Electrocardiogram Classification Research Articles

Concordance analysis between deep learning model predictions and electrocardiography thresholds from cardiology guidelines

Abstract Deep learning models for the classification of electrocardiograms (ECGs) are able to learn disease-specific patterns, but they are rarely implemented in medical practice due to their "black box" nature. Post-hoc explainable artificial intelligence (XAI) methods compute regions of interest (ROI) which are of importance for a model’s decision making. However, it needs to be further analyzed whether a model focuses on the morphological or rhythmical information within the ROIs. We evaluate a pre-trained ResNet for sinus bradycardia (SB) and sinus tachycardia (ST) classification on the PTBXL dataset using the XAI method Integrated Gradients. We compare the confidence of the model predictions to ECG features used by clinicians using correlation analysis. Correlation is highest for RR intervals (SB: 0.44) and atrial as well as ventricular heart rates (ST: 0.51), with the majority exceeding clinical thresholds for both disorders, indicating that the model learned rhythmical features. Except for QT intervals in ST classification, morphological features such as duration and amplitudes of P-/T-waves do not show any correlation.

Decorrelative network architecture for robust electrocardiogram classification

Decorrelative network architecture for robust electrocardiogram classification

A deep neural network model with spectral correlation function for electrocardiogram classification and diagnosis of atrial fibrillation

Atrial Fibrillation (AF) is a common type of irregular heartbeat, and early detection can significantly improve treatment outcomes and prognoses. Single-lead Electrocardiogram (ECG) devices are under extensive scrutiny for monitoring patients' heart health worldwide. Standardized ECG signal monitoring has demonstrated a significant reduction in mortality rates associated with severe cardiovascular diseases. However, the automatic detection method for AF requires significant improvement. This study presents a novel approach that utilizes the cyclostationary analysis of ECG signals, uncovering a spectral hidden periodicity between the QRS-T (the main wave components representing electrical activity in the heart) complexes of the ECG signal through the Spectral Correlation Function (SCF). To validate the proposed method's performance, the single ECG's SCF coefficients are applied to the Convolutional Recurrent Neural Network (CRNN), which consists of convolutional and long short-term memory (LSTM) layers, on the 2017 PhysioNet challenge dataset. The obtained results demonstrate that the proposed approach efficiently represents ECG signals through SCF coefficients, leading to the accurate detection of AF with an average accuracy of 92.76% and an average F1-score of 89.1%.

Aprimoramento do diagnóstico automatizado de eletrocardiograma (ECG) por meio de pré-treinamento multimodal com laudos em texto

Objective: Heart diseases are the leading cause of death worldwide, and the electrocardiogram (ECG) is the primary diagnostic tool for assessing cardiac activity. Automated and remote ECG diagnosis can help the healthcare system with timely and high-quality cardiac assessments, especially for peripheral regions and rural areas. Automatic ECG classification has been extensively researched, but it is still challenging to build accurate models for such a wide spectrum of scenarios. Method: This study enhances the performance of ECG deep learning classification models using a multimodal pre-training stage with physician's reports. Results: Our approach improves the state-of-the-art model and achieves a mean F1 score of 0.755 over six categories using the full dataset, which is a relevant improvement for a relatively larger unlabeled corpus. Conclusion: The results demonstrate the potential to improve automated cardiac assessment with text pretraining.

Federated Learning With Deep Neural Networks: A Privacy-Preserving Approach to Enhanced ECG Classification.

In response to increasing data privacy regulations, this work examines the use of federated learning for deep residual networks to diagnose cardiac abnormalities from electrocardiogram (ECG) data. This approach allows medical institutions to collaborate without exchanging raw patient data. We utilize the publicly available data from the PhysioNet/Computing in Cardiology Challenge 2021, featuring diverse ECG databases, to compare the classification performance of three federated learning methods against both central training with data sharing and isolated training scenarios. We show that federated learning outperforms ECG classifiers trained in isolation. In particular, our findings demonstrate that a globally trained model fine-tuned to specific local datasets surpasses non-collaborative approaches. This shows that models trained in federation learn general features that can be tailored to specific tasks. Furthermore, federated learning almost matches the performance of central training with data sharing on out-of-distribution data from non-participating institutions. These results highlight the ability of federated learning in developing models that generalize well across diverse patient data, without the need to share data among institutions, thus addressing data privacy concerns.

EGCNet: a hierarchical graph convolutional neural network for improved classification of electrocardiograms

The automatic classification of electrocardiograms (ECGs) plays a crucial role in the early diagnosis of cardiovascular diseases. In recent research, deep neural network (DNN)-based methods have garnered significant attention due to their exceptional feature extraction capabilities. However, these methods face challenges in dealing with the complex inter- and intra-class dependencies inherent in different arrhythmias. We propose a classification model for electrocardiograms based on the Graph Convolutional Neural Network (EGCNet) to address this. In our EGCNet framework, we first employ the row and column 1-dimensional residual convolution to learn intra-class dependency. Then, we construct a hierarchical graph structure to model the inter-class dependency. In such a graph, each node represents one type of disease, and the edges describe the relationship between diseases. To reduce the complexity of inter-class dependency, we build a hierarchical graph structure by treating similar nodes as the same type. Thus, we can learn the discriminant dependency in the higher-level graph space and feed this high-level knowledge back to the lower-level graph space. Finally, we evaluated our method using publicly available datasets from the PhysioNet/Computing Challenge 2020 ECG data. For instance, our model achieved robust performance in the 24-label experiment with a precision of 0.753 and an F1 score of 0.722.

A MIL-based framework via contrastive instance learning and multimodal learning for long-term ECG classification

Recently, deep learning-based models are widely employed for electrocardiogram (ECG) classification. However, classifying long-term ECGs, which contain vast amounts of data, is challenging. Due to the limitation of memory with respect to the original data size, preprocessing techniques such as resizing or cropping are often applied, leading to information loss. Therefore, introducing multi-instance learning (MIL) to address long-term ECG classification problems is crucial. However, a major drawback of employing MIL is the destruction of sample integrity, which consequently hinders the interaction among instances. To tackle this challenge, we proposed a multimodal MIL neural network named CIMIL, which consists of three key components: an instance interactor, a feature fusion method based on attention mechanisms, and a multimodal contrastive instance loss. First, we designed an instance interactor to improve the interaction and keep continuity among instances. Second, we proposed a novel feature fusion method based on attention mechanisms to effectively aggregate multimodal instance features for final classification, which selects key instances within each class, not only enhances the performance of our model but also reduces the number of parameters. Third, a multimodal contrastive instance loss is proposed to enhance the model’s ability to distinguish positive and negative multimodal instances. Finally, we evaluated CIMIL on both intrapatient and interpatient patterns of two commonly used ECG datasets. The experimental results show that the proposed CIMIL outperforms existing state-of-the-art methods on long-term ECG tasks.

Evaluating the Impact of Windowing Techniques on Fourier Transform-Preprocessed Signals for Deep Learning-Based ECG Classification

(1) Background: Arrhythmias, or irregular heart rhythms, are a prevalent cardiovascular condition and are diagnosed using electrocardiogram (ECG) signals. Advances in deep learning have enabled automated analysis of these signals. However, the effectiveness of deep learning models depends greatly on the quality of signal preprocessing. This study evaluated the impact of different windowing techniques applied to Fourier transform-preprocessed ECG signals on the classification accuracy of deep learning models. (2) Methods: We applied three windowing techniques—Hamming, Hann, and Blackman—to transform ECG signals into the frequency domain. A one-dimensional convolutional neural network was employed to classify the ECG signals into five arrhythmia categories based on features extracted from each windowed signal. (3) Results: The Blackman window yielded the highest classification accuracy, with improved signal-to-noise ratio and reduced spectral leakage compared to the Hamming and Hann windows. (4) Conclusions: The choice of windowing technique significantly influences the effectiveness of deep learning models in ECG classification. Future studies should explore additional preprocessing methods and their clinical applications.

121 year after Einthoven: show me the normal ECG?

Abstract Background The clinical electrocardiogram (ECG) is measured about 120 years. The interpretations of the ECG is an art mastered by only few. Rhythm and Interval analysis is relative reliable and well supported by the available ECG machines. However, the interpretation of the ECG waveform and where it starts to deviate from the normal waveforms is poorly defined for the clinical practice. To support the anatomical relationship of the ECG, the PathECG has been developed as part of the CineECG tool, expanding the view on the ECG. In this study normal ECG distributions were created from a public available database to enable a) visual comparison of the ‘normality’ of the ECG and PathECG and b) build a predictive model to classify an ECG as normal or abnormal. Purpose To create an objective and 3D visual comparison of an ECGs to "normal ECGs". Methods From the (Physionet) PTB-XL database (Germany) all verified ECGs with known gender were selected. From the these ECGs a median ventricular beat was automatically constructed and subsequently resampled to have a standardized number of samples (Figure: Normal ECGs on top of abnormal ECGs). These resampled ECGs were also converted to their equivalent PathECG. From the normal labeled ECGs (3124 female, 2283 male) distributions were created, representing the normal ECG for male, female or the combined group. With these distributions a percentage of ECG segments (QRS, ST, T wave) can be identified as being outside this distribution, and thus as potentially abnormal. A logistic regression model was fitted to determine the optimal outlier removal per segment for the ECG waveforms and the PathECG to classify and ECG as normal or abnormal per gender group. The performance was tested on the CPSC (Hong Kong) public available ECG database. Results The abnormal ECG data in both databases contained approximately 1/3 normal ECG, atrial and ventricular arrhythmias and 1/3 ischemia and conduction disorders. The performance (AUC) of the logistic regression model on the PTB-XL and CPSC data respectively was MALES: 87.5% - 81.5%, FEMALES: 86.7% - 83.7%, and COMBINED: 86.8%-82.9%. The visual overlay of the ECG significantly supports the identification of abnormal segments (Figure: Normal ECG distribution example). Conclusions The overlay of an ECG and PathECG on the normal distribution supports the identification and verification of abnormal ECGs. This method might increase the ability of less experienced ECG readers, like general practitioners, to identify patients that need to be transferred to a cardiologist. This method could also be used for the classification of P-waves, paediatric ECGs or ECGs in specific patient groups.Normal ECGs on top of abnormal ECGsNormal ECG distribution example

A dual-branch convolutional neural network with domain-informed attention for arrhythmia classification of 12-lead electrocardiograms

The automatic classification of arrhythmia is an important task in the intelligent auxiliary diagnosis of an electrocardiogram. Its efficiency and accuracy are vital for practical deployment and applications in the medical field. For the 12-lead electrocardiogram, we know that the comprehensive utilization of lead characteristics is key to enhancing diagnostic accuracy. However, existing classification methods (1) neglect the similarities and differences between the limb lead group and the precordial lead group; (2) the commonly adopted attention mechanisms struggle to capture the domain characteristics in an electrocardiogram. To address these issues, we propose a new dual-branch convolutional neural network with domain-informed attention, which is novel in two ways. First, it adopts a dual-branch network to extract intra-group similarities and inter-group differences of limb and precordial leads. Second, it proposes a domain-informed attention mechanism to embed the critical domain knowledge of electrocardiogram, multiple RR (R wave to R wave) intervals, into coordinated attention to adaptively assign attention weights to key segments, thereby effectively capturing the characteristics of the electrocardiogram domain. Experimental results show that our method achieves an F1-score of 0.905 and a macro area under the curve of 0.936 on two widely used large-scale datasets, respectively. Compared to state-of-the-art methods, our method shows significant performance improvements with a drastic reduction in model parameters.

Source Localization and Classification of Pulmonary Valve-Originated Electrocardiograms Using Volume Conductor Modeling with Anatomical Models.

Premature ventricular contractions (PVCs) are a common arrhythmia characterized by ectopic excitations within the ventricles. Accurately estimating the ablation site using an electrocardiogram (ECG) is crucial for the initial classification of PVC origins, typically focusing on the right and left ventricular outflow tracts. However, finer classification, specifically identifying the left cusp (LC), anterior cusp (AC), and right cusp (RC), is essential for detailed preoperative planning. This study aims to improve the accuracy of cardiac waveform source estimation and classification in 27 patients with PVCs originating from the pulmonary valve. We utilized an anatomical human model and electromagnetic simulations to estimate wave source positions from 12-lead ECG data. Time-series source points were identified for each measured ECG waveform, focusing on the moment when the distance between the estimated wave source and the pulmonary valve was minimal. Computational analysis revealed that the distance between the estimated wave source and the pulmonary valve was reduced to less than 1 cm, with LC localization achieving errors under 5 mm. Additionally, 74.1% of the subjects were accurately classified into the correct origin (LC, AC, or RC), with each origin demonstrating the highest percentage of subjects corresponding to the targeted excitation origin. Our findings underscore the novel potential of this source localization method as a valuable complement to traditional waveform classification, offering enhanced diagnostic precision and improved preoperative planning for PVC ablation procedures.

ECG classification based on guided attention mechanism

Background and ObjectiveIntegrating domain knowledge into deep learning models can improve their effectiveness and increase explainability. This study aims to enhance the classification performance of electrocardiograms (ECGs) by customizing specific guided mechanisms based on the characteristics of different cardiac abnormalities. MethodsTwo novel guided attention mechanisms, Guided Spatial Attention (GSA) and CAM-based spatial guided attention mechanism (CGAM), were introduced. Different attention guidance labels were created based on clinical knowledge for four ECG abnormality classification tasks: ST change detection, premature contraction identification, Wolf-Parkinson-White syndrome (WPW) classification, and atrial fibrillation (AF) detection. The models were trained and evaluated separately for each classification task. Model explainability was quantified using Shapley values. ResultsGSA improved the F1 score of the model by 5.74%, 5%, 8.96%, and 3.91% for ST change detection, premature contraction identification, WPW classification, and AF detection, respectively. Similarly, CGAM exhibited improvements of 3.89%, 5.40%, 8.21%, and 1.80% for the respective tasks. The combined use of GSA and CGAM resulted in even higher improvements of 6.26%, 5.58%, 8.85%, and 4.03%, respectively. Moreover, when all four tasks were conducted simultaneously, a notable overall performance boost was achieved, demonstrating the broad adaptability of the proposed model. The quantified Shapley values demonstrated the effectiveness of the guided attention mechanisms in enhancing the model's explainability. ConclusionsThe guided attention mechanisms, utilizing domain knowledge, effectively directed the model's attention, leading to improved classification performance and explainability. These findings have significant implications in facilitating accurate automated ECG classification.

AutoNet-Generated Deep Layer-Wise Convex Networks for ECG Classification.

The design of neural networks typically involves trial-and-error, a time-consuming process for obtaining an optimal architecture, even for experienced researchers. Additionally, it is widely accepted that loss functions of deep neural networks are generally non-convex with respect to the parameters to be optimised. We propose the Layer-wise Convex Theorem to ensure that the loss is convex with respect to the parameters of a given layer, achieved by constraining each layer to be an overdetermined system of non-linear equations. Based on this theorem, we developed an end-to-end algorithm (the AutoNet) to automatically generate layer-wise convex networks (LCNs) for any given training set. We then demonstrate the performance of the AutoNet-generated LCNs (AutoNet-LCNs) compared to state-of-the-art models on three electrocardiogram (ECG) classification benchmark datasets, with further validation on two non-ECG benchmark datasets for more general tasks. The AutoNet-LCN was able to find networks customised for each dataset without manual fine-tuning under 2 GPU-hours, and the resulting networks outperformed the state-of-the-art models with fewer than 5% parameters on all the above five benchmark datasets. The efficiency and robustness of the AutoNet-LCN markedly reduce model discovery costs and enable efficient training of deep learning models in resource-constrained settings.

Visual interpretation of deep learning model in ECG classification: A comprehensive evaluation of feature attribution methods

Feature attribution methods can visually highlight specific input regions containing influential aspects affecting a deep learning model's prediction. Recently, the use of feature attribution methods in electrocardiogram (ECG) classification has been sharply increasing, as they assist clinicians in understanding the model's decision-making process and assessing the model's reliability. However, a careful study to identify suitable methods for ECG datasets has been lacking, leading researchers to select methods without a thorough understanding of their appropriateness. In this work, we conduct a large-scale assessment by considering eleven popular feature attribution methods across five large ECG datasets using a model based on the ResNet-18 architecture. Our experiments include both automatic evaluations and human evaluations. Annotated datasets were utilized for automatic evaluations and three cardiac experts were involved for human evaluations. We found that Guided Grad-CAM, particularly when its absolute values are utilized, achieves the best performance. When Guided Grad-CAM was utilized as the feature attribution method, cardiac experts confirmed that it can identify diagnostically relevant electrophysiological characteristics, although its effectiveness varied across the 17 different diagnoses that we have investigated.

FPGA‐Based Implementation of Real‐Time Cardiologist‐Level Arrhythmia Detection and Classification in Electrocardiograms Using Novel Deep Learning

ABSTRACTCardiac arrhythmia refers to irregular heartbeats caused by anomalies in electrical transmission in the heart muscle, and it is an important threat to cardiovascular health. Conventional monitoring and diagnosis still depend on the laborious visual examination of electrocardiogram (ECG) devices, even though ECG signals are dynamic and complex. This paper discusses the need for an automated system to assist clinicians in efficiently recognizing arrhythmias. The existing machine‐learning (ML) algorithms have extensive training cycles and require manual feature selection; to eliminate this, we present a novel deep learning (DL) architecture. Our research introduces a novel approach to ECG classification by combining the vision transformer (ViT) and the capsule network (CapsNet) into a hybrid model named ViT‐Cap. We conduct necessary preprocessing operations, including noise removal and signal‐to‐image conversion using short‐time Fourier transform (SIFT) and continuous wavelet transform (CWT) algorithms, on both normal and abnormal ECG data obtained from the MIT‐BIH database. The proposed model intelligently focuses on crucial features by leveraging global and local attention to explore spectrogram and scalogram image data. Initially, the model divides the images into smaller patches and linearly embeds each patch. Features are then extracted using a transformer encoder, followed by classification using the capsule module with feature vectors from the ViT module. Comparisons with existing conventional models show that our proposed model outperforms the original ViT and CapsNet in terms of classification accuracy for both binary and multi‐class ECG classification. The experimental findings demonstrate an accuracy of 99% on both scalogram and spectrogram images. Comparative analysis with state‐of‐the‐art methodologies confirms the superiority of our framework. Additionally, we configure a field‐programmable gate array (FPGA) to implement the proposed model for real‐time arrhythmia classification, aiming to enhance user‐friendliness and speed. Despite numerous suggestions for high‐performance FPGA accelerators in the literature, our FPGA‐based accelerator utilizes optimization of loop parallelization, FP data, and multiply accumulation (MAC) unit. Our accelerator architecture achieves a 57% reduction in processing time and utilizes fewer resources compared to a floating‐point (FlP) design.

Incidence and risk factors associated with atrioventricular block in the general population: the atherosclerosis risk in communities study and Cardiovascular Health Study

ObjectivesTo identify risk factors correlated with atrioventricular block (AVB) in the general population.MethodsParticipants in the Atherosclerosis Risk in Communities study (ARIC) and the Cardiovascular Health study (CHS) were enrolled. The presence of AVB was confirmed at an electrocardiogram (ECG) reading center using Minnesota ECG Classification. Cox proportional hazards models were performed to investigate potential risk factors of AVB, after adjustment for age, sex, race and traditional cardiovascular risk factors.ResultsDuring the 17 years of follow-up, a total of 731 high-degree AVB cases were identified. Age and sex-standardized rate of AVB was 2.79 and 2.35 per 1000 person-years in the white and the black population, respectively. With the increase of the geriatric population, the incidence of high-degree AVB will increase from 378,816 in 2020 to 535,076 in 2060, and most increment would occur among the elderly. Older age, male sex, the white race, overweight, comorbidities, declined forced vital capacity (FVC), elevated inflammation biomarkers, left bundle branch block and bifascicular block were independently associated with the incidence of high-degree AVB.ConclusionTo conclude, older age, male sex, white population, overweight, combined diabetes or chronic kidney disease, impaired FVC, elevated inflammation biomarkers, left bundle branch block and bifascicular block were independent predictors for high-degree AVB. The next 40 years would witness a dramatic increase in the incidence of high-degree AVB.

A Review of ECG Classification Techniques Based FPGA

Electrocardiogram (ECG) signals are particularly used in the diagnosis and supervision of different cardiovascular diseases. The accurate and efficient classification of ECG signals is of critical importance in clinical applications. FPGA-based systems have been identified to offer an efficient solution to the implementation of real-time ECG signal classification systems because of their parallel processing, reprogrammable nature, and low power consumption. Thus, this review paper aims to explore the existing ECG classification techniques with a focus on FPGA implementation. The review discussed several classification methods, including adaptive algorithms and neural networks, which have been optimized for arrhythmia detection. Notably, Neural network techniques like ANN, CNN, and spiking neural networks are preferable due to their ability to provide superior accuracy. For instance, Mathias et al. developed a high-precision neural network with an impressive accuracy of 99.82%. Similarly, Arona et al. deployed a DCNN on FPGA, achieving accuracies reaching up to 99.67%. Researchers widely used the MIT-BIH dataset to evaluate ECG classification techniques, This paper presents new trends and focuses on future research prospects in this field. Hence, the findings of the present review can help researchers and engineers in developing effective and efficient FPGA-based ECG monitoring and diagnostic systems for clinical and healthcare applications.

Classification of imbalanced ECGs through segmentation models and augmented by conditional diffusion model.

Electrocardiograms (ECGs) provide essential data for diagnosing arrhythmias, which can potentially cause serious health complications. Early detection through continuous monitoring is crucial for timely intervention. The Massachusetts Institute of Technology-Beth Israel Hospital (MIT-BIH) arrhythmia dataset employed for arrhythmia analysis research comprises imbalanced data. It is necessary to create a robust model independent of data imbalances to classify arrhythmias accurately. To mitigate the pronounced class imbalance in the MIT-BIH arrhythmia dataset, this study employs advanced augmentation techniques, specifically variational autoencoder (VAE) and conditional diffusion, to augment the dataset. Furthermore, accurately segmenting the continuous heartbeat dataset into individual heartbeats is crucial for confidently detecting arrhythmias. This research compared a model that employed annotation-based segmentation, utilizing R-peak labels, and a model that utilized an automated segmentation method based on a deep learning model to segment heartbeats. In our experiments, the proposed model, utilizing MobileNetV2 along with annotation-based segmentation and conditional diffusion augmentation to address minority class, demonstrated a notable 1.23% improvement in the F1 score and 1.73% in the precision, compared to the model classifying arrhythmia classes with the original imbalanced dataset. This research presents a model that accurately classifies a wide range of arrhythmias, including minority classes, moving beyond the previously limited arrhythmia classification models. It can serve as a basis for better data utilization and model performance improvement in arrhythmia diagnosis and medical service research. These achievements enhance the applicability in the medical field and contribute to improving the quality of healthcare services by providing more sophisticated and reliable diagnostic tools.

Conv-RGNN: An efficient Convolutional Residual Graph Neural Network for ECG classification

Background and objective:Electrocardiogram (ECG) analysis is crucial in diagnosing cardiovascular diseases (CVDs). It is important to consider both temporal and spatial features in ECG analysis to improve automated CVDs diagnosis. Significant progress has been made in automated CVDs diagnosis based on ECG with the continuous development of deep learning. Current most researches often treat 12-lead ECG signals as synchronous sequences in Euclidean space, focusing primarily on extracting temporal features while overlooking the spatial relationships among the 12-lead. However, the spatial distribution of 12-lead ECG electrodes can be more naturally represented using non-Euclidean data structures, which makes the relationships among leads more consistent with their intrinsic characteristics. Methods:This study proposes an innovative method, Convolutional Residual Graph Neural Network (Conv-RGNN), for ECG classification. The first step is to segment the 12-lead ECG into twelve single-lead ECG, which are then mapped to nodes in a graph that captures the relationships between the different leads through spatial connections, resulting in the 12-lead ECG graph. The graph is then used as input for Conv-RGNN. A convolutional neural network with a position attention mechanism is used to extract temporal sequence information and selectively integrate contextual information to enhance semantic features at different positions. The spatial features of the 12-lead ECG graph are extracted using the residual graph neural network. Results:The experimental results indicate that Conv-RGNN is highly competitive in two multi-label datasets and one single-label dataset, demonstrating exceptional parameter efficiency, inference speed, model performance, and robustness. Conclusion:The Conv-RGNN proposed in this paper offer a promising and feasible approach for intelligent diagnosis in resource-constrained environments.

Classification of Prehospital Electrocardiograms Performed in Ambulances According to Severity Using a Deep Learning Neural Network

Prehospital electrocardiogram (PH-ECG) transmission is an important technology for reducing door-to-balloon time, but the decision to transmit often depends on the discretion of emergency medical technicians (EMTs). Additionally, studies based on real-world data remain insufficient. This study reports a machine learning-based method for classifying the severity of PH-ECG images and explores its feasibility. PH-ECG data were compiled from 120 patients between September 2017 and September 2020. The model we created from these data was the first classification model for PH-ECG images using data from a Japanese study population and showed a weighted F1-score of 0.85 and an Area Under the Curve (AUC) of 0.93. This result can be interpreted as having an excellent balance of sensitivity and specificity. The Cohen’s Kappa coefficient between AI’s inferences and the correct labels created by two cardiologists was 0.68 (p < 0.05), which is considered “substantial” according to the guidelines presented by Landis and Koch. In this study, although we were not able to remove noise caused by patient movement or electrode detachment, the results indicate that image-based abnormality detection from PH-ECGs is feasible and effective, particularly in regions like Japan where ECG data are often stored and transmitted as images. In addition, in our region, paramedics follow a multi-step process to decide whether to transmit an ECG, which takes time for the first screening. However, if the ECG is transmitted when either the paramedics or the deep learning model detects an abnormality, it is expected to reduce reading time and door-to-balloon time, as well as decrease false negatives.