Exploring innovations in cardiac arrhythmia detection: Evaluating models, performance, and future paths

A significant portion of embolic strokes occurs without documented atrial fibrillation (AF), challenging the traditional paradigm of cardioembolism. This review addresses the emerging concept of “atrial cardiopathy” as a distinct clinical entity—an underlying atrial substrate abnormality, characterized by fibrosis and dysfunction, that promotes thromboembolism independent of AF. We posit that AF is often a late-stage manifestation of atrial cardiopathy, not the sole trigger for thrombosis. This paper synthesizes the growing evidence linking biomarkers of atrial cardiopathy to Embolic Stroke of Undetermined Source (ESUS). This new framework has profound clinical implications, suggesting a shift from arrhythmia detection to assessing atrial substrate health for stroke risk stratification. Recognizing atrial cardiopathy is fundamental for developing novel “upstream” therapies, such as targeted anticoagulation, aimed at preventing both AF and its devastating thromboembolic consequences. This review critically evaluates the evidence and translational gaps in the field, synthesizing the emerging role of advanced computational modeling as a key future tool for personalized risk stratification.

Cardiovascular diseases (CVDs) constitute a foremost global health challenge, with cardiac arrhythmias significantly increasing both mortality and morbidity. Early and precise detection of these arrhythmias from Electrocardiogram (ECG) signals is paramount but inherently complex due to the vast volume, diverse characteristics and variability of ECG data. While Deep Learning (DL) models offer transformative potential for automated ECG analysis, their widespread clinical adoption is hindered by issues such as susceptibility to overfitting, high computational demands and a notable lack of interpretability, resulting in black-box systems. This paper presents an explainable DL framework for accurate and reliable arrhythmia detection. Our innovative approach integrates advanced DL architectures, specifically Convolutional Neural Network (CNN) and Dense Neural Network (DNN), within a sophisticated multi-stage pipeline. This pipeline encompasses meticulous data preparation, state-of-the-art signal preprocessing and robust multi-strategy data balancing techniques, including ADASYN, SMOTE, SMOTETomek and Random Over-Sampling (ROS), to maximize model performance and generalization. Crucially, the framework incorporates Explainable Artificial Intelligence (XAI) methodologies-namely SHAP, LIME and Feature Importance Analysis (FIA) to provide transparent insights into the model's decision-making process. Rigorous evaluation on benchmark ECG datasets such as MITDB, PTBDB and NSTDB, demonstrates superior classification accuracy, with our ROS+CNN model achieving 99.74%, 99.43% and 99.98%, respectively. The embedded XAI components offer actionable interpretability, fostering clinical trust and paving the way for more reliable and impactful AI-driven cardiovascular diagnostics.

Editorial: Artificial intelligence for arrhythmia detection and prediction

Class imbalance poses a significant challenge in biomedical classification tasks, particularly when abnormal conditions such as arrhythmias occur infrequently. While oversampling methods like SMOTE, ADASYN, and Active SMOTE attempt to alleviate this by generating synthetic minority samples, they often overlook model uncertainty during sampling. In this paper, we propose an enhanced Active SMOTE framework that integrates a lightweight uncertainty-aware module. The module measures prediction confidence through softmax probabilities, identifies the most ambiguous minority-class instancesthose with predicted probabilities close to 0.5and prioritises them for synthetic augmentation. To generate new samples, a k-nearest-neighbour interpolation mechanism is applied, producing diverse yet informative synthetic data near decision boundaries. This design strengthens the classifiers ability to learn from critical borderline cases and reduces wasted computation on confidently classified samples. We evaluate the method on two biomedical datasets with 12 features: a large-scale ECG dataset (80,000 samples) and a smaller Gas Sensor Drift dataset (~13,000 samples). Each dataset is processed in a five-stage incremental learning setup, simulating gradual data arrival as in real-world biomedical systems. Across both datasets, our uncertainty-aware strategy consistently outperforms traditional methods (SMOTE, ADASYN, Active SMOTE) in F1-score and recall, with particularly strong gains in early learning stages when data is scarce. The approach is efficient, interpretable, and easily integrable with existing classifiers, offering a practical and deployable improvement for biomedical applications such as arrhythmia detection or sensor drift monitoring.

Background: Abdominal fetal electrocardiography (fECG) is an emerging noninvasive tool for fetal monitoring. A key challenge is separating the weak fetal signal from overlapping maternal ECG and external noise. Recent advances in machine learning (ML), particularly in deep learning, hold promise for enhancing signal extraction and detecting subtle fetal cardiac issues, such as arrhythmias. These advances could standardize diagnosis and support early intervention, necessitating a structured review of current practices. Hypothesis: We hypothesize that advanced AI techniques and intensive learning models significantly improve the extraction and interpretation of abdominal fECG signals. These models are expected to outperform traditional signal processing in detecting key fetal cardiac features and abnormalities, including arrhythmias and congenital heart disease (CHD), thereby improving diagnostic accuracy and clinical decision-making. Aims: To review and synthesize studies utilizing AI methods for fECG signal processing, with a focus on diagnostic performance in fetal monitoring, arrhythmia detection, and CHD diagnosis. Methods: A systematic search of PubMed, Scopus, Web of Science, and Cochrane databases was conducted for studies published from 2015 to January 2025. Included studies applied AI techniques to fECG data in prenatal evaluation. Data were extracted using a standardized form. Results: Sixty-two studies were included, with sample sizes ranging from 5 to 757 (median: 68) and gestational ages from 21 to 42 weeks. The most commonly used preprocessing methods are bandpass filtering and resampling. Accuracy ranged from 71% to 100%, F1-score from 70.15% to 100%, precision from 75.33% to 100%, and sensitivity from 63% to 100%. CNN-BiLSTM achieved perfect performance across all metrics. Other models (CNN, DenseNet, W-NETR, U-Net) showed near-perfect results, especially for QRS and R-peak detection. For arrhythmia detection, accuracy reached 98.56%, with one CNN model achieving 100% specificity and sensitivity. CHD detection studies have shown accuracies of 71–95%, with the top models achieving 94% across precision, sensitivity, and F1-score. Hybrid and deep learning models consistently outperformed simpler approaches. Conclusion: AI-driven, intensive learning shows strong potential for improving fECG signal quality and diagnostic accuracy. While the results are promising, further clinical validation is necessary for routine implementation in prenatal care.

Background: Long-term continuous ambulatory cardiac monitoring (LTCM) is a widely used diagnostic tool for arrhythmia detection, outperforming other modalities. COVID-19 accelerated adoption of home enrollment (HE) for LTCM, which includes mailing devices to patients for self-application and activation, highlighting the need for patient-centered solutions that optimize usability and comfort. HE was recently made available for a next-generation LTCM, which is smaller and lighter than prior designs, with a breathable adhesive, and has demonstrated superior performance. Aims: We assessed wear compliance and ECG signal quality for next generation LTCM devices applied in-clinic by a technician vs. HE. Additionally, we evaluated the impact of a smartphone app on wear compliance and ECG quality. Methods: U.S. adults prescribed the Zio Monitor (iRhythm Technologies, San Francisco, CA) for 14 days between December 2, 2024 - March 16, 2025, were included, corresponding to the initial availability of HE for Zio Monitor. Outcomes compared between in-clinic and HE devices included mean wear time, mean analyzable time (% free from artifact), early wear terminations (≤ 2 days), and actionable arrhythmia yield. Additional analyses evaluated outcomes among patients opting to use a smartphone app (MyZio), which provides onboarding, digitized instructions, and reminders for wear and return, vs. those who did not. Results: Of 304,735 LTCM devices worn, 276,142 (91.6%) were applied in-clinic and 28,593 (9.4%) were HE. Mean age was 61.5±17.9 years; 56.0% were female. App use was higher in the HE group (54% vs 17%, p < 0.0001). Mean wear time and % analyzable time were high and comparable for in-clinic and HE. Early wear terminations were infrequent in both groups and arrhythmia yield was comparable. App use was associated with lower % of early wear terminations and greater analyzable time in both groups (Table). Among prescribed devices, return compliance (activated, worn and returned ≤ 45 days) was higher in app users for both in-clinic (96.0% vs. 93.2%) and HE (90.4% vs. 71.1%) devices. Conclusion: Wear compliance and percent analyzable time for a next-generation LTCM were high and comparable when applied in-clinic by a technician vs. HE, indicating that HE achieves comparable arrhythmia detection while eliminating in-clinic visits and reducing provider burden. Patient apps as medical device adjuncts may further improve enrollment and compliance with home-based or ambulatory diagnostics.

Background: In the DRAI MARTINI study, the DeepRhythmAI (DRAI) algorithm had a superior sensitivity for arrhythmia detection compared to ECG technicians. However, it has not been reported whether the additional arrhythmias that were detected were directly relevant to the indication for ambulatory monitoring or were incidental. Methods: We included n=14,606 patients with 14±10 days of continuous ambulatory ECG that was analysed beat-to-beat by both DRAI and ECG technicians (n=167). In patients monitored for tachyarrhythmia (known or suspected atrial fibrillation (AF), palpitations, or transient ischemic attack or stroke) AF, supraventricular tachycardias (SVTs), ectopic atrial rhythm (EAR), ventricular tachycardia (VT) or idioventricular rhythm (IVR) were considered relevant findings, while 2 nd or 3 rd degree atrioventricular block (AVB) and pauses/asystoles >2.0/3.5s incidental. In patients monitored for bradycardia (syncope or dizziness) any finding of AVB, pause/asystole, VT, and IVR were considered relevant, whereas EAR, SVT and AF incidental. DRAI and technicians were compared to annotations by a panel of three experts, and confidence intervals (CIs) were derived using bootstrapping with 1,000 replications. Results: The sensitivity for both monitoring indications (tachyarrhythmia and bradycardia) was superior for DRAI compared to technicians. In patients monitored for tachyarrhythmia, the sensitivity for relevant arrhythmias was 99.5% (95%CI 98.8-100.0%) for AI vs. 67.9% (95%CI 62.9-71.8%) for technicians. The corresponding rates of arrhythmia detection were 221/1,000 patient-recordings (95%CI 209-234) for AI vs. 142/1000 patient-recordings (95%CI 131-153) for technicians. In patients monitored for bradycardia, the sensitivity for relevant arrhythmias was 99.4% for DRAI vs 54.6% (95%CI 45.3-61.9%) for technicians, and the corresponding rates of arrhythmia detection 115/1,000 patient-recordings (95%CI 104-126) vs 64/1,000 patient-recordings (95%CI 57-71). DRAI also had higher sensitivity for incidental findings, 98.5% (95%CI 96.3-100%) vs 49.2% (95%CI 41.3-56.2%) in patients monitored for tachyarrhythmias and 99.7% (95%CI 99.3-99.9%) vs 77.2% (95%CI 65.8-86.5%) in patients monitored for bradycardia. True positive rates for relevant and incidental findings are shown in Figure 1a-d. Conclusion: Analysis with DRAI has a superior sensitivity compared to technicians both for arrhythmias relevant to the monitoring indication and for incidental findings.

Electrocardiogram (ECG) analysis is one of the most essential tools for identifying cardiac rhythm abnormalities and assessing cardiovascular health [1], [2]. Accurate interpretation of ECG signals is crucial for detecting disorders such as arrhythmias, atrial fibrillation, and other irregular heart activities that can lead to severe medical complications [3], [4]. However, traditional diagnostic methods heavily rely on manual inspection by clinicians, which is not only labor-intensive but also susceptible to subjective bias and diagnostic inconsistencies [5]. To address these limitations, this research presents an intelligent and automated ECG signal analysis framework utilizing Deep Recurrent Neural Networks (DRNNs) for reliable classification of heart rhythm disorders [6], [7]. The proposed approach leverages the temporal learning capabilities of recurrent architectures such as Long ShortTerm Memory (LSTM) and Gated Recurrent Unit (GRU) networks to model the complex time-dependent characteristics inherent in ECG sequences [8], [9]. By doing so, the system effectively captures both short-term and long-term dependencies in heartbeat patterns, enabling more precise identification of abnormal cardiac activity [10]. Prior to model training, several signal preprocessing techniques—including denoising, baseline drift removal, segmentation, and feature scaling—are applied to enhance signal quality and reduce artifacts caused by noise or motion interference [11], [12]. Subsequently, a feature extraction phase converts raw ECG data into time– frequency representations to enrich the model’s understanding of rhythm morphology and improve classification accuracy [13], [14]. The proposed model is trained and validated using publicly available benchmark datasets such as MIT-BIH Arrhythmia and PTB Diagnostic ECG Database [15], [16], ensuring generalizability across diverse cardiac conditions and patient profiles. Comprehensive performance evaluation reveals that the DRNN-based system significantly outperforms traditional machine learning algorithms—including Support Vector Machines (SVMs), Random Forests, and Decision Trees—achieving superior results in accuracy, sensitivity, and specificity [17], [18]. Moreover, the system demonstrates robust generalization on unseen ECG data, confirming its potential for real-world deployment in both clinical environments and wearable health monitoring devices [19], [20]. Overall, this study presents a data-driven, intelligent ECG analysis solution capable of real-time arrhythmia detection and continuous cardiac monitoring [21], [22]. The proposed framework not only enhances diagnostic precision but also supports physicians in clinical decision-making, ultimately contributing to faster, more reliable, and accessible cardiac healthcare systems. With further refinement, integration with IoT-based platforms and edge computing technologies could transform this framework into a cornerstone of next-generation AI-assisted cardiovascular care [23]–[25].

Introduction: Several ECG parameters accessible via ambulatory Holter recordings may help risk-stratify coronary heart disease (CHD) risk, but their clinical use is mostly limited to arrhythmia detection. Candidate ECG-based measures for CHD include autonomic markers like heart rate variability (HRV) and cardiac repolarization measures involving the T wave and QT interval. We explored several potential ECG digital biomarkers for CHD classification in a cohort of veteran twins to help inform future research on real-time CHD risk assessment with ambulatory ECG monitoring. Methods: We collected ambulatory Holter data (1,000 Hz, 3 channels) for one week from veteran twins recruited from the Vietnam Era Twins Registry. We assessed CHD if they described either previous MI, CABG, cardiac arrest/ICD shock, or PCI during a clinical history conducted by a trained clinician. We used the Physionet Cardiovascular Signal Toolbox to derive several ECG features from sliding 5-minute windows (table 1). For each ECG feature, we calculated several different statistical summary measures. This resulted in 164 total features. We performed logistic regression with forward selection and corrected for multiple comparisons with Bonferroni correction. We also tested the performance of the newly discovered ECG markers within twin CHD discordant pairs to control for family-level confounding. Lastly, we controlled for traditional CHD risk factors using multivariable models. Results: We studied 514 men with mean (SD) age 73.4 (2.9) years; 7% were Black, and 23% had a history of CHD (table 2). The mean QT interval and T amplitude in lead I were the most strongly associated with CHD (p<0.0001 for both), and none of the other markers met criteria for statistical significance. The ECG-CHD relationships were similar within 33 twin pairs discordant for CHD (AUC 0.69). Adjustment for traditional risk factors did not reduce the observed ECG-CHD associations. The QT interval (427 ms) was longest in the cardiac arrest/ICD shock subgroup, while the T-amplitude (0.054 mV) was lowest in the subgroup with CABG history (table 3). Discussion: In this cross-sectional study of older veteran twins, mean QT interval and T amplitude (lead I) were the only two significant Holter parameters associated with CHD after correction for multiple testing. These results can help to inform future studies of CHD risk with Holter monitors and ecological momentary assessments of stress and behavioral exposures.

Background: Detection of very low amplitude ("fine") ventricular fibrillation (VF) remains a challenge for implantable cardioverter-defibrillators (ICDs), as such signals may fall below programmed sensitivity thresholds. To address this, Abbott ICDs incorporate the VF Therapy Assurance (VFTA) algorithm, which adjusts detection criteria when low-amplitude signals are sensed, aiming to reduce delayed therapy for hemodynamically unstable arrhythmias. However, this case illustrates a novel pitfall: inappropriate ICD shock triggered by misclassification of supraventricular tachycardia (SVT) as fine VF due to far-field signal distortion and VFTA activation. Case: A 53-year-old male with heart failure with recovered ejection fraction (LVEF 45–50%) from non-ischemic cardiomyopathy, paroxysmal atrial fibrillation, and an Abbott CRT-D, experienced an unexpected ICD shock while performing light housework. He had no preceding symptoms. Device interrogation revealed a regular tachycardia at 160 bpm with low-amplitude ventricular signals on the far-field Coil-Can vector. These were misclassified by the VFTA algorithm as fine VF. Once triggered, VFTA collapsed all detection zones into a single therapy zone and suppressed supraventricular discriminators. Only six ventricular beats labeled “F,” each with a cycle length faster than 400 ms, were required to fulfill VF detection criteria, and a shock was delivered. SecureSense markers showed low-amplitude R-waves annotated as “VS2,” especially before “F” annotations (Figure 1). After identifying the mechanism, the sensing vector was reprogrammed from Coil-Can to Tip-Can, improving signal fidelity and preventing recurrence. Discussion: This case highlights how VFTA, though designed to enhance VF detection, can override SVT discriminators and result in inappropriate shocks in the setting of low-amplitude far-field signals. The algorithm modifies detection criteria in response to signal dropout, but in patients with intrinsically low amplitudes, especially on far-field vectors, this may paradoxically increase misclassification risk. Reprogramming the sensing vector proved a simple and effective solution. As ICD algorithms become more sophisticated, individualized vector selection and device programming remain critical to ensuring accurate arrhythmia detection.

Cardiovascular disease is a leading cause of maternal morbidity and mortality. As cardio-obstetric care evolves, digital health technologies including telehealth, wearable devices and remote monitoring are playing an increasingly critical role. These innovations have the potential to enhance cardiovascular screening, risk assessment and disease management across the perinatal continuum. This article explores the role of digital health technologies in advancing cardio-obstetrics care. It focuses on the impact of telehealth, wearable sensors and emerging digital tools on improving maternal cardiovascular outcomes and reducing disparities in care delivery. A narrative review approach was used to synthesize existing literature and clinical insights related to telecardiology, wearable monitoring and digital innovations. Emphasis was placed on applications in pregnancy and post-partum care, as well as on evaluating implementation challenges and equity concerns. Digital health tools improve access to care and facilitate early diagnosis of cardiovascular conditions. Telehealth increases care continuity and patient satisfaction, especially in underserved populations. Wearable devices equipped with photoplethysmography and electrocardiogram sensors enable intermittent, non-invasive monitoring, aiding early detection of arrhythmias and hypertensive disorders. Furthermore, novel technologies such as digital twins, natural language processing and virtual reality show potential to personalize care and support medical education. Despite these advancements, key barriers persist, including data privacy concerns, unequal access to technology and algorithmic bias. Digital health technologies are transforming cardio-obstetrics care by enabling proactive management and expanding access. However, for these tools to deliver equitable benefits, targeted efforts are needed to address privacy, infrastructure and literacy challenges. Future efforts should focus on integrating digital health into routine maternal care, promoting digital literacy, ensuring equitable technology access and improving interoperability with electronic health records. Additionally, ongoing evaluation through clinical trials in high-risk pregnancies and ethical safeguards to mitigate algorithmic bias will be essential to ensure safe, scalable and inclusive implementation of these innovations in maternal cardiovascular care.

Trends in cardiovascular mortality in Hodgkin and non-Hodgkin lymphoma: A nationwide analysis from 1999 to 2020

Graph-enhanced deep learning for ECG arrhythmia detection: An integration of CNN-GNN-BiLSTM approach.

HLS-compiled PYNQ-based cardiac arrhythmia detection system leveraging quantized ECG beat images

It was sought to compare the effectiveness of two methods 7-day ECG Holter or 14-day event-Holter monitoring in detection of arrythmias and ischemia in a heart failure (HF) population far from academic centers treated by a primary care physician under cardiologist supervision. In the prospective, non-randomized, 3-month pilot program carried out between June and December 2023 recruited were 429 HF patients from villages and small cities in 14 primary care units, far from academic centers. Of them, 124 (28.9%) patients were additionally monitored by either 7-day ECG Holter (7H-group) or 14-day event Holter (14eH-group). The cumulative percentage of patients with non-sustained ventricular tachycardia, new atrial fibrillation or ischemic changes was a primary composite endpoint. Of 126 patients, 54 (43.5%) were monitored by 7-day ECG Holter while 70 (56.5%) by 14-day event Holter. At baseline, there were no significant differences between 7H- vs. 14eH-group in terms of demographics and cardiovascular risk factors. A history of PCI was more frequent in 7H- vs. 14eH-group (33 vs. 15%, p = 0.039). The cumulative percentage of the primary composite endpoint was significantly higher in 7H- vs. 14eH-group (24 vs. 2.9%, p < 0.001) and was driven by silent ischemia. The number of therapeutic interventions, including introduction of an oral anticoagulant or coronary angioplasty was numerically higher in 7H- vs. 14eH-group (11.1 vs. 4.3%, p = 0.27). In this pilot study, 7-day ECG Holter was more effective in detection of non-sustained ventricular tachycardia, new atrial fibrillation or ischemic changes than 14-day event-Holter in HF patients in social exclusion regions.

Background/Objectives: Current guidelines recommend 24-48 h Holter for risk stratification and atrial fibrillation (AF) screening in hypertrophic cardiomyopathy (HCM). However, the limited duration of this approach may not provide optimal sensitivity. In addition, extended ECG monitoring has been demonstrated to be more effective in detecting arrhythmias in other clinical entities. We aimed to assess the utility of extended ECG monitoring for 30 days in a non-high-risk cohort of HCM patients. Methods: We conducted a prospective multicentre study with 113 non-high-risk HCM patients who underwent 30-day ECG monitoring with a dedicated device. We compared the detection of relevant arrhythmias (AF, atrial flutter, and non-sustained ventricular tachycardia) during 30-day monitoring with the findings observed during the first 24 h. Results: Extended ECG monitoring detected relevant arrhythmias in 63.7% of patients, compared with 12.4% during the first 24 h (p < 0.001). This difference was mainly driven by non-sustained ventricular tachycardia (NSVT) (61.1% vs. 8.9%, p < 0.001). Atrial fibrillation episodes were detected in 10.6% of patients after completing prolonged monitoring vs. 6.2% during the first 24 h (p = 0.066). Extended monitoring resulted in a reclassification of 21.2% of patients to a higher sudden cardiac death (SCD) risk category using the HCM-SCD calculator. Conclusions: Extended ECG monitoring significantly enhances the detection of arrhythmias in HCM. Using this technique, NSVT were detected in most patients of a non-high-risk HCM cohort. Further investigation is warranted to determine the role of extended monitoring in SCD risk stratification and AF screening.

Bruton tyrosine kinase inhibitors (BTKis) have revolutionized treatment for chronic lymphocytic leukemia (CLL), but cardiovascular (CV) toxicities pose significant challenges. Second-generation BTKis offer improved target specificity, yet CV risks persist. This expert opinion review evaluates current evidence and offers guidance for managing BTKi-associated CV events, particularly in patients with comorbidities. A hematologist, specialist in CLL, and a cardiologist with specific expertise in managing cardiac complications arising from oncologic therapies conducted a systematic literature review. Subsequently, clinical findings synthesized through multidisciplinary expert discussions led to the development of practical recommendations for CV risk stratification and management in patients with CLL receiving BTKi therapy. Regular CV monitoring is essential for early detection of atrial fibrillation (AF), hypertension, and ventricular arrhythmias. A multidisciplinary approach between hematologists and cardiologists is recommended for comprehensive care. Identifying hematological biomarkers for predicting cardiotoxicity and exploring cardioprotective therapies for high-risk patients should be prioritized. It is essential to educate healthcare providers about the CV risks associated with BTKis. Additionally, clinical guidelines should be regularly updated to reflect the latest evidence, ensuring effective prevention and management strategies. In patients with a heightened risk of CV complications, the use of second-generation BTKis should be prioritized. Ongoing cardiovascular monitoring is also recommended to reduce the risk of adverse events and minimize treatment discontinuation.

BackgroundTimely and accurate detection of arrhythmias from electrocardiograms (ECGs) is crucial for improving patient outcomes. While artificial intelligence (AI)-based ECG classification has shown promising results, limited transparency and interpretability often impede clinical adoption.MethodsWe present ECG-XPLAIM, a novel deep learning model dedicated to ECG classification that employs a one-dimensional inception-style convolutional architecture to capture local waveform features (e.g., waves and intervals) and global rhythm patterns. To enhance interpretability, we integrate Grad-CAM visualization, highlighting key waveform segments that drive the model's predictions. ECG-XPLAIM was trained on the MIMIC-IV dataset and externally validated on PTB-XL for multiple arrhythmias, including atrial fibrillation (AFib), sinus tachycardia (STach), conduction disturbances (RBBB, LBBB, LAFB), long QT (LQT), Wolff-Parkinson-White (WPW) pattern, and paced rhythm detection. We evaluated performance using sensitivity, specificity, and area under the receiver operating characteristic curve (AUROC), and benchmarked against a simplified convolutional neural network, a two-layer gated recurrent unit (GRU), and an external, pre-trained, ResNet-based model.ResultsInternally (MIMIC-IV), ECG-XPLAIM achieved high diagnostic performance (sensitivity, specificity, AUROC > 0.9) across most tasks. External evaluation (PTB-XL) confirmed generalizability, with metric values exceeding 0.95 for AFib and STach. For conduction disturbances, macro-averaged sensitivity reached 0.90, specificity 0.95, and AUROC 0.98. Performance for LQT, WPW, and pacing rhythm detection was 0.691/0.864/0.878, 0.773/0.973/0.895, and 0.96/0.988/0.993 (sensitivity/specificity/AUROC), respectively. Compared to baseline models, ECG-XPLAIM offered superior performance across most tests, and improved sensitivity over the external ResNet-based model, albeit at the cost of specificity. Grad-CAM revealed physiologically relevant ECG segments influencing predictions and highlighted patterns of potential misclassification.ConclusionECG-XPLAIM combines high diagnostic performance with interpretability, addressing a key limitation in AI-driven ECG analysis. The open-source release of ECG-XPLAIM's architecture and pre-trained weights encourages broader adoption, external validation, and further refinement for diverse clinical applications.

Real-world evaluation of arrhythmias detection in hemodialysis patients using wearable electrocardiogram monitor: a two-phase multicenter observational study