Development of a semi–real-time electrocardiogram monitoring system integrating artificial intelligence and wearable devices for atrial fibrillation screening

Abstract

Background

Atrial fibrillation (AF) is a common arrhythmia associated with substantial morbidity, particularly ischemic stroke. Conventional AF screening using rule-based algorithms is limited by suboptimal accuracy and the need for extensive clinician review when processing large volumes of electrocardiogram (ECG) data. Artificial intelligence (AI) has emerged as a promising approach to improve detection efficiency and scalability.

Objectives

To develop and internally evaluate a deep learning model for AF screening using real-world 24-hour Holter ECG data, and to evaluate a semi–real-time monitoring workflow integrating AI and wearable ECG devices in high-risk patients.

Methods

This retrospective study included 1,489 Holter ECG recordings collected at Nguyen Trai Hospital. A Residual Network (ResNet) model was trained on 1,089 recordings and evaluated on an independent dataset of 400 recordings using case-level classification with patient-level data splitting. A semi–real-time screening workflow integrating AI analysis and wearable ECG devices was prospectively implemented in a high-risk cohort.

Results

From the training dataset, a total of 29,765 minutes of cardiologist-annotated AF episodes were identified across 82 AF-positive recordings. These episode-level annotations were subsequently mapped to fixed 60-second segments for model training using a predefined sampling strategy. On the independent evaluation dataset (n = 400; AF prevalence 6.3%), the model achieved 100.0% sensitivity, 88.0% specificity, and 35.7% positive predictive value for case-level AF detection. In the pilot cohort (n = 167), AF was detected in 24 patients (14.4%), including 20 (12.0%) identified after 24 hours of monitoring.

Conclusions

The proposed framework demonstrated high sensitivity for AF detection in real-world Holter ECG data. Integration with wearable devices in a semi–real-time clinician-in-the-loop workflow was feasible and may support AF screening in high-risk populations.

Keywords

atrial fibrillation artificial intelligence ambulatory ECG monitoring wearable multi-use device semi–real-time protocol

1. Introduction

Atrial fibrillation (AF) is one of the most prevalent cardiac arrhythmias and is strongly associated with an increased risk of thromboembolic events, particularly ischemic stroke. Timely diagnosis and appropriate anticoagulation therapy substantially reduce stroke risk, whereas delayed or missed detection exposes patients to preventable cerebrovascular events. Conversely, overdiagnosis may lead to unnecessary anticoagulation and associated bleeding risks.^1,2 Recent guidelines have therefore expanded indications for AF screening, particularly among patients with prior ischemic stroke or transient ischemic attack.³

The diagnosis of AF is typically established using 12-lead electrocardiography (ECG); however, paroxysmal or subclinical AF episodes are frequently missed in routine clinical evaluation. Ambulatory ECG monitoring, including 24-hour Holter monitoring, is widely used to improve detection yield.⁴ Nevertheless, traditional Holter analysis relies on rule-based algorithms with limited accuracy in noisy or complex recordings, often requiring time-intensive expert review.⁵

Recent advances in deep learning have enabled artificial intelligence (AI) models to automatically extract complex temporal and morphological ECG features, offering improved accuracy compared with conventional approaches. These models provide a scalable solution for AF screening and rhythm interpretation in large datasets.⁶

In this context, the present study aimed to (1) develop and internally evaluate a deep learning–based AI model for AF detection using real-world 24-hour Holter ECG data, and (2) evaluate a semi–real-time monitoring framework integrating AI analysis with wearable ECG devices for screening in high-risk populations.

2. Methods

2.1. Study objectives

This study had three main objectives: (1) to develop a deep learning–based classifier for atrial fibrillation (AF) detection using Holter ECG data; (2) to evaluate the diagnostic performance of the model at the case (Holter recording) level using real-world datasets; and (3) to implement and assess the feasibility of a semi–real-time AF monitoring workflow integrating wearable ECG devices and AI analysis in high-risk patients.

The study was conducted at Nguyen Trai Hospital, Ho Chi Minh City, Vietnam, between March 2024 and September 2025. All data were anonymized prior to analysis in accordance with institutional ethical standards.

2.2. AI model development

2.2.1. Population and dataset

A total of 1,489 24-hour Holter ECG recordings were retrospectively collected during the study period. Of these, 135 recordings were excluded from demographic analysis because of incomplete demographic information. All 1,489 recordings were retained for model development and evaluation (Figure 1).

Figure 1.

Dataset flow diagram showing inclusion, exclusion, and allocation of Holter ECG recordings for demographic analysis and model development.

2.2.2. Data splitting

The dataset was split at the patient (subject) level to prevent data leakage. All recordings were assigned exclusively to either the training dataset (n = 1,089) or the evaluation dataset (n = 400), with no overlap. Each patient contributed only one Holter recording.

ECG signals were preprocessed using a bandpass filter of 0.5–30 Hz before segmentation. No notch filtering, resampling, normalization, or feature scaling was applied. ECG segments were generated only after dataset splitting, ensuring that no segment from the same recording appeared in both datasets.

Within the training dataset, a patient-level validation subset was used for hyperparameter tuning, model selection, and early stopping, and remained fully disjoint from the evaluation dataset. No cross-validation was performed.

2.2.3. Annotation and labeling

Two independent cardiologists reviewed all recordings and annotated AF episodes at the episode level. Inter-rater agreement between the two cardiologists was assessed prior to consensus adjudication using Cohen’s kappa coefficient, which was κ = 0.92. Continuous AF episodes were labeled from onset to termination, whereas intermittent AF was annotated as separate discrete episodes. Transition points were defined based on the earliest visually identifiable rhythm change.

In this study, annotated rhythm strips referred to cardiologist-identified continuous rhythm intervals or episodes with variable duration, rather than fixed-length model input segments. The reported AF duration therefore reflects cumulative annotated AF time across recordings. These episode-level annotations were subsequently mapped to fixed 60-second segments for model training.

The non-AF category included sinus rhythm, sinus tachycardia, sinus bradycardia, atrial tachycardia, supraventricular ectopy (PAC), and ventricular ectopy (PVC). No atrial flutter cases were present in the dataset. Frequent PACs or PVCs with irregularity were classified as non-AF when preserved P-wave morphology or consistent ectopic patterns were observed. Recordings with implanted pacemakers were excluded. Recordings with variable signal quality were retained for model development; potential noise-related misclassification was mitigated through clinician verification.

2.2.4. Segmentation and aggregation

After patient-level splitting, ECG recordings were segmented into 60-second windows. In the training dataset, overlapping segmentation was first applied to AF episodes using a fixed stride smaller than 60 seconds, whereas non-AF signals were segmented using non-overlapping windows (stride = 60 seconds). Segment-level sampling was then applied, retaining all AF-positive segments and randomly sampling non-AF segments to achieve an approximately balanced class distribution.

This sampling was applied only in the training dataset and does not reflect dataset-level AF prevalence. In the evaluation dataset, non-overlapping segmentation (stride = 60 seconds) was applied uniformly to avoid redundancy and ensure unbiased performance estimation.

Each segment contained 12,000 samples (200 Hz × 60 seconds). A segment was defined as AF-positive if it contained ≥6 consecutive seconds of confirmed AF. The model generated 60 probability outputs per segment (one per second). A segment was classified as AF-positive if ≥6 consecutive seconds had predicted AF probability ≥0.5. At the case level, a Holter recording was classified as AF-positive if at least one segment met the AF-positive criterion.

2.2.5. Model architecture and training

A Residual Network (ResNet) architecture was used to model temporal ECG features (Figure 2). Input segments consisted of 60s ECG signals sampled at 200 Hz. A single ECG lead was selected from the available channels and used as input to the model for both training and inference. Convolutional layers with progressively increasing filters (16–112) were used to extract hierarchical features, followed by batch normalization, ReLU activation, and dropout layers.

Figure 2.

Residual network (ResNet) architecture.

The classification threshold was determined using the training/validation dataset with prioritization of sensitivity for screening purposes and fixed prior to evaluation. External datasets (MIT-BIH and AFDB) were used for technical benchmarking only and were not involved in model training, hyperparameter tuning, or evaluation on the primary dataset.^7,8

The model was trained using the Adam optimizer (learning rate 1×10^-4, batch size 32) with binary cross-entropy loss. Early stopping was applied based on validation loss, and the best-performing model was selected according to the lowest validation loss. No learning rate scheduler was used. The number of training epochs was not fixed a priori and was determined by the early stopping criterion. The model architecture and training parameters are summarized in Table 1.

Table 1.

Model architecture, training, and evaluation framework.

Parameter	Value/description
Model architecture	Residual Network (ResNet)
Input signal	Single-channel ECG
Signal preprocessing	Bandpass filtering 0.5–30 Hz; no notch filtering, resampling, normalization, or feature scaling
Sampling frequency	200 Hz
Segment length	60 seconds
Samples per segment	12,000
Segment overlap	Overlapping windows applied to AF segments in the training dataset; no overlap in the evaluation dataset
Stride	Stride 60 seconds for non-overlapping windows; smaller stride used for overlapping AF segments during training
Model output	60 probability outputs per segment (one per second)
Ground truth labeling	Episode-level annotation by cardiologists
Segment-level AF definition	Segment contains ≥6 seconds of cardiologist-confirmed AF
Segment-level prediction rule	AF-positive if ≥6 consecutive seconds have predicted AF probability ≥0.5
Case-level aggregation	Holter recording AF-positive if ≥1 AF-positive segment
Optimizer	Adam
Learning rate	1 × 10^-4
Batch size	32
Loss function	Binary cross-entropy
Early stopping	Based on validation loss
Model selection	Best-performing model selected according to validation loss
Class imbalance handling	Segment-level sampling applied to the training dataset only
Data split	Patient-level split
Validation strategy	Validation subset derived from training data (patient-level separation)
External datasets	MIT-BIH and AFDB used for technical benchmarking only
Threshold selection	Determined using training/validation data and fixed prior to evaluation
Clinical role	Decision-support system; all outputs reviewed and verified by cardiologists

2.3. AI performance evaluation

The evaluation was designed to estimate diagnostic performance, particularly sensitivity, at the case (Holter recording) level. No formal a priori sample size calculation was performed, and the evaluation dataset size was determined based on available real-world data.

All diagnostic performance metrics were calculated at the case level following aggregation of segment-level predictions. Sensitivity, specificity, positive predictive value, and negative predictive value were reported with 95% confidence intervals using the Wilson score method. AUROC and PR-AUC were reported descriptively.

2.4. Clinical application: Semi–real-time AF screening

2.4.1. Design of the pilot study

Following model development, a semi–real-time atrial fibrillation (AF) screening protocol was deployed in a cohort of high-risk patients without previously diagnosed AF using wearable multi-use 7-day ambulatory ECG devices. The high-risk population was defined as individuals meeting at least one of the following criteria: (1) history of ischemic stroke, (2) age ≥ 75 years, or (3) CHA₂DS₂-VA score ≥ 2 points.² Monitoring duration was determined by clinicians based on individual clinical judgment.

The primary endpoint was feasibility of the AI-assisted screening workflow, defined as successful completion of ECG acquisition, data upload, automated analysis, and clinician verification. Secondary endpoints included AF detection yield, monitoring duration, timing of AF detection, and AF burden (cumulative duration of AF episodes).

2.4.2. Wearable device

A wearable multi-use ECG device, OCTOBEAT (OCTOMED Co., Ltd., Vietnam), was used to support ambulatory rhythm monitoring (Figure 3). The device enables up to 7 days of continuous recording using a 3-lead configuration and supports Bluetooth-based data transmission to a mobile application with subsequent upload to a hospital server. The system provides basic device status monitoring (e.g., battery level and electrode contact) and allows patient-triggered symptom annotation. The device complies with IEC 60601 safety standards and was used solely for data acquisition in this study; all clinical decisions were made by cardiologists.

Figure 3.

Wearable multi-use device and patch electrodes.

2.4.3. Operational definition of semi–real-time workflow

In this study, “semi–real-time” refers to automated AI-based analysis performed after ECG data upload, with clinician review completed within 24 hours. Accordingly, the expected detection-to-review latency is within 24 hours from data upload.

This daily review cadence was selected to balance detection timeliness with clinical workflow feasibility in a screening context. No continuous real-time alerting was implemented. Instead, AI-detected AF events were communicated to clinicians at a predefined daily review time, representing a scheduled alerting workflow. No automated alert performance metrics (e.g., alert frequency or response time) were systematically recorded in this study.

2.4.4. AI-integrated screening workflow

ECG data were transmitted to a central server via wireless communication, preprocessed using the same bandpass filtering pipeline, segmented into 60-second non-overlapping windows, and input into the trained AI model for AF probability inference.

All AI-generated outputs were reviewed by cardiologists before clinical decision-making. The clinician-in-the-loop framework supports scalable ECG analysis with cardiologist oversight in real-world clinical settings. The AI-integrated screening workflow is illustrated in Figure 4 and includes segment prioritization, cardiologist confirmation, and subsequent clinical action following AF detection. This study was reported in accordance with TRIPOD-AI principles to ensure transparency and reproducibility (Appendix).

Figure 4.

AI-integrated semi–real-time atrial fibrillation screening workflow.

3. Results

A total of 1,489 24-hour Holter ECG recordings were included, with 1,089 assigned to the training dataset and 400 to the evaluation dataset. A total of 135 recordings with missing demographic information were excluded from descriptive analyses.

3.1. Demographics of study population

The demographic characteristics of the study population are summarized in Table 2. The study population was predominantly older and female. There were no significant differences in age or sex distribution between the training and evaluation cohorts.

Table 2.

Demographics of study population.

	Overall^* (N=1354)	Training (N=992)	Evaluation (N=362)	p-value
Age (years)	67 (58–75)	68 (59–76)	66 (57–74)	0.050
Age ≥ 75	355 (26.2%)	271 (27.3%)	84 (23.2%)	0.146
Female sex	843 (62.3%)	605 (61.0%)	238 (65.7%)	0.125

*A total of 135 cases with missing demographic information were excluded.

The overall prevalence of AF across the full dataset (n = 1,489) was 7.2% and did not differ significantly between the training and evaluation datasets (Table 3).

Table 3.

Atrial fibrillation prevalence.

	Overall (N=1489)	Training (N=1089)	Evaluation (N=400)	p-value
Atrial fibrillation	107 (7.2%)	82 (7.5%)	25 (6.3%)	> 0.05

3.2. AI model training

From the 1,089 Holter ECG recordings in the training dataset, cardiologist-annotated rhythm strips were identified and classified into AF and non-AF categories. A total of 3,218 AF rhythm strips were annotated across 82 AF-positive recordings, corresponding to a cumulative AF duration of 29,765 minutes (Table 4). Annotated rhythm strips were mapped to fixed 60-second segments using a sliding window approach. Non-AF rhythm intervals were more fragmented because annotation focused on rhythm transitions and clinically relevant intervals rather than continuous labeling of all non-AF periods.

Table 4.

Annotated rhythm intervals and cumulative duration in the training dataset.

Rhythm type	Number of cases	Cumulative annotated duration (minutes)	Number of annotated rhythm strips
Atrial fibrillation	82	29,765	3,218
Non-AF total	1,007	8,112	2,631
– Sinus rhythm	1,023	5,600	1,867
– Tachycardia	246	1,885	505
– Bradycardia	146	627	259

Annotated rhythm strips are variable-duration rhythm intervals identified by cardiologists.

AF: Atrial fibrillation.

The model was additionally benchmarked on two external ECG datasets (MIT-BIH and AFDB). Performance was assessed using segment-based and duration-based metrics derived from standardized PhysioNet WFDB evaluation tools. These results are presented solely for technical benchmarking and not as indicators of clinical diagnostic performance (Table 5).

Table 5.

External benchmark performance on MIT-BIH and AFDB datasets using segment-based and duration-based metrics.

Dataset	ESe	E+P	DSe	D+P
MIT-BIH	84.0%	72.0%	95.0%	87.0%
AFDB	90.0%	95.0%	82.0%	99.0%

ESe: sensitivity, segment-based; E+P: precision, segment-based; DSe: sensitivity, duration-based; D+P: precision, duration-based.

3.3. AI model performance

3.3.1. Evaluation dataset

Independent Holter ECG recordings were used for case-level evaluation. In the final evaluation dataset, 25 AF-positive cases were observed among 400 recordings (6.3%).

3.3.2. Diagnostic performance evaluation

All performance metrics were calculated at the case level following aggregation of segment-level predictions. The confusion matrix is presented in Table 6.

Table 6.

Confusion matrix and diagnostic performance on the evaluation dataset (case-level).

	Actual AF	Actual non-AF
Predicted AF	25	45
Predicted non-AF	0	330
Sensitivity	100.0% (95% CI: 86.7–100.0%)
Specificity	88.0% (95% CI: 84.4–90.9%)
PPV	35.7% (95% CI: 25.8–47.0%)
NPV	100.0% (95% CI: 98.9–100.0%)
AUROC	0.95
PR-AUC	0.56

CI: confidence interval, PPV: positive predictive value, NPV: negative predictive value, AUROC: area under the receiver operating characteristic curve, PR-AUC: precision–recall AUC.

3.4. Clinical implementation

A pilot implementation of the AI-assisted screening workflow was conducted in 179 high-risk patients. The AI-assisted workflow was successfully completed in 167 of 179 patients (93.3%), including ECG acquisition, data upload, automated analysis, and clinician verification. The remaining 12 patients were excluded due to early device removal, insufficient recording duration, or inadequate signal quality.

AF was detected in 24 patients (14.4%). Of these, 4 (2.4%) were identified within the first 24 hours, whereas 20 (12.0%) were detected after 24 hours. The median cumulative AF duration was 528 minutes (interquartile range: 186–666 minutes) (Table 7).

Table 7.

Clinical application of the semi–real-time AF screening protocol.

	N = 167
Age (years)	65 (54–77)
Female	75 (44.9%)
Screening time (hours)	65 (43–107)
AF detected	24 (14.4%)
– Within 24 hours	4 (2.4%)
– After 24 hours	20 (12.0%)
AF duration (minutes)	528 (186–666)

Patients with detected AF were allowed to discontinue monitoring early and were referred for clinical evaluation. Oral anticoagulation therapy was initiated when clinically indicated, given the high-risk baseline characteristics of the cohort.

4. Discussion

In this study, we developed and evaluated a deep learning–based model for atrial fibrillation (AF) detection using real-world 24-hour Holter ECG data and implemented a semi–real-time screening workflow integrating wearable ECG devices and AI analysis. The model achieved high sensitivity (100.0%) and moderate specificity (88.0%) at the case level. The high sensitivity reflects the case-level aggregation rule (≥1 AF-positive segment) combined with a threshold prioritizing detection of AF episodes, consistent with prior AI-based AF detection studies.^9,10 In addition, the pilot implementation in a high-risk cohort demonstrated the feasibility of integrating AI-assisted ECG analysis into routine clinical workflows.

The relatively low positive predictive value observed in this study is primarily attributable to the low AF prevalence in the evaluation dataset (6.3%), which is consistent with a screening context in a low-prevalence population. In addition, the sensitivity-prioritized classification strategy may contribute to an increased false-positive rate. In real-world screening implementation, this may increase the number of AI-flagged recordings requiring clinician review and potentially contribute to additional workload. Further refinement of classification thresholds and post-processing strategies may help improve precision while maintaining adequate screening sensitivity.

Conventional ECG interpretation remains dependent on clinician expertise and is subject to interobserver variability. AI-based ECG analysis enables scalable review of large volumes of ambulatory recordings, reducing reliance on continuous manual interpretation.¹¹ The AI system was designed as a decision-support tool rather than an autonomous diagnostic system, with all outputs reviewed by cardiologists prior to clinical decision-making.

The semi–real-time workflow implemented in this study enables structured batch analysis of ECG data with clinician review within 24 hours, providing a practical balance between timely detection and clinical workload constraints. By prioritizing segments with high predicted AF probability, the system facilitates efficient identification of clinically relevant AF episodes.¹²

Continuous ECG monitoring is essential for detecting paroxysmal and asymptomatic AF. While conventional 24-hour Holter monitoring provides high-quality multi-lead recordings, it is limited by short monitoring duration.¹³ Extended monitoring using single-use patch devices improves detection yield but may be associated with higher cost, limited reusability, and reduced signal diversity.¹⁴ In contrast, the reusable multi-lead wearable ECG device used in this study supports extended monitoring and flexible ambulatory use. This approach may facilitate repeated or longitudinal rhythm monitoring in high-risk populations, where intermittent AF episodes are often missed with short-duration recordings.

The pilot implementation further demonstrated the feasibility of this approach in a real-world high-risk cohort. A high completion rate (93.3%) was achieved, and AF was detected in 14.4% of patients, with the majority of cases identified after 24 hours of monitoring. These findings underscore the importance of extended monitoring duration for detecting intermittent AF. The workflow endpoint was defined at the system level, including successful ECG acquisition, data transmission, automated analysis, and clinician verification. The ability to discontinue monitoring early after AF detection represents an additional practical advantage, enabling timely clinical intervention while optimizing resource utilization. Patients with detected AF were referred for clinical evaluation, and subsequent management, including initiation of oral anticoagulation therapy, was determined based on individual clinical assessment.

This study has several important limitations. First, it was conducted at a single center using retrospective data for model development, which may limit generalizability. External validation in independent and multicenter datasets is required. Second, although sensitivity was high, specificity could be further improved through refinement of classification thresholds or post-processing strategies. Third, workflow-level metrics, including alert frequency, false-positive alert rate, clinician review time, and time-to-notification, were not systematically recorded and should be evaluated in future prospective studies. Fourth, calibration performance was not assessed and warrants further investigation using calibration curves and related metrics. Fifth, no formal signal quality validation against a reference standard (e.g., 12-lead ECG) was performed. Sixth, no dedicated model interpretability analysis (e.g., saliency maps or attention mechanisms) was conducted, which may limit understanding of model decision processes. Finally, potential failure modes include misclassification in rhythms with irregular RR intervals (e.g., atrial tachycardia, frequent ectopy) and in low-quality recordings; these risks were partially mitigated through expert annotation and clinician verification.

5. Conclusions

This study demonstrates the feasibility of integrating deep learning–based ECG analysis with wearable monitoring devices in a semi–real-time clinician-in-the-loop workflow for atrial fibrillation screening. Using real-world Holter ECG data, the proposed framework achieved high sensitivity for AF detection while enabling structured prioritization and review of ambulatory ECG recordings. In a high-risk pilot cohort, AI-assisted wearable ECG monitoring was successfully incorporated into routine clinical workflows. Further prospective multicenter studies are warranted to evaluate broader clinical implementation and workflow performance.

Footnotes

Acknowledgements

The authors thank Minh Van Le from Nguyen Trai Hospital for his assistance and support in data collection.

ORCID iD

Si Van Nguyen

Ethical considerations

The study protocol was reviewed and approved by the Institutional Review Board of the University of Medicine and Pharmacy at Ho Chi Minh City (IRB-VN01002/IRB00010293/FWA00023448). All procedures were conducted in accordance with the ethical standards of the institutional research committee and with the 1964 Helsinki declaration and its later amendments.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work has received support from the Korea International Cooperation Agency (KOICA) under the project entitled “Education and Research Capacity Building Project at University of Medicine and Pharmacy at Ho Chi Minh City” conducted from 2024 to 2025 (Project No. 2021-00020-3).

Declaration of conflicting interests

The authors declared the following potential conflicts of interest with respect to the research, authorship, and/or publication of this article: Minh Khac Ho is affiliated with OCTOMED Co., Ltd., which developed the wearable ECG device used in this study. The device was used solely for data acquisition, and all diagnostic decisions were made by cardiologists. The other authors declare no conflicts of interest.

Contributorship

Si Van Nguyen: Conceptualization, study design, data analysis, manuscript drafting, and critical revision; Minh Khac Ho: AI model development, critical revision; Dat Vu Nguyen: Data labeling; Canh Quang Nguyen: Data labeling; An Le Pham: Supervision, critical revision; Hung Thanh Quach: Supervision, critical revision. All authors read and approved the final manuscript.

Guarantor

Si Van Nguyen, M.D., Ph.D.

Appendix

References

Joglar

Chung

Armbruster

, et al. ACC/AHA/ACCP/HRS Guideline for the Diagnosis and Management of Atrial Fibrillation: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation 2024; 149(1): e1–e156. https://doi.org/10.1161/CIR.0000000000001193

Van Gelder

Rienstra

Bunting

, et al. ESC Guidelines for the management of atrial fibrillation developed in collaboration with the European Association for Cardio-Thoracic Surgery (EACTS). Eur Heart J 2024; 45(36): 3314–3414. https://doi.org/10.1093/eurheartj/ehae176

Spooner

Messé

Chaturvedi

, et al. ACC Expert Consensus Decision Pathway on Practical Approaches for Arrhythmia Monitoring After Stroke: A Report of the American College of Cardiology Solution Set Oversight Committee. J Am Coll Cardiol 2025; 85(6): 657–681. https://doi.org/10.1016/j.jacc.2024.10.100

Lown

Garrard

Irving

, et al.

Should we screen for atrial fibrillation?

Br J Gen Pract 2017; 67(660): 296–297. https://doi.org/10.3399/bjgp17X691613

Hwan Bae

Hoon Lee

Heon Yang

, et al. Erroneous computer electrocardiogram interpretation of atrial fibrillation and its clinical consequences. Clin Cardiol 2012; 35(6): 348–353. https://doi.org/10.1002/clc.22000

Manetas-Stavrakakis

Sotiropoulou

Paraskevas

, et al. Accuracy of Artificial Intelligence-Based Technologies for the Diagnosis of Atrial Fibrillation: A Systematic Review and Meta-Analysis. J Clin Med 2023; 12(20): 6576. https://doi.org/10.3390/jcm12206576

Seo

Kim

, et al. ECG data dependency for atrial fibrillation detection based on residual networks. Sci Rep 2021; 11(1): 18256. https://doi.org/10.1038/s41598-021-97308-1

Moody

Mark

. The impact of the MIT-BIH arrhythmia database. IEEE Eng Med Biol Mag 2001; 20(3): 45–50. https://doi.org/10.1109/51.932724

Zhang

Xing

, et al. Robust Artificial Intelligence Tool for Atrial Fibrillation Diagnosis: Novel Development Approach Incorporating Both Atrial Electrograms and Surface ECG and Evaluation by Head-to-Head Comparison With Hospital-Based Physician ECG Readers. J Am Heart Assoc 2024; 13(3): e032100. https://doi.org/10.1161/JAHA.123.032100

10.

Chang

Wen

Chou

, et al. Atrial fibrillation detection using ambulatory smartwatch photoplethysmography and validation with simultaneous holter recording. Am Heart J 2022; 247: 55–62. https://doi.org/10.1016/j.ahj.2022.02.002

11.

Kassem

Folkes

Mukherjee

, et al. Performance of an artificial intelligence-powered smartphone application in the UK clinical settings: ECG automation compared to healthcare professionals. Egypt Heart J 2025; 77(1): 97. https://doi.org/10.1186/s43044-025-00689-1

12.

Szabo

Raisi-Estabragh

Salih

, et al. Clinician's guide to trustworthy and responsible artificial intelligence in cardiovascular imaging. Front Cardiovasc Med 2022; 9: 1016032. https://doi.org/10.3389/fcvm.2022.1016032

13.

Steinberg

Varma

Cygankiewicz

, et al. ISHNE-HRS expert consensus statement on ambulatory ECG and external cardiac monitoring/telemetry. Heart Rhythm 2017; 14(7): e55–e96. https://doi.org/10.1016/j.hrthm.2017.03.038

14.

Reynolds

Stein

Sun

, et al. Cost-Effectiveness of AF Screening With 2-Week Patch Monitors: The mSToPS Study. Circ Cardiovasc Qual Outcomes 2023; 16(11): e009751. https://doi.org/10.1161/CIRCOUTCOMES.122.009751