To evaluate and visualize how artificial intelligence (AI) recognizes electrocardiography (ECG) characteristics of each heart failure (HF) subtype. We developed a visually explainable and interpretable AI-ECG algorithm for various HF subtypes to meet this need. This study proposes the ShapeExplainer (SE) to interpret models. From 33,920 patients at two hospitals, we trained a convolutional neural network to identify patients with HF subtypes using ECG and clinical diagnosis. HF subtypes include heart failure with reduced ejection fraction (HFrEF), heart failure with mid-range ejection fraction (HFmrEF, and heart failure with preserved ejection fraction (HFpEF).To interprete each HF classifier model, SE, which converts non-HF group data to be recognized as each HF subtype by each classifier, was trained. The converted fake ECG shows what the change is for the classifier to recognize as each HF subtype. SE from the non-HF group of the external validation set shows the result by comparing the shape before and after transformation. When tested on an independent set of 6,784 patients, the network model yielded values for the area under the curves of HFrEF of 0.945, HFmrEF of 0.884, and HFpEF of 0.859, respectively. The results of each classifier were visualized by overlapping the original and newly generated ECGs. Comparing each subtype and its fake ECG, there was no clear trend of electrocardiographic features. However, fake ECG showed a tendency of qrs amplitude as it progressed to non-HF, HFpEF, HFmrEF, and HFrEF, especially in T wave or ST-change. The proposed fake ECG generating method is a new approach for interpreting the AI-ECG model, helping physicians understand how AI recognizes HF subtypes.

To assess whether electrocardiography (ECG)-based artificial intelligence (AI) algorithm developed to detect 1-year mortality in patients with HF with reduced ejection fraction (HFrEF). Data from two hospital visits (2016 Oct to 2021 Apr) was used. Participants with good quality baseline ECGs and HFrEF were included. We tested the hypothesis that the application of AI to the ECG could identify 1-yr mortality in HFrEF patients. A total of 16,228 (64% male, mean age of 58.5) were eligible. When tested on an independent set of 16,228 patients, the network model yielded values for the area under the curve, sensitivity, specificity, and accuracy of 0.83, 83.3%, and 64.2%, respectively. Those with a positive AI screen were at 2.6 times the risk (p<0.0001) of developing future mortality compared with those with a negative screen . We used a sensitivity map to visualize the ECG region used in the AI model to detect high 1-yr mortality in HFrEF. The map shows that the AI model focused on the QRS complex, particularly the R-wave, in most patients.

Atrial fibrillation (AF) is the most cause of cardioembolic source causing cryptogenic stroke. In these, anticoagulation therapy could reduce recurrence of stroke. However, paroxysmal AF would not be detected even by 24 hours Holter monitoring. Deep learning-based electrocardiogram (ECG) analysis models were recently developed to detect AF during sinus rhythm.We aimed to develop a deep learning algorithm (DLA) to detect AF during sinus rhythm and validate the model in patients with cryptogenic stroke who underwent implantable cardiac monitoring (ICM) to diagnose paroxysmal AF. This cohort study involved three hospitals (A, B, and C). We developed a DLA to detect AF using sinus rhythm 10 s 12-lead ECG. We included adult patients aged ≥18 years from hospital A and B. We used development data from AF adult patients who had at least one atrial fibrillation rhythm in the study period (Jan 2016 to Dec 2021) and non-AF patients who had no reference to AF in the ECG and electronic medical record. DLA was based on convolutional neural network (CNN) using 10 s 12-lead. For external validation, the ECGs from 217 patients (hospital C) with cryptogenic stroke who underwent ICM were analyzed by using the DLA for validating the accuracy in the real-world clinical situations. We included 10,605 AF adult patients and 50,522 non-AF patients as development data. During the internal validation, the area under the curve (AUC) of the final DLA based on CNN was 0.793 (95% Confidence interval 0.778–0.807). In external validation data from cryptogenic stroke patients, the mean ICM duration was 15.1 months, and AF >5 mins was detected in 32 patients (14.5%). The diagnostic accuracy of DLA was 0.793 to detect AF during sinus rhythm, and AUC was 0.824. The sensitivity, specificity, positive predictive value, and negative predictive value of the model were 0.844, 0.784, 0.403, and 0.967, respectively, which outperformed other conventional predictive methods based on clinical factors, such as CHARGE-AF, C2hest, and HATCH. In this study, DLA accurately detected paroxysmal AF using 12-leads normal sinus rhythm ECG in patients with cryptogenic stroke and outperformed the conventional models. The DLA could be used as a screening tool to identify the cause of stroke in the future.

To develop deep learning models for predicting Interoperative hypotension (IOH) using waveforms from arterial blood pressure (ABP), electrocardiogram (ECG), and electroencephalogram (EEG), and to determine whether combination ABP with EEG or CG improves model performance. Data were retrieved from VitalDB, a public data repository of vital signs taken during surgeries in 10 operating rooms at Seoul National University Hospital from January 6, 2005, to March 1, 2014. Retrospective data from 14,140 adult patients undergoing non-cardiac surgery with general anaesthesia were used. The predictive performances of models trained with different combinations of waveforms were evaluated and compared at time points at 3, 5, 10, 15 minutes before the event. The performance was calculated by area under the receiver operating characteristic (AUROC), area under the precision-recall curve (AUPRC), sensitivity and specificity. The model performance was better in the model using both ABP and EEG waveforms than in all other models at all time points (3, 5, 10, and 15 minutes before an event) Using high-fidelity ABP and EEG waveforms, the model predicted IOH with a AUROC and AUPRC of 0.935 [0.932 to 0.938] and 0.882 [0.876 to 0.887] at 5 minutes before an IOH event. The output of both ABP and EEG was more calibrated than that using other combinations or ABP alone. The results demonstrate that a predictive deep neural network can be trained using ABP, ECG, and EEG waveforms, and the combination of ABP and EEG improves model performance and calibration.

Electrocardiogram (ECG) signals are beneficial in diagnosing cardiovascular diseases, which are one of the leading causes of death. However, they are often contaminated by noise artifacts and affect the automatic and manual diagnosis process. Automatic deep learning-based examination of ECG signals can lead to inaccurate diagnosis, and manual analysis involves rejection of noisy ECG samples by clinicians, which might cost extra time. To address this limitation, we present a two-stage deep learning-based framework to automatically detect the noisy ECG samples. Through extensive experiments and analysis on two different datasets, we observe that the deep learning-based framework can detect slightly and highly noisy ECG samples effectively. We also study the transfer of the model learned on one dataset to another dataset and observe that the framework effectively detects noisy ECG samples.

In the lack of data, an appropriate inductive bias is an key solution for the successful model training. Applying the inductive bias, the knowledge distillation is one way to train the student network by predicting the output of the teacher network. Data-efficient image Transformers (DeiT) have demonstrated this effect on the ImageNet benchmark dataset. However, we observe that the performance of DeiT is degraded when evaluated in not the ImageNet dataset. Based on the in-depth analysis, we identify that DeiT fails to bring the inductive bias of its teacher network. In order to efficiently transfer the inductive bias, we proposed the block-by-block matching based on the feature-based knowledge distillation, instead of the response-based knowledge distillation. This approach is efficient when data is scarce, or the number of classes is limited. We identified that the proposed method effectively transfers the inductive bias of the teacher network to the student network. Also, the proposed approach outperforms the existing method.

We present the taxonomy of data augmentation for electrocardiogram (ECG) after reviewing various ECG augmentation methods. On the basis of the taxonomy, we demonstrate the effect of augmentation methods on the ECG classification via extensive experiments. Initially, we examine the performance trend as the magnitude of distortion increases and identify the optimal distortion magnitude. Secondly, we investigate the synergistic combinations of the transformations and identify the pairs of transformations with the greatest positive effect. Finally, based on our experimental findings, we propose an efficient augmentation policy and demonstrate that it outperforms previous augmentation policies.

Peripartum cardiomyopathy (PPCM) is a fatal maternal complication, with left ventricular systolic dysfunction (LVSD; Left ventricular ejection fraction 45% or less) occurring at the end of pregnancy or in the months following delivery. The scarcity of screening tools for PPCM leads to a delayed diagnosis and increases its mortality and morbidity. We aim to evaluate an electrocardiogram (ECG)-deep learning model (DLM) for detecting cardiomyopathy in the peripartum period. For the DLM development and internal performance test for detecting LVSD, we obtained a dataset of 122,733 ECG-echocardiography pairs from 58,530 male and female patients from two community hospitals. For the DLM external validation, this study included 271 ECG-echocardiography pairs (157 unique pregnant and postpartum period women) examined in the Ajou University Medical Center (AUMC) between January 2007 and May 2020. All included cases underwent an ECG within two weeks before or after the day of transthoracic echocardiography, which was performed within a month before delivery, or within five months after delivery. Based on the diagnostic criteria of PPCM, we analyzed the area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) to evaluate the model effectiveness. The ECG-based DLM detected PPCM with an AUROC of 0.877. Moreover, its sensitivity, specificity, PPV, and NPV for the detection of PPCM were 0.877, 0.833, 0.809, 0.352, and 0.975, respectively. An ECG-based DLM non-invasively and effectively detects cardiomyopathies occurring in the peripartum period and could be an ideal screening tool for PPCM.

We developed and validated an artificial intelligence (AI)-enabled smartwatch ECG to detect heart failure-reduced ejection fraction (HFrEF). This was a cohort study involving two hospitals (A and B). We developed the AI in two steps. First, we developed an AI model (ECGT2T) to synthesize ten-lead ECG from the asynchronized 2-lead ECG (Lead I and II). ECGT2T is a deep learning model based on a generative adversarial network, which translates source ECGs to reference ECGs by learning styles of the reference ECGs. For this, we included adult patients aged ≥18 years from hospital A with at least one digitally stored 12-lead ECG. Second, we developed an AI model to detect HFrEF using a 10 s 12-lead ECG. The AI model was based on convolutional neural network. For this, we included adult patients who underwent ECG and echocardiography within 14 days. To validate the AI, we included adult patients from hospital B who underwent two-lead smartwatch ECG and echocardiography on the same day. The AI model generates a 10 s 12-lead ECG from a two-lead smartwatch ECG using ECGT2T and detects HFrEF using the generated 12-lead ECG. We included 137,673 patients with 458,745 ECGs and 38,643 patients with 88,900 ECGs from hospital A for developing the ECGT2T and HFrEF detection models, respectively. The area under the receiver operating characteristic curve of AI for detecting HFrEF using smartwatch ECG was 0.934 (95% confidence interval 0.913-0.955) with 755 patients from hospital B. The sensitivity, specificity, positive predictive value, and negative predictive value of AI were 0.897, 0.860, 0.258, and 0.994, respectively. An AI-enabled smartwatch 2-lead ECG could detect HFrEF with reasonable performance.

In recent years, self-supervised learning methods have shown significant improvement for pre-training with unlabeled data and have proven helpful for electrocardiogram signals. However, most previous pre-training methods for electrocardiogram focused on capturing only global contextual representations. This inhibits the models from learning fruitful representation of electrocardiogram, which results in poor performance on downstream tasks. Additionally, they cannot fine-tune the model with an arbitrary set of electrocardiogram leads unless the models were pre-trained on the same set of leads. In this work, we propose an ECG pre-training method that learns both local and global contextual representations for better generalizability and performance on down-stream tasks. In addition, we propose random lead masking as an ECG-specific augmentation method to make our proposed model robust to an arbitrary set of leads. Experimental results on two downstream tasks, cardiac arrhythmia classification and patient identification, show that our proposed approach outperforms other state-of-the-art methods.