enormous significance in medical diagnostics and healthcare (18, 26). A key advantage of deep learning over more conventional machine learning approaches for ECG beat classification is its ability to automatically derive hierarchical features from raw ECG data, eliminating the need for features to be manually extracted (18, 26). In contrast, deep learning models, specifically CNNs and RNNs, are exceptionally proficient at automatically extracting complex features from raw ECG data, which allows for the accurate classification of a wide range of cardiac disorders and arrhythmias (18, 19).

DL algorithms can improve diagnostic ability by utilizing extensive datasets, allowing them to generalize across various cardiac diseases and adjust to differences in patient demographics and recording settings (18, 26). Over the past few years, CNNs have been utilized extensively in ECG diagnosis and attained remarkable performance. CNNs are superior to other methods for extracting spatial characteristics from input signals; this allows them to detect local patterns that may indicate different cardiac problems. In contrast, RNNs can understand the temporal correlations between subsequent ECG samples since they specialize in modeling sequential data (25–27). By utilizing the combined advantages of convolutional neural networks and recurrent neural networks, particularly in hybrid structures such as CNN-RNN, ECG classification algorithms can attain exceptional performance in terms of accuracy and robustness (26–28).

DL methods have various advantages over standard ECG beat classification algorithms, including automatically learning discriminative features from raw ECG data without manual feature extraction (18). With this automated feature extraction technique, DL models can better classify cardiac diseases by detecting intricate patterns and tiny variations in the data. Further, the temporal and sequential nature of ECG signals is well-suited to DL methods, especially RNNs and CNNs, enabling them to detect the long-range relationships and temporal patterns in heartbeats (25). Furthermore, DL models have also shown excellent generalization abilities, which means they can adapt to different patient groups and recording situations, which is very important for real-world clinical applications.

In the field of ECG beat classification, deep learning algorithms have some significant limitations and challenges. Overfitting is a significant challenge in ECG beat classification, particularly in deep learning-based models, where the model captures noise and irrelevant patterns from training data, leading to poor generalization to new, unseen data. One of the primary causes of overfitting in ECG classification is the use of complex models with high-dimensional feature spaces, especially when datasets are imbalanced. Studies have shown that utilizing all 12 ECG leads in classification models can introduce redundancy and unnecessary complexity, thereby reducing generalizability and increasing the risk of overfitting (176). In addition, the presence of linear dependencies within ECG data has been identified as a major issue that skews the learning process towards the majority class, leading to biased model predictions and lower accuracy in minority class detection (57). This issue is particularly evident in patient-specific ECG classification, where models tend to overfit to the training patients, failing to generalize effectively across different individuals (29). Moreover, deep learning models trained on high-dimensional spectro-temporal ECG features have been found to suffer from overfitting due to their inability to leverage beat-to-beat variations effectively (120).

To mitigate overfitting in ECG beat classification, researchers have proposed various strategies. Feature selection and dimensionality reduction techniques, such as selecting an optimal subset of ECG leads instead of using all 12 leads, have been found to enhance classification accuracy while maintaining interpretability (176). Sample reduction techniques, such as QR and SVD, have also been effective in eliminating redundant data and addressing class imbalance, thereby reducing bias and improving generalization (57). In addition, the use of ensemble methods, such as multiple classifier architectures and hybrid machine learning approaches, has shown promise in improving classification robustness and interpretability (60, 175). Regularization techniques, including dropout and L2 regularization, are widely used to constrain deep learning models and prevent overfitting. Cost-sensitive learning approaches, such as modifying the classification threshold based on class imbalance, have also been shown to significantly enhance ECG beat classification performance by reducing bias towards majority classes (98). Furthermore, data augmentation strategies, such as adding noise, applying transformations, or generating synthetic ECG beats using generative models, have been successful in increasing the diversity of training data and improving model generalization (120). Collectively, these approaches ensure the reliability and robustness of ECG beat classification models, making them more effective for real-world medical applications.

In addition to the overfitting, large annotated datasets are also necessary for effective deep neural network training, which is a significant challenge. Generating high-quality annotated ECG data, particularly for rare cardiac disorders, can be costly and time-consuming (9). Furthermore, DL models are computationally demanding and require enormous computational resources, particularly during the training and optimization phases. This may limit their scalability and practical implementation in contexts with limited resources, such as point-of-care settings. Despite these challenges, continuous research is endeavoring to overcome these constraints by creating interpretable DL architectures, data-efficient learning methods, and hardware optimizations (24). The ultimate goal is to maximize the benefits of DL for ECG beat classification and simultaneously reduce its potential drawbacks.

Deep learning has become the leading paradigm for ECG beat classification by automatically learning spatial-temporal features from raw signals. Among various models, CNN-LSTM hybrids and attention-enhanced networks consistently demonstrate superior performance, achieving accuracies of 98%–99% across multiple ECG databases. CNNs capture morphological patterns, while LSTM and GRU layers model temporal dependencies. Transformer-based models further improve interpretability and long-range context handling. Data augmentation with GANs and representation learning via autoencoders effectively address class imbalance and noise. In addition, transfer-learning and lightweight architectures, such as MobileNet, enable real-time deployment on wearable devices. Overall, hybrid and attention-driven frameworks represent the state of the art in robust, generalizable ECG beat classification systems.

ECG beat classification using XAI leverages advanced machine learning techniques to enhance transparency in automated diagnostic systems. Traditional classification methods often function as “black boxes,” making it difficult for clinicians to interpret or trust decisions, especially in critical applications such as arrhythmia detection. Explainable AI addresses this limitation by employing methods such as saliency maps, feature importance visualization, and decision boundary analysis. These approaches allow for better understanding and verification of the features contributing to a model’s predictions. Techniques such as layer-wise relevance propagation (LRP) and gradient-weighted class activation mapping (Grad-CAM) have been applied to visualize critical regions of ECG signals, helping practitioners align model decisions with medical insights (177, 178). This not only improves diagnostic reliability but also aids in identifying potential errors in classification. Incorporating XAI into ECG beat classification aligns automated systems with real-world clinical requirements. Explainable methods enable the assessment of model predictions in handling class imbalances and morphological complexities in ECG datasets. Studies integrating XAI with neural networks, such as convolutional and recurrent architectures, have shown improved interpretability without compromising classification accuracy (58, 121). For instance, combining ICA with neural networks has demonstrated robust classification results with enhanced feature interpretability, achieving over 98% accuracy on benchmark datasets (98, 179). By bridging the gap between machine intelligence and clinician trust, explainable AI ensures that automated ECG diagnostic tools are not only effective but also transparent and clinically viable. A detailed comparison of the different features is summarized in Table 8.

XAI has emerged as a crucial frontier in ECG beat classification research, bridging the gap between high-performing deep-learning models and clinical interpretability. While deep neural networks, such as CNNs, LSTMs, and transformers, have demonstrated outstanding diagnostic accuracy, their “black-box” nature continues to limit real-world clinical trust and regulatory acceptance. Among the 106 reviewed studies, approximately 14% (15 papers) explicitly incorporated XAI methods to interpret model predictions and visualize decision reasoning. Most of these studies applied saliency-based techniques, including Grad-CAM (180) and LRP (60), to highlight the waveform segments that contribute most strongly to the classification outcomes. Such visualization maps frequently revealed correspondence between algorithmic focus regions and clinically significant components of the ECG—particularly the QRS complex, P-wave, and ST–T segments—demonstrating the potential of XAI to validate model reasoning in physiological terms.

Beyond saliency methods, a smaller number of studies utilized attention mechanisms within hybrid CNN-LSTM architectures to provide interpretability by weighting critical temporal regions. Others employed the SHapley Additive exPlanations (SHAP) (181) and Local Interpretable Model-Agnostic Explanations (LIME) to quantify feature importance in models trained on handcrafted attributes. However, the overall adoption of such explainability frameworks remains limited. Approximately 86% of the analyzed deep-learning studies focused solely on predictive performance metrics—accuracy, precision, recall, or F1-score—without providing interpretive insight into model behavior or physiological relevance. This significant gap highlights that the ECG-AI community remains primarily performance-driven, with interpretability often treated as an auxiliary consideration rather than a core design principle. The limited integration of XAI tools underscores an urgent need for standardization in interpretability reporting. Future research should emphasize quantitative explainability benchmarks, such as region relevance overlap with annotated ECG segments, and clinician-in-the-loop validation to assess whether the model explanations align with expert reasoning. Integrating XAI at the design stage can enhance clinical transparency, regulatory compliance, and physician confidence, ultimately enabling safer and more explainable deployment of AI-driven cardiac diagnostic systems.

As the traditional methods use handcrafted features, they are unable to provide reliable and adequate information from similar ECG beats due to the action of morphological patterns, which may lead to misclassification (11, 43–46).

ECG signals exhibit significant intra- and inter-patient variability in beat morphology due to differences in heart anatomy, electrode placement, noise, artifacts, and physiological factors. This variability makes it challenging for deep learning models to generalize across different individuals and accurately classify ECG beats (18, 20, 28, 116).

From the 106 selected articles, we identified consistently successful practices for ECG beat classification, particularly regarding data imbalance, generalization, model design, and evaluation (55). A key observation across numerous studies is that models trained and tested on the same patient cohort (intra-patient splits) tend to exhibit inflated performance, as they inadvertently learn subject-specific morphology (127). In contrast, inter-patient validation—where all beats from a given subject appear in only one data split—offers a more realistic assessment by ensuring generalization to unseen individuals. Therefore, inter-patient (or leave-one-subject-out) evaluation should be considered the primary reporting standard, with any intra-patient results clearly labeled as ablation or sensitivity analyses. The synthesis of reviewed literature highlights that rigorous data curation and proper dataset partitioning contribute significantly more to trustworthy and generalizable results than methodological novelty alone. These best practices address four major recurring challenges in ECG beat classification: data imbalance, architectural design, training and evaluation methodology, and reproducibility. Therefore, it is increasingly accepted that inter-patient evaluation should serve as the primary reporting protocol, with intra-patient results presented only as supplementary ablation tests. Equally important is the transparent documentation of any pre-processing, i.e., band-pass filtering (BPF) parameters, baseline-wander removal, R-peak detection methods, and beat segmentation window lengths, since small differences in these steps can significantly alter results (57). Explicitly reporting such settings is now recognized as best practice for reproducibility.

The most persistent problem identified across datasets such as MIT-BIH Arrhythmia (8), INCART (53), and MIT-BIH Supraventricular Arrhythmia (3) is the severe class imbalance between normal and abnormal beats. To mitigate this, studies have increasingly combined algorithmic and data-level solutions. Among the algorithmic solutions, loss-function re-weighting or dynamic minority-biased batch weighting consistently improved sensitivity to rare arrhythmias without harming majority class precision. These methods dynamically scale gradients according to class frequency, preventing the network from converging toward trivial majority predictions. Complementary strategies such as feature-fusion architectures, which integrate temporal, frequency-domain, and time-frequency representations, were also effective, as they expose the classifier to richer discriminative cues and mitigate bias towards dominant waveform shapes (197). A smaller but influential group of studies used DBA or F1-based early-stopping criteria instead of overall accuracy, thereby explicitly optimizing for balanced recognition across classes. Collectively, these practices demonstrate that metric-aware optimization is more faithful to clinical relevance than simple accuracy maximization. At the data level, the combination of ECG-specific augmentation and transfer learning emerged as another reliable route to improved minority-class recall. Augmentations such as controlled time-shifting, amplitude scaling, additive Gaussian noise, and mild temporal warping expand training diversity while preserving physiological plausibility. Overly aggressive distortions, such as heavy frequency modulation, were found to degrade performance by altering beat morphology. Consequently, augmentations that respect signal physiology are now favored. For smaller or domain-specific datasets, transfer learning—pre-training on a large balanced corpus and fine-tuning on the target data—consistently reduced overfitting and variance, especially for rare arrhythmias. Ensemble learning, though computationally heavier, repeatedly provided robustness against minority misclassifications by averaging the decisions of differently initialized or architected models (21). Together, these results establish a hierarchy of preferred imbalance-handling practices, namely, loss-level weighting, physiologically consistent augmentation, transfer learning, and lightweight ensembles, in order of practicality. Architectural trends in the literature show a clear evolution from manual feature engineering to end-to-end deep networks capable of automatically learning morphological and temporal patterns. Among these, 1D-CNNs remain the dominant backbone for local feature extraction because they effectively capture QRS-complex morphology and achieve high efficiency. When CNNs are extended with recurrent layers—especially LSTM or Bi-LSTM units—they gain temporal awareness, enabling them to exploit rhythm dependencies between consecutive beats. Hybrid CNN-LSTM architectures, sometimes referred to as hierarchical temporal models, consistently occupied the upper performance tier, with F-scores typically between 95% and 99% in benchmark datasets. These combinations proved particularly successful in distinguishing morphologically similar beats that differ primarily in timing rather than shape.

A further architectural enhancement repeatedly associated with improved interpretability and accuracy is the attention mechanism. Attention layers help the network emphasize diagnostically important regions of the ECG, such as the QRS onset and offset, while down-weighting redundant background segments. Across multiple studies, attention modules increased sensitivity to subtle morphological differences between ventricular and supraventricular ectopic beats, and they provided intuitive saliency maps that facilitate clinical interpretability. Where data and computational resources permitted, transformer-style multi-head attention further improved long-range context modeling, capturing dependencies across entire cardiac cycles rather than single beats. In parallel, modern convolutional refinements such as residual connections, squeeze-and-excitation blocks, and frequency-channel attention enhanced optimization stability and feature discrimination. These structural elements have therefore become standard components of competitive ECG classifiers. Beyond architecture, the review reveals consensus on several training and evaluation practices that determine whether good models generalize in clinically realistic conditions. Regularization through dropout, weight decay, and batch normalization is nearly universal among successful deep networks, preventing overfitting to small patient subsets. Early stopping based on validation F1-score or AUPRC proved more reliable than using loss reduction alone, because it guards against bias towards majority classes. Many articles also emphasize aligning model capacity with hardware constraints. For example, for wearable or edge deployment, compact 1D-CNNs with attention heads offer an attractive trade-off between interpretability, accuracy, and latency. Reporting inference time per beat or per signal window is increasingly viewed as essential, since real-time performance is critical for embedded medical systems.

The proposed ECG beat classification techniques in the literature provide promising performance, but there is a need to develop such methods to overcome the limitations of the current techniques. In this section, we discuss some of the open possible extensions of the techniques proposed in this article.

Personalized, adaptive models for individualized ECG classification: One significant future direction is the development of personalized and adaptive models that tailor the ECG classification to the individual patient’s data. Traditional models often struggle with variability in ECG signals across different patients due to factors such as age, physiology, and comorbid conditions. By incorporating patient-specific data into the training process, models can adjust their parameters to better reflect individual cardiac patterns. Transfer learning and online learning can fine-tune models using a patient’s historical ECG data, leading to more accurate and reliable diagnoses (198, 199).

The review process followed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines to ensure methodological rigor and reproducibility. Nevertheless, one inherent limitation stems from the restriction to peer-reviewed, English-language publications indexed in major databases such as PubMed, IEEE Xplore, ScienceDirect, and Springer. While this criterion ensured quality and accessibility, it potentially excluded significant research reported in non-English journals, regional repositories, theses, and conference proceedings not indexed in the selected databases. Excluding non-English studies may have introduced a language bias, especially given that ECG research is conducted globally, with active contributions from Asian, European, and South American institutions. Consequently, important innovations, particularly those published in national medical journals or non-indexed conference proceedings, may have been overlooked. This review paper provides a detailed analysis of machine learning and deep learning techniques for the classification of ECG beats, focusing on the advancements from 2014 to 2024, but it may have excluded earlier pioneering work that laid the groundwork for more advanced ECG analysis methods. A systematic approach was adopted to analyze the 106 studies, offering a comprehensive evaluation of methodologies and their applications in ECG classification. Another limitation relates to the deliberate exclusion of grey literature, including technical reports, dissertations, white papers, and non-peer-reviewed conference abstracts. While this approach enhanced the credibility of the included studies by ensuring they met peer-review standards, it may have resulted in the omission of innovative but unpublished work or early-stage algorithmic implementations. Grey literature often contains valuable methodological insights, comparative evaluations, or negative findings that are less likely to appear in journal publications due to publication bias.

A major methodological limitation of this review arises from the heterogeneity across the datasets used in the included studies. The three most frequently employed databases—MIT-BIH Arrhythmia (8), St. Petersburg INCART (53), and MIT-BIH Supraventricular Arrhythmia (3)—differ substantially in sampling frequency, signal resolution, lead configuration, and patient demographics. Consequently, performance metrics such as accuracy, sensitivity, and F1-score cannot be directly compared across studies, as the underlying data distributions vary considerably. For instance, the MIT-BIH Arrhythmia database contains 48 half-hour two-lead ECG recordings from 47 subjects sampled at 360 Hz, whereas the INCART dataset (53) includes 75 recordings at 257 Hz from 25 patients with different arrhythmia profiles. These discrepancies introduced variability in the reported results, even when identical algorithms are used. Furthermore, some studies utilized patient-specific evaluation protocols (intra-patient validation), while others employed inter-patient cross-validation strategies. The choice between these validation schemes has a profound effect on reported accuracy. Models evaluated using intra-patient splits typically yield inflated performance metrics since the test data shares morphological characteristics with the training data. Conversely, inter-patient validation better represents real-world generalization but usually yields lower accuracy. The lack of standardized evaluation protocols across the studies makes it difficult to provide an entirely uniform assessment of algorithmic performance.