Generative Counterfactual Explainable Artificial Intelligence (XAI) offers a novel approach to understanding how AI models interpret electrocardiograms (ECGs). Traditional explanation methods focus on highlighting important ECG segments but often fail to clarify why these segments matter or how their alteration affects model predictions. In contrast, the proposed framework explores “what-if” scenarios, generating counterfactual ECGs that increase or decrease a model’s predictive values. This approach has the potential to increase clinicians’ trust specific changes—such as increased T wave amplitude or PR interval prolongation—influence the model’s decisions. Through a series of validation experiments, the framework demonstrates its ability to produce counterfactual ECGs that closely align with established clinical knowledge, including characteristic alterations associated with potassium imbalances and atrial fibrillation. By clearly visualizing how incremental modifications in ECG morphology and rhythm affect artificial intelligence-applied ECG (AI-ECG) predictions, this generative counterfactual method moves beyond static attribution maps and has the potential to increase clinicians’ trust in AI-ECG systems. As a result, this approach offers a promising path toward enhancing the explainability and clinical reliability of AI-based tools for cardiovascular diagnostics.

We hypothesized that AI-enabled ECG scores would show distinct temporal patterns after hospital discharge in patients with HF, and that these patterns would differ between patients who experienced rehospitalization and those who did not. This single-center retrospective study analyzed ECG data from patients hospitalized for HF between March 2017 and January 2025 in South Korea. Post-discharge, ECGs were processed using AI-ECG models for left ventricular systolic dysfunction (LVSD), diastolic dysfunction (LVDD), and myocardial infarction (MI). We compared AI-ECG patterns in patients readmitted within six months vs. those not (Figure 1). Temporal trends in AI-ECG scores were assessed using a mixed-effects linear regression model with group and time as fixed effects, and patient as a random effect. Among 1,007 patients, 1,539 hospitalization events were identified. A total of 1,674 ECGs from 269 rehospitalized and 4,066 ECGs from 917 non-rehospitalized patients were collected from 180 days before to 60 days after the index readmission or follow-up end. The mean age was 65.2 years, and 63.1% were male. Diabetes mellitus and chronic kidney disease were significantly more prevalent in the rehospitalization group, whereas other comorbidities were comparable. Significant differences in ECG intervals and axes were also observed, with no notable difference in heart rate. In the LVSD model, rehospitalized patients showed higher scores overall (β = 7.96, 95% CI: 3.18–12.75, p = 0.001) (Figure 2). Time since discharge was associated with decreasing scores (β = –0.096/day, 95% CI: –0.104 to –0.087, p<0.001), but this decline was attenuated in the rehospitalization group (interaction β = 0.092, 95% CI: 0.069–0.115, p<0.001). The LVDD model demonstrated a similar trend, while the MI model exhibited no statistically significant differences in scores (Figure 3). AI-ECG models show potential as dynamic biomarkers for detecting early physiological deterioration and predicting readmission risk in HF patients. These findings support their use in future patient monitoring strategies.

Our team previously developed an AI-ECG method for diagnosing ST-segment elevation myocardial infarction (STEMI) and non-ST-segment elevation myocardial infarction (NSTEMI) using 12-lead electrocardiograms (ECGs), demonstrating superior performance compared to cardiologists (Sci Rep 10, 20495 [2020]). In 2023, this approach was approved as an innovative technology in South Korea (AiTAMI v1.00.00). External validation was conducted across 18 emergency centers (ROMIAE study). Building on these findings, we introduce the “10-minute rule” for early risk assessment of acute myocardial infarction (AMI). We trained AiTAMI v2.00.00 using a foundation model and ECG data from the ROMIAE cohort collected across 14 hospitals. The model was validated at four additional centers, comprising 1,480 patients (Non-AMI = 1,150; NSTEMI = 198; STEMI = 132). Model performance and risk stratification were evaluated using AUROC, clinical endpoints, and decision rule performance. The updated model improved AUROC from 0.887 to 0.906 and AUPRC from 0.760 to 0.795. The 10-minute rule-out strategy identified 23.2% of patients with a negative predictive value (NPV) of 99.7%, while the rule-in strategy identified 24.4% of patients with a positive predictive value (PPV) of 68.5%. AI-ECG utilizing the 10-minute rule can classify 47.6% of chest pain patients early in emergency settings, indicating a potential paradigm shift in the management of AMI.

Accurate electrocardiogram (ECG) delineation is essential for automated cardiac diagnosis, enabling the precise identification of key waveforms such as the P wave, QRS complex, and T wave. This study presents the first comprehensive benchmarking of neural network-based semantic segmentation models for ECG delineation, evaluating their accuracy, resource efficiency, and robustness across both public and private datasets. Our results demonstrate that convolutional neural network (CNN)-based approaches consistently achieve superior accuracy compared to Transformer-based approaches. Additionally, we observed the presence of fragmented segments in the delineation results. To address this issue, we explored post-processing techniques to consolidate or eliminate fragmented segments using an optimal configuration, leading to performance improvements. Furthermore, by analyzing performance variations across different waveform labels, we provide critical insights into key considerations for ECG segmentation tasks. Notably, our findings also reveal that larger model sizes do not necessarily correlate with better performance. Based on our findings, we propose a set of practical guidelines for leveraging segmentation models in ECG delineation, offering valuable direction for future research and clinical applications.

Test-Time Adaptation (TTA) methods have been widely used to enhance model robustness by continuously updating pre-trained models with unlabeled target data. However, in real-world biosignal applications-where factors such as age, lifestyle, and comorbidities induce significant variability—traditional TTA often falls short in addressing personalization needs. To satisfy such needs, we introduce a novel Test-Time Calibration (TTC) framework that integrates continuous self supervised adaptation on unlabeled samples with periodic supervised calibration using the sporadically available ground-truth labels. Our approach leverages a model equipped with dual heads for supervised learning (SL) and self-supervised learning (SSL), and further incorporates a dual buffer along with a weighted batch sampling strategy to effectively manage and utilize both data types during the test phase. We evaluate our framework on two distinct datasets: the publicly available PulseDB, a benchmark for cuff-less blood pressure estimation, and a private ICU dataset collected from critically ill patients. Experimental results demonstrate that our approach improves blood pressure prediction accuracy and robustness, highlighting its suitability for dynamic, personalized biosignal applications.

We present a new perspective on scaling neural networks for electrocardiograms (ECG). Although ResNet-based models are widely used in ECG classification, the potential benefits of network scaling remain unexplored. Our research investigates the impact of changes in the depth of layers, the number of channels, and the dimensions of the convolution kernels on performance. Contrary to computer vision practices, we found that shallower networks, with more channels and smaller kernels, lead to better performance for ECG classifications. Based on these findings, we provide insights that can guide the efficient development of models in practice. Finally, we explore why scaling hyperparameters affects ECG and computer vision differently. Our findings suggest that the inherent periodicity of the ECG signals plays a crucial role in this difference.

Online Test-Time Adaptation (OTTA) enhances model robustness by updating pretrained models with unlabeled data during testing. In healthcare, OTTA is vital for realtime tasks like predicting blood pressure from biosignals, which demand continuous adaptation. We introduce a new test-time scenario with streams of unlabeled samples and occasional labeled samples. Our framework combines supervised and self-supervised learning, employing a dual-queue buffer and weighted batch sampling to balance data types. Experiments show improved accuracy and adaptability under real-world conditions.

The real-world effectiveness of the artificial intelligence model based on electrocardiogram (AI-ECG) signals from portable devices for detection of left ventricular systolic dysfunction (LVSD) requires further exploration. In this prospective, single-centre study, we assessed the diagnostic performance of AI-ECG for detecting LVSD using a six-lead hand-held portable device (AliveCor KardiaMobile 6L). We retrained the AI-ECG model, previously validated with 12-lead ECG, to interpret the 6-lead ECG inputs. Patients aged 19 years or older underwent six-lead ECG recording during transthoracic echocardiography. The primary outcome was the area under the receiver operating characteristic curve (AUROC) for detecting LVSD, defined as an ejection fraction below 40%. Of the 1716 patients recruited prospectively, 1635 were included for the final analysis (mean age 60.6 years, 50% male), among whom 163 had LVSD on echocardiography. The AI-ECG model based on the six-lead portable device demonstrated an AUROC of 0.924 [95% confidence interval (CI) 0.903–0.944], with 83.4% sensitivity (95% CI 77.8–89.0%) and 88.7% specificity (95% CI 87.1–90.4%). Of the 1079 patients evaluated using the AI-ECG model based on the conventional 12-lead ECG, the AUROC was 0.962 (95% CI 0.947–0.977), with 90.1% sensitivity (95% CI 85.0–95.2%) and 91.1% specificity (95% CI 89.3–92.9%). The AI-ECG model constructed with the six-lead hand-held portable ECG device effectively identifies LVSD, demonstrating comparable accuracy to that of the conventional 12-lead ECG. This highlights the potential of hand-held portable ECG devices leveraged with AI as efficient tools for early LVSD screening.

Despite the proliferation of heart failure (HF) mortality prediction models, their practical utility is limited. Addressing this, we utilized a significant dataset to develop and validate a deep learning artificial intelligence (AI) model for predicting one-year mortality in heart failure with reduced ejection fraction (HFrEF) patients. The study’s focus was to assess the effectiveness of an AI algorithm, trained on an extensive collection of ECG data, in predicting one-year mortality in HFrEF patients.

Emerging evidence supports artificial intelligence–enhanced electrocardiogram (AI-ECG) for detecting acute myocardial infarction (AMI), but real-world validation is needed. The aim of this study was to evaluate the performance of AI-ECG in detecting AMI in the emergency department (ED). The Rule-Out acute Myocardial Infarction using Artificial intelligence Electrocardiogram analysis (ROMIAE) study is a prospective cohort study conducted in the Republic of Korea from March 2022 to October 2023, involving 18 university-level teaching hospitals. Adult patients presenting to the ED within 24 h of symptom onset concerning for AMI were assessed. Exposure included AI-ECG score, HEART score, GRACE 2.0 score, high-sensitivity troponin level, and Physician AMI score. The primary outcome was diagnosis of AMI during index admission, and the secondary outcome was 30 day major adverse cardiovascular event (MACE). The study population comprised 8493 adults, of whom 1586 (18.6%) were diagnosed with AMI. The area under the receiver operating characteristic curve for AI-ECG was 0.878 (95% CI, 0.868–0.888), comparable with the HEART score (0.877; 95% CI, 0.869–0.886) and superior to the GRACE 2.0 score, high-sensitivity troponin level, and Physician AMI score. For predicting 30 day MACE, AI-ECG (area under the receiver operating characteristic, 0.866; 95% CI, 0.856–0.877) performed comparably with the HEART score (0.858; 95% CI, 0.848–0.868). The integration of the AI-ECG improved risk stratification and AMI discrimination, with a net reclassification improvement of 19.6% (95% CI, 17.38–21.89) and a C-index of 0.926 (95% CI, 0.919–0.933), compared with the HEART score alone. In this multicentre prospective study, the AI-ECG demonstrated diagnostic accuracy and predictive power for AMI and 30 day MACE, which was similar to or better than that of traditional risk stratification methods and ED physicians.