Machine learning (ML) is the technique of enabling computers to mimic human learning behaviors to update their existing knowledge frame and acquire new knowledge to progressively advance their ability to complete specific tasks (22). In cardiology, ML methods have been extensively used in medical imaging [e.g., CT (23), MRI (24), chest X-ray (25), echocardiogram (26)] and ECG (27). It is especially beneficial in pre-hospital settings, where paramedics may lack the knowledge of emergency physicians or cardiologists when it comes to interpreting ECGs. ECG interpretation using ML, including traditional ML, deep learning (DL), and a combination of the two. For traditional ML methods, extra hand-crafted feature extraction and selection steps are needed. Morphological features are computed by the demarcation of major ECG characteristic points such as the QRS complex, T wave, and the J point (28). Wavelet transforms decompose ECG signals in the time-frequency domain. Principal component analysis (PCA) (29), empirical mode decomposition (EMD) (30), and the hidden Markov model (31) are all commonly employed to extract representative features. The goal of feature selection is to lower the complexity of the computational process and ensure that ML algorithms only employ the most informative features. Conventional threshold-based support vector machines (SVM), random forest (RF), naive Bayes, decision tree (DT), k-nearest neighbor (KNN), and neural network (NN) are the commonly used classifiers, and almost all these techniques have achieved good performance beyond that of cardiologists, who achieved an average accuracy of 75% for detecting ECG pathologies (32). To name a few, Kora (33) proposed a hybrid firefly (FF) and PSO (FFPSO) algorithm to optimize ECG features. Then, an ANN model named the Levenberg–Marquardt Neural Network with optimization algorithm achieved the best accuracy of 99.3% compared with the other two ML algorithms, KNN and SVM. Acharya et al. (34) investigated a four-level discrete wavelet transform to extract 12 types of non-linear features. An ANOVA analysis was used to rank these features and obtain optimal features. The performance of the proposed model was evaluated using a KNN classifier, which had an accuracy of 98.8% for MI detection and 98.73% for MI localization. Some recent conventional ML methods with good performance for MI classifications are shown in Table 2. However, these ML techniques are also confronted with some challenges, which are concluded as follows:

Deep learning (DL) is a subfield of ML. DL models can be trained on huge datasets to learn the relationship between the input features and the results automatically (47) Thus, they can directly learn features from given input of raw data without a specific step of feature extraction and have the capacity to maintain good generalizability. The hidden layers of DL models are a black box, which is responsible for automatic features learning. The deep hidden layers make a deep neural network that can potentially map to any function, which allows it to solve exceedingly complicated functions. A system with 5–20 non-linear layers may implement extraordinarily complex functions of its inputs (47). DL has been rapidly evolving and having an imperative impact on the accuracy in the classification of heart diseases. Data-driven DL models rely significantly on the quality of data and can be promoted as more data is gathered. Since 2010, due to aging populations, the availability of easy-to-use ECG monitoring devices in the form of wireless, mobile, and remote technologies have greatly expanded the capture of ECG data, and DL algorithm-based interpretation software automatically interprets ECG data. Wearable technology, wireless sensors, and deep learning techniques can all collaborate to create innovative approaches to improve healthcare services (48, 49). The improvement of AI may greatly promote the development of ECG and expand the interoperability of healthcare. For the limitations of conventional ML approaches mentioned in Section Machine Learning, the advantages of DL methods compared with the ML approaches are concluded as follows:

Based on the summary of Ansari et al. (51) we add the DL methods for MI detection and localization emerging from the recent 5 years to construct the timeline in Figure 2.

There exist other six related works that focus on automatic ECG analysis for the prediction of structural cardiac pathologies, including two systematic reviews (52, 53), one meta-analysis (54), and three comprehensive reviews (44, 51, 55). Al Hinai et al. (52) assess the evidence for DL-based analysis of resting ECGs to predict cardiac diseases such as left ventricular (LV) systolic dysfunction, myocardial hypertrophy, and ischemic heart disease. Joloudari et al. (53) focus on ML and DL techniques for myocardial infarction disease (MID) diagnosis but just cover 16 papers regarding DL methods. Grün et al. (54) include a total of five reports to provide an overview of the ability of AI to predict heart failure based on ECG signals. Attia et al. (55) discuss AI ECG algorithms for cardiac screening including LV dysfunction, silent atrial fibrillation, hypertrophic cardiomyopathy, and other structural and valvular diseases. Jothiramalingam et al. review papers that consider ECG signal pre-processing, feature extraction and selection, and classification techniques to diagnose heart disorders such as LV Hypertrophy, Bundle Branch Block, and MI. Ansari et al. (51) comprehensively evaluate several hundred publications that analyzed the ECG signal and electronic health records (EHR) to diagnose myocardial ischemia and infarction automatically and point out that DL methods have not specifically been used to detect MI and ischemia prior to 2017.

In recent 5 years, there has been an emerging number of studies focusing on state-of-the-art DL methods for MI detection and localization. Compared to the related work mentioned before, we only focus on DL techniques for MI detection and localization to consider more related papers. Covering a wide variety of studies focused on this topic in the review, we have set the following detailed goals: First, to describe and evaluate all public databases and newly collected datasets, as well as the commonly accepted assessment measures employed in the reviewed publications. Second, to assess and compare the different DL methods based on their respective performance in detecting and locating MI to find the most popular one. Finally, to discuss how these different DL methods contribute to MI detection and localization, as well as the primary potential obstacles of using DL algorithms in clinical practice for MI detection.

We search technical papers that deployed DL methods [deep neural networks (DNNs)] for MI detection and localization on Google Scholar, PubMed, and Web of Science. The following general search terms are used: (“deep learning” OR “deep neural network” OR “artificial neural network” OR “convolutional neural network” OR “CNN” OR “recurrent neural network” OR “RNN” OR “long short-term memory” OR “LSTM” OR “autoencoder” OR “deep belief network” OR “DBN” OR “generative adversarial networks” OR “GAN” OR “Restricted Boltzmann machines” OR “RBM”) AND (“electrocardiogram” OR “ECG” OR “EKG” OR “electrocardiography” OR “electrocardiograph” OR “electrocardiology”) AND (“myocardial infarction” OR “MI” OR “acute myocardial infarction” OR “AMI”). To avoid missing related papers, we also include some references that have been cited by others. Figure 3 shows the flow diagram of paper selection.

A high-quality dataset can boost the improvement of data-driven model performance and generalizability, making DL techniques possible to be used in clinical environments. A total of 13 datasets are considered in all the reviewed papers. The largest database is the ECG-ViEW II database consisting of 979,273 recordings from 461,178 patients, which was collected over a 19-year study period in South Korea. The smallest dataset is the Long-Term ST Database (LTST), only covering 86 recordings from 80 subjects, which were collected in Slovenia. Thus, these datasets differ significantly in the number of ECG recordings. Among the investigated papers, 47 (more than 79%) research trained DL models on the public Physikalisch-Technische Bundesanstalt (PTB) Diagnostic ECG database and achieved high performance for MI classification. It makes the problem of the generalizability of these existing DL models worth considering, and clinical data is needed to verify the diagnostic efficacy of such models. Other public databases such as the LTST, the PTB-XL, and the ECG-ViEW II database are alternatives for some researchers. Notably, 7 studies are using the new ECG collections from some medical institutes in recent two years. For instance, Tadesse et al. (56) collected the ECGs from 17,381 patients (11,853 MI and 5,528 Normal cases) in the Provincial Key Laboratory of Coronary Heart Disease, Guangdong Cardiovascular Institute (GCI), and three subgroups: acute, recent, and old MI were divided by cardiologists based on patients' medical history with a combination of ECGs. In 12/13 of these datasets, ECGs are collected as voltage amplitude time-series signals in one dimension, however, there is also one ECG images collection for CNN model training. Khan et al. manually collected 11,148 ECG images from Ch. Pervaiz Elahi Institute of Cardiology Multan in Pakistan and cardiologists annotated the images. All collected datasets are listed in Table 3.

Instead of relying on the complex steps of feature extraction, DL methods learn the intrinsic characteristics directly from raw or low-level processed data. Relatively little preprocessing such as resample, denoise, segmentation, and data balance are also necessary for some research. Resampling ECG data to obtain a consistent sampling rate is essential to maintain consistency. In this stage, the ECG beats are changed into the same periodic length for the DL models. ECG waveforms can be subsequently deformed by two major artifacts including high-frequency noise as well as baseline wander (BW) caused by breathing and patient movement (75). Wavelet transform methods performed well in eliminating the ubiquitous ECG noises. Daubechies wavelet 6(db6) mother wavelet basis function and Savitzky-Golay (SG) smoothing filter are commonly used to filter out high-frequency noise and correct baseline wanders. In Acharya et al. (76) and Liu et al. (77), the ECG signals with noise and without noise are both used for CNN models to compare the performance, and the results illustrated that the higher quality denoised ECG signals can improve the performance of models. However, in Zhang et al. (78), heartbeats with and without noise are adopted in contrast experiments to verify the model's robustness to noise, and the research results indicated that the experiment using ECG heartbeats with noise achieved better performance. The main reason for this anomaly is that some informative features will be missed while using wavelet to denoise since the noise domain has an unescapable overlap with the information domain. Heartbeats are segmented by distinctive points, notably fiducial R-point since R-wave has an extremely high amplitude and plainly visible peak. Pan Tompkins algorithm is the most widely used R-peak and QRS-wave detection method and modified Nagatomo's method was conducted to automatically update the algorithm of R-peak threshold in Sugimoto et al. (79). According to the position of R-peak, sliding windows are generated with the same number of heartbeats to control the size of ECG records which inputted into the model due to the length of ECG signals varies in datasets. Data imbalance occurs when the number of samples for different classes varies greatly in a classification task, and it is another common issue of influencing model performance since it almost exists in all datasets. Low-degree imbalanced data does not matter much whereas there is a great deal of highly imbalanced data in medical diagnostics of the clinical setting. Large data skew can affect the prediction results. Several approaches have been discussed in reviewed literature to best address the problem of imbalance. Rai and Chatterjee (80) proposed a Synthetic Minority Over-sampling Technique (SMOTE) to create synthetic samples for minor classes instead of copies. SMOTE is a representative algorithm of the oversampling method. Hammad et al. (81) and Dai et al. (82) added a new loss function called focal loss (FL) to address data imbalance and results in Tripathy et al. (83) indicated that FL increased MI detection accuracy by 9%. Cao et al. (84) chose Balanced Cross-Entropy, which is a modified version of the cross-entropy loss function. To forcibly strengthen the robustness of DL models, data augmentation is also used to deal with the scarce class data and to supplement the trained model with more diverse and representative data. Alghamdi et al. (85) investigated MI ECG segments with and without the augmentation technique in the proposed model, and results indicated that the model with data augmentation technique achieved better performance than that without one. Besides that, in Darmawahyuni et al. (86), researchers introduced two performance metrics Balanced Accuracy (BACC) and Matthew's Correlation Coefficient (MCC) to evaluate the performance in their proposed study which the imbalanced ratio reaches 4.57. Notably, the balanced data helps to improve the model performance, nevertheless, there is a gap between balanced distribution and real condition (87).

Furthermore, to solve the problems of amplitudes scaling and eliminating offset effect, some researchers centralized and normalized each heartbeat, respectively (43, 68, 84). Rare studies converted one-dimensional (1-D) ECG data into two-dimensional (2-D) images and treated them as computer vision tasks. In this research, signal preprocessing steps can be nearly avoided.

A plethora of studies have utilized DL algorithms to vastly improve ECG waveform-based heart diseases classification. With a focus on MI classifications, aside from binary classifiers for detection of MI and non-MI, more studies have also accomplished multiple-class classification for localizing the distinct injury parts of the myocardium. According to a content review, six main types of DL models are concluded: CNN, LSTM, CRNN, ResNet, AE, and GRU. The popularity of each type of model in MI detection and localization is illustrated in Figures 5A,B, respectively. The usage changes over the years of the six methods are illustrated in Figure 5C. The popularity of the CNN model can be reflected in terms of research purposes and annual classification models. In the following sections, each of these models will be discussed in more detail.

Each original research covered in this review yielded encouraging diagnostic performances regarding the utility of DL in ECG interpretation. The advantages and disadvantages of the six different DL methods are reviewed in Table 10. To compare the most reported performance concerning the six different DL methods, Figures 6A,B display the maximum, average, and minimum accuracies of each DL technique for MI detection and localization, respectively. The maximum accuracies of DL techniques that were trained on reviewed ECG databases and used the different number of leads are summarized in Figure 6C. It provides us reference about selecting the appropriate data leads and datasets according to the purpose of the research since some smart bands and smartwatches record single-lead ECG and multi-lead Holter monitors. It is a little difficult and arbitrary to make a comparison of model performance among studies according to a single measurement due to variability that exists between different datasets and various metrics for measurements. However, some authors have conducted studies that compare different classification algorithms in their research, and using the same dataset makes the model performance more comparable. Table 11 summarizes the only six pairs of DL models reported in the investigated paper. This is also a flaw in MI studies. Studies comparing multiple DL techniques are expected to be conducted in future works.

Wearable devices have shown their potential for risk assessment, screening, early diagnosis, and patient management in some cardiovascular diseases such as hypertension, atrial fibrillation, heart failure, peripheral vascular disease, coronary artery disease, and so on (113). The MiCORE study in Marvel et al. (114) proved wearables' value of ECG monitoring in MI prevention. Two hundred patients with type I MI are enrolled, and the wearables include an Apple Watch and a Bluetooth BP cuff. Preliminary results showed a 43% lower likelihood of the first readmission within 30 days among patients in the Corrie Health Digital Platform (Corrie) than among participants in a historical comparison group. In the investigated studies, most researchers just build the DL models for MI detection and localization by using a fixed ECG dataset and few have deployed their developed models with great performance in wearable devices. Rashid et al. (95) designed a Binary Convolutional Neural Network (BCNN) for low-power and low-memory wearable devices with a single lead sensor, the microcontroller is similar to that of the SmartCardia device. The energy consumption evaluation shows that their model achieves 8 × and 12 × energy efficiency compared to other related works, so it is energy efficient and memory-saving for wearables. The current power consumption constraints of wearables might make researchers try to compress the DL models since the memory of digital devices is limited, and models using different lead data can be used in different devices.

Cloud computing offers some help to healthcare providers, such as hospitals and clinics, who need quick access to large storage facilities and computing. With cloud computing, healthcare providers can easily share healthcare data across regions, eliminating delays in patient care to realize smart medical monitoring. Cloud computing can be integrated into a platform for interaction and data sharing, especially for remote and personalized patient care in the medical (115). In medical diagnosis, various algorithms with promising performance have been proposed in large amounts of research, but they are rarely applied to an interactive platform and non-related professional and technical personnel have trouble using these models. Recently, an easy-to-operate platform LINDA which is designed for online medical image recognition has been proposed for common users without specific knowledge in the diagnostic area (136). The contribution of the LINDA system reflects in three aspects: firstly, it uses computational intelligence to recognize abnormalities instead of knowledge-based feature extraction. Second and most importantly, an application of image processing can be shared with other professionals to obtain more shared data for the improvement of model performance. Data privacy must be considered when involves data sharing. In the LINDA system, the hash code is used to prevent parameters from passing, and the number of requests to the prediction API is limited to reduce service attacks. In addition, the traffic assessment tools are also conducted. The third contribution is that it supports creating a large image database to self-improve. Similarly, in MI diagnosis, DL techniques can also be integrated into such a cloud platform to achieve data sharing and remote diagnosis, which further works are expected to explore.

We are in the era of data explosion, the usage of smartphones, wearables, and IoT devices makes new data generate all the time. There is a lot of information that can be mined by DL algorithms in the massive data, but it also raises challenges about privacy. DL methods are data-driven methods. In the medical domain, the ethics related to patients' data that are being used for DL model prediction is of utmost importance. It makes sense to develop a solution that tries to find a balance between data utilization and privacy protection. The four stages of the life cycle model of big data in healthcare involve data collection, storage, sharing, and analysis process (137). Healthcare data subject involves individual patients, healthcare institutions, government, research institutions, and industry. Privacy leakage concerns can exist in every aspect and need to take the corresponding technical approaches to deal with. In data collection, data anonymity and differential privacy are the main technologies. Due to the size of sensors and wearables are getting smaller, lightweight, and less complex differential privacy algorithms are needed to adapt to such devices. Cloud computing is largely used to store big data in healthcare, and it is common to use encryption and auditing to ensure data confidentiality and integrity. The risks from data sharing could be effectively controlled through access control technology. In this way, novel technologies such as blockchain have been considered in encryption. In Khalid et al. (138), a decentralized authentication and access control mechanism is proposed and can apply to many scenarios. In Parah et al. (139), a novel high payload and reversible EHR embedding framework based on left data mapping (LDM), pixel repetition method (PRM), RC4 encryption, and checksum computation is proposed. In addition to technical support, policies, laws, and regulations are also necessary such as the European General Data Protection Regulation (GDPR). Hedlund et al. draw up a data sharing policy to provide guidelines for activities in the Analytic Imaging Diagnostics Arena (AIDA) (140). Privacy protection techniques in data analysis are especially important since lots of hidden information can be dug for only after data is analyzed. A research hotspot of model privacy is machine unlearning (141), and it is expected to realize model forgetting with less computational. To achieve smooth interoperability between the different parties, making data management is important.

In the medical domain, the input samples are always accompanied by various artifacts which lead to the lack of annotated data of admissible quality. However, the limited available input data in real-world practice make ML models may capture specifically these artifacts which are nothing to do with the diagnostic task (142). In terms of the problem of imperfect data quality, solutions including data augmentation techniques, transfer learning, and domain adaptation provide help. How to generate high-quality input samples instead of simply producing more training data through data augmentation has become the key issue. Recently, there are quite a few related studies, such as the classic Synthetic Minority Oversampling Technique (SMOTE), MixUp (ReMix), Hard Negative Mixing (143), and good-enough example extrapolation (144) all aim to improve the quality of resampling of small samples. The transfer learning emerges for solving the problem of insufficient annotation data, so it can be carried out under fast modeling, annotation deficiency, and small data modeling. However, the parametric transfer learning methods are time- and cost-consuming to tune, hence restricting the wide applicability in computational constraints of wearables devices. In Wang et al. (145), propose a first easy non-parametric transfer learning approach, the result of the extensibility experiment shows this easy approach can efficiently achieve good performance without requiring feature learning algorithms. Domain adaptation (DA) is an important part of transfer learning to solve the problem that data distribution between the source domain and target domain does not correspond with the independent and identically distributed (IID) condition. Gradient reversal layer (146) and generative adversarial network (GAN) (147) have been successfully applied in unsupervised domain adaptation.

In terms of input samples with artifacts mentioned before, another helpful solution is the interpretability of the models. The interpretability could enable it to notice abnormal predictions before being widely used as a diagnostic tool. The methods to achieve interpretability enable people with specific experience to have the maximum understanding of a specific model when given specific data and tasks. Based on the existing methods, we can divide the methods to achieve interpretability into three stages, that is, before, during, and after model training. Before model training, the objective of interpretability lies in data analysis, and what we're trying to do is learn about the data to the maximum extent. The methods used generally include data visualization and statistical analysis such as maximum mean discrepancy (MMD). In modeling, directly building interpretable models is a vital method. Linear regression models and specific neural networks with an explicit pooling structure are all self-explanatory models. Linear models and some of their variants have good explainability due to a very solid statistical foundation. Attention mechanisms are always adopted to provide insight into model prediction. In Samek et al. (148), four post-hoc explanation techniques have been summarized as follows: interpretable local surrogates, occlusion analysis, gradient-based techniques, and layerwise relevance propagation (LRP). A popular approach of local surrogates is the Locally Interpretable Model-agnostic Explanations (LIME) (149), which proposes to explain single predictions of any ML classifiers, and attribution method based on Shapley values which is a framework firstly proposed in game theory is a kind of occlusion analysis (150).