ECG data were acquired using a PS201-01 wearable Holter ECG device [10] with either shirt-type or belt-type garment electrodes (Figure 1).

For model development, the RRI segments were extracted from raw ECG waveforms and annotated by a cardiologist.

Two-dimensional feature images used as inputs to the DL model were generated from the RRI time series data to construct the training and evaluation datasets.

The data used for training or validation were handled according to the following rules:

The data used for test were handled according to the following rules:

A supervised DL model was constructed to classify the generated 2D RRI images into non-AF, AF, and noise classes.

Between March 1, 2023, and November 20, 2025, 129 patients underwent monitoring with a single-lead garment-type wearable Holter ECG device. Of these, 12/129 (9.3%) patients with documented AT/AFL or paced rhythm were excluded according to the predefined task definition. Consequently, 117/129 (90.7%) patients were included in model development and internal validation. Their baseline characteristics are summarized in Table 1.

The institutional dataset comprised a total of 2850 hours of ECG recordings from 117 patients. After segmentation and label aggregation, the number of unique window-level samples before data augmentation for the 1.5-, 3-, and 6-minute settings was 114,000, 56,969, and 28,452, respectively (Figure 4). Across all window length settings, non-AF windows were the most frequent class, whereas noise windows were the least frequent, indicating class imbalance in the generated datasets.

To characterize the rhythm composition of the institutional cohort, patient-level AF burden was calculated from the original 10-second annotations as the proportion of AF-labeled intervals among analyzable intervals. The AF burden distribution was bimodal: 53 (45.3%) patients had an AF burden of 0%, whereas 49 (41.9%) patients had an AF burden of ≥80%; patients with intermediate AF burden were relatively infrequent (Table S1 in Multimedia Appendix 2).

In the rudimentary semiquantitative review of ectopic burden among 53 patients with an AF burden of 0%, estimated ectopic burden was <1% in 36 (67.9%) patients, 1% to <5% in 8 (15.1%) patients, 5% to <10% in 3 (5.7%) patients, and ≥10% in 6 (11.3%) patients. This review suggested that the non-AF recordings were not limited to exceptionally clean sinus rhythm. The distribution is summarized in Table S2 in Multimedia Appendix 2.

Noise-labeled intervals were commonly observed in the institutional recordings. The total number of noise-labeled 10-second intervals was 249,285, corresponding to approximately 692.5 hours of noise-labeled recording. At the patient level, the median noise burden was 17.6% (IQR 7%‐36%), and only 2 of 117 patients had a noise burden of 0%. These findings indicate that uninterpretable or unreliable segments were commonly present in long-term garment-type wearable Holter ECG recordings. The detailed distributions of noise burden are provided in Table S3 in Multimedia Appendix 2.

In the supplementary review of 300 randomly sampled 30-second ECG segments, 238 (79.3%) were considered annotatable. All pure non-AF segments and all pure AF segments were annotatable, and all these were graded as acceptable. Almost all pure noise segments (49/50, 98%) were judged nonannotatable and unreadable. Among noise-containing segments, 37 (74%) of 50 were annotatable and readable. Across the 238 annotatable segments, 237 (99.6%) were graded as acceptable, and 1 (0.4%) was graded as minor errors, whereas none were graded as unreliable.

In the supplementary annotation verification analysis, 150 randomly sampled ECG segments from the in-house dataset were reassessed, including 50 non-AF segments, 50 AF segments, and 50 noise segments. Agreement with the original study annotations was high for both the reevaluation by the original annotator and the independent review by an additional cardiologist. Cohen κ was 0.91 for the comparison between the original annotations and the reevaluation by the original annotator and 0.95 for the comparison between the original annotations and the independent cardiologist, with corresponding agreement percentages of 94% and 96.7%, respectively. Detailed pairwise agreement (Cohen κ) and class-stratified agreement by label are provided in Table S4 in Multimedia Appendix 2.

The RRI histogram images were generated for each window length. Representative images of each class are shown in Figure 5. The proportion of candidate training windows excluded because they contained both AF and non-AF labels was less than 0.1% of all candidate windows. Exact ties requiring the application of the predefined tie-breaking rule were uncommon in the test data label aggregation process, occurring in 0% of 1.5-minute windows, 0.4% of 3-minute windows, and 0.3% of 6-minute windows. These low frequencies indicate that the tie-breaking rule had minimal influence on the overall evaluation. The mean performance metrics across the 5 cross-validation folds for each window length are presented in Table 2, and the detailed class-wise metrics, calculated in a one-versus-rest manner for the non-AF, AF, and noise classes, are provided in Table S5 in Multimedia Appendix 2.

The 1.5- and 3-minute windows achieved similarly high overall performance, with mean (SD) cross-validation accuracies of 96.6% (SD 1.2) for both models, macro F₁-scores of 96.3% (SD 1.2) and 96.4% (SD 1.1), and macro AUROCs of 99.6% (SD 0.3) and 99.6% (SD 0.3), respectively. For the AF class, the 3-minute model showed slightly better balance than the 1.5-minute model, with comparable sensitivity of 96.7% (SD 2.7) versus 96.7% (SD 3.2), and slightly higher specificity of 98.2% (SD 0.8) versus 98.0% (SD 0.9) and PPV of 95.9% (SD 2.1) versus 95.6% (SD 2.2). In contrast, the 6-minute model showed slightly lower overall accuracy of 96.1% (SD 1.1), macro F1-score of 95.8% (SD 1.2), and AF sensitivity of 94.9% (SD 4.2), although AF specificity remained high at 98.4% (SD 0.4). Performance for the non-AF class was also consistently high across window lengths, indicating stable discrimination of AF from non-AF rhythms in the 3-class framework. Noise class performance appeared favorable in the internal validation, although its interpretation should remain cautious in light of the lower external noise sensitivity discussed in the following paragraph. Receiver operating characteristic curves for each window length are shown in Figure S3 in Multimedia Appendix 1.

In an exploratory post hoc subgroup analysis by garment type using pooled out-of-fold predictions from the 3-minute model, performance was similar between shirt-type and belt-type recordings. Accuracy was 96.8% in shirt-type recordings and 96.5% in belt-type recordings, and AF sensitivity was 97.4 and 96.7, respectively. However, the shirt-type subgroup included fewer patients and showed substantial class imbalance; therefore, definitive conclusions regarding garment-specific performance differences cannot be drawn. The detailed subgroup results are provided in Table S6 in Multimedia Appendix 2.

The final DL models, trained using the 117-patient institutional dataset, were evaluated using the MIT-BIH AFDB to assess their generalizability. In the primary external validation, AFL-annotated intervals were excluded to align the test-label definition with the training task, in which AT/AFL episodes had been excluded. The results are summarized in Table 3.

The external validation dataset consisted of 23 independent AFDB recordings (approximately 230 h in total). As the AFDB did not originally include an explicit noise label, visually uninterpretable windows were additionally annotated as noise. The resulting reference-label distribution was imbalanced across all window length settings: for the 1.5-minute setting, 5549 non-AF (59.7%), 3701 AF (39.8%), and 49 noise (0.5%) windows; for the 3-minute setting, 2785 non-AF (60%), 1834 AF (39.5%), and 20 noise (0.5%) windows; and for the 6-minute setting, 1385 non-AF (59.8%), 922 AF (39.8%), and 10 noise (0.4%) windows.

In the AFDB, the 3-minute model achieved the window-level accuracy of 97.3%, with an AF sensitivity of 96.9% and AF specificity of 97.7%. The 1.5-minute model achieved an overall accuracy of 96.6%, AF sensitivity of 94.2%, and AF specificity of 98.6%, representing a modest reduction in overall accuracy and AF sensitivity compared to the 3-minute model. The 6-minute model showed overall accuracy of 96.6%, AF sensitivity of 94.7%, and AF specificity of 98.1%, which were slightly lower than those of the 3-minute window. Accordingly, in the external dataset, the 3-minute window showed the highest overall accuracy and AF sensitivity while maintaining high AF specificity, although the differences across window lengths were modest and no formal statistical comparison was performed. Across all window lengths, however, sensitivity for the noise class was low: 79.6% (39/49), 70% (14/20), and 80% (8/10) for the 1.5-, 3-, and 6-minute models, respectively. This result should be interpreted cautiously because the number of noise-labeled windows in the external dataset was small and the noise characteristics in the AFDB may not have fully matched those in the institutional wearable dataset. Consistent with the class-wise metrics, the confusion matrices showed that AF and non-AF were generally well discriminated, whereas noise windows were relatively infrequent and more often misclassified than the other classes (Figure 6).

To further characterize the low noise sensitivity, we reviewed the false-negative noise windows in the 3-minute external validation analysis. Among the 20 noise-labeled windows, 6 were misclassified as non-AF or AF. Visual review of the corresponding ECG waveforms and RRI histograms suggested 2 common patterns. In 3 windows, the ECG waveform was visually nondiagnostic for rhythm interpretation, but the automatic R-peak detection algorithm generated structured RRI sequences, resulting in RRI histograms that did not show a typical scattered noise pattern. In the remaining 3 windows, visually noisy ECG portions were mixed with analyzable rhythm portions within the same 3-minute window, leaving structured RRI features in the model input despite the window-level noise label. Representative examples are shown in Figure S4 in Multimedia Appendix 1.

We visually inspected the windows misclassified by the 3-minute model to characterize common patterns in the corresponding images. Among the AF windows misclassified as non-AF, 87.5% (49/56) corresponded to images in which AF and non-AF rhythms coexisted within the same 3-minute window, whereas windows in which AF clearly persisted beyond 3 minutes but was not detected accounted for 12.5% (7/56) of the cases. These findings indicate that most AF-to-non-AF misclassifications were associated with mixed-rhythm windows. Representative misclassified images are shown in Figure 7.

The results of the secondary analysis in which AFL intervals were mapped to the AF class are provided in Table S7 in Multimedia Appendix 2. Across all window lengths, performance in this secondary analysis was slightly lower than in the primary AFL-excluded analysis; for example, for the 3-minute model, accuracy decreased from 97.3% to 96.1%, and AF sensitivity decreased from 96.9% to 93.3%.

When AF burden was treated as a continuous patient-level variable derived from aggregated window-level classifications, the model showed a strong correlation with the reference AF burden across all 3 window length settings (Figure 8). Pearson correlation coefficients were 0.995 (R²=0.991), 0.991 (R²=0.982), and 0.989 (R²=0.979) for the 1.5-, 3-, and 6-minute models, respectively, and the corresponding Spearman rank correlation coefficients were 0.988, 0.982, and 0.979. Overall, patient-level AF burden correlation was high across all window lengths, with only small differences between models.

In this study, we developed a DL model that classifies RRI-derived 2D histogram images into non-AF, AF, and noise classes using a 2D-CNN. Internal and external validation suggested that the 3-minute window provided a clinically reasonable balance across the evaluated window lengths, although the differences were modest and should be interpreted as exploratory rather than definitive. In the external validation, the model also showed high correlation with the reference AF burden at the patient level. The contribution of this study lies not in the use of a novel network architecture, but in the application and evaluation of an RRI-based 3-class classification framework that explicitly incorporates noise in real-world garment-type wearable Holter ECG recordings.

For automated AF diagnosis, DL approaches can be broadly grouped into waveform-based models, RRI/heart rate variability (HRV)–based models, and multimodal frameworks that combine ECG-derived features with additional clinical information. Waveform-based models trained directly on raw ECG signals have shown high diagnostic performance in a variety of datasets, including residual network–based approaches, but their performance may depend substantially on recording format, window length, and dataset characteristics [18-22,31]. In contrast, RRI/HRV-based approaches focus on the beat-to-beat irregularity that is fundamental to AF and have been explored using transition-matrix features, entropy-based indices, multivariate classifiers, visual irregularity mapping, Lorenz/Poincaré-type representations, and image-based DL frameworks [4,23-25,32-34]. Recent multimodal work has further suggested that combining ECG with HRV and demographic information may improve AF detection performance in selected settings [35]. These studies provide useful context for the present findings, but direct numerical comparison should be interpreted cautiously because the underlying datasets, recording durations, input modalities, and intended use cases differ substantially. In this context, this study does not seek novelty in the network architecture itself; rather, its contribution lies in evaluating an RRI-based, 3-class, noise-aware framework in long-term garment-type wearable Holter ECG recordings, a setting in which noisy and clinically heterogeneous data are common. By converting the RRI time series into histogram images, this approach preserves AF-related irregularity patterns in a form suitable for CNN-based analysis while remaining conceptually closer to established HRV-based rhythm assessment than to raw waveform classification. As RRI-based representations use interval-derived information rather than full raw waveform input, converting an RRI time series into 2D images for CNN-based analysis may provide a conceptually simple and practical framework for AF screening in noisy wearable ECG recordings.

With a single-lead Holter or wearable ECG, reliably identifying P waves and fibrillatory waves (f waves), which are important for conventional AF diagnosis, can be challenging in the presence of noise. This problem is particularly pronounced in garment-type devices (eg, shirt- or belt-type systems), where artifacts arising from body movements and variable electrode contacts are common [8,36]. Garment-type Holter systems are expected to reduce the skin burden associated with adhesive electrodes, but they may be associated with more frequent noise contamination because of motion artifacts and variable electrode contact. Compared to curated public databases, real-world long-term Holter recordings contain a greater proportion of segments affected by noise artifacts. If such noise is not adequately identified, model performance, particularly specificity, may deteriorate, leading to an increased number of false positives [37]. RRI-based analysis offers a practical advantage in this setting; provided that R-peaks are accurately detected, it may be less dependent on raw waveform morphology than waveform-based approaches, making it attractive for handling noisy wearable ECG recordings [38].

To better handle noise-contaminated wearable ECG data, we implemented two additional strategies: (1) histogram binning of the RRI time series and (2) explicit annotation and learning of noise as a separate class. Histogram binning may provide a degree of local aggregation, thereby reducing the effects of minor timing fluctuations or isolated R-peak detection errors [39]. In addition, by formulating the task as a 3-class problem (non-AF, AF, and noise), the model was designed to distinguish uninterpretable segments from AF and non-AF rhythms without requiring a separate preprocessing step to exclude noisy data in advance. This framework may be advantageous for real-world wearable Holter ECG recordings, in which noise contamination is often unavoidable. Several previous studies have reported higher classification accuracies after rigorously excluding noisy segments from analysis [4,34,40]. However, such approaches may be less applicable to daily wearable monitoring, where preprocessing and manual cleaning are not always feasible. In this study, the AF diagnostic performance remained high despite the inclusion of noise-containing recordings, suggesting that this 3-class noise-aware framework is practical for AF screening in garment-type wearable ECG monitoring.

At the same time, noise sensitivity was substantially lower in the external validation than in the internal validation. One possible explanation is the very small proportion of noise-labeled segments in the external dataset, which makes sensitivity estimates unstable. Another possibility is a domain shift in noise characteristics: many noise segments in the institutional wearable dataset were associated with relatively prolonged signal degradation, such as electrode detachment or sustained signal loss, whereas the noise segments additionally annotated in the AFDB were retrospectively identified as visually apparent artifacts in a public dataset that did not originally include an explicit noise label. Additional review of false-negative noise windows suggested that some visually nondiagnostic ECG segments still yielded structured RRI histograms after automatic R-peak detection, while others contained mixed noisy and analyzable ECG portions within the same window. Thus, waveform-level visual noise annotation did not always correspond to degradation of the RRI-based feature representation used by the model, limiting the generalizability of noise classification across datasets. Although lower noise sensitivity should be regarded as a limitation, the consistently high noise specificity indicates that the model rarely assigned analyzable AF or non-AF windows to the noise class. This may be relevant in clinical review workflows, where excessive exclusion of interpretable rhythm segments could also be undesirable. The noise class in the present framework should therefore be interpreted as an operational class to avoid forcing unreliable windows into AF or non-AF categories, rather than as a universal artifact detector. Future implementations may benefit from complementary signal quality assessment, wear status detection, or preprocessing procedures.

The choice of the analysis window length for AF detection must be aligned with the model performance and the clinical significance of AF episodes. Previous studies using cardiac implantable electronic devices have suggested that stroke risk may increase when atrial high-rate episodes persist for ≥6 minutes, and 6 minutes have therefore been proposed as a clinically relevant threshold [28,29]. In this study, considering this evidence, we evaluated window lengths of 1.5, 3, and 6 minutes. Both internal and external validations suggested that the 3-minute window provided a clinically reasonable balance between AF sensitivity and specificity. However, the differences across window lengths were modest, and no formal statistical comparison was performed; therefore, this interpretation should be regarded as exploratory rather than definitive. The 3-minute model showed a high patient-level correlation with the reference AF burden in the AFDB, with performance comparable to that reported in previous studies [4,41]. A plausible explanation for the slightly lower AF sensitivity of the 6-minute model is that longer windows are more likely to include mixed rhythm or transitional segments, which may dilute AF-related irregularity within a single window; however, the high patient-level AF burden correlation suggests that the practical impact of this effect on overall burden estimation was limited. These window-level classification metrics and patient-level AF burden estimates address complementary aspects of performance: the former reflects the model’s ability to classify individual ECG windows, whereas the latter reflects the clinical interpretability of aggregating those predictions over an entire recording. Although small differences in AF burden estimates may not necessarily alter patient management in every clinical context, this analysis was included to show that the aggregated model outputs remained clinically meaningful beyond segment-level classification accuracy alone. From a practical monitoring perspective, a 3-minute window also allows AF status to be updated at relatively short intervals while still permitting consecutive-window aggregation to identify clinically relevant sustained episodes. Collectively, these findings indicate that a 3-minute RRI-based model may provide a practical balance for AF screening and AF burden assessment in ambulatory long-term Holter monitoring while allowing consecutive-window aggregation to support the detection of sustained AF episodes.

As the present model was specifically designed to detect AF on the basis of RRI irregularity, AT/AFL episodes were excluded from the primary training and validation framework. AT/AFL often exhibit relatively regular RRIs and therefore do not consistently share the irregular rhythm pattern that characterizes AF, making them difficult to detect reliably using an RRI irregularity–based approach alone [26]. In a secondary analysis, in which AFL was grouped with the AF class for broader clinical categorization, diagnostic performance decreased, further supporting the limited suitability of the present RRI-based framework for identifying AT/AFL as AF. Therefore, future studies should explore a 2-stage algorithm in which RRI-based screening is followed by a detailed waveform-based analysis to identify AT/AFL among segments classified as non-AF. Incorporating additional features, such as waveform morphology and periodicity measures, may help construct a more comprehensive framework for diagnosing supraventricular arrhythmias using wearable ECG devices [42].

As the demand for AI-based diagnosis from single-lead ECG data increases, AF detection algorithms embedded in standard smartwatches still face limitations, including “unclassifiable” recordings and suboptimal sensitivity and specificity [43,44]. Conversely, the direct application of DL to smartwatch ECG waveforms can yield a performance that markedly surpasses that of built-in algorithms, suggesting that the potential for AI applications in wearable ECG will continue to grow. Within this context, the RRI-based 2D-CNN model developed in this study appears to be a promising AF screening approach that combines noise-aware design with potential practical applicability in prolonged high noise settings, such as garment-type wearable Holter monitoring. Our findings may contribute to the future implementation of garment-type wearable diagnostic systems and the development of efficient AF screening tools in primary care and community-based health care settings.

Although this study demonstrated the effectiveness of an RRI-based automatic AF detection model using single-lead Holter ECG recordings, it has several limitations.

First, although the CNN-based DL model achieved high classification performance, its decision-making process relied on internal latent representations and was not directly interpretable at the individual-prediction level. This inherent “black-box” nature reflects the broader challenge of explainability in medical AI and underscores the need for complementary approaches, such as model-agnostic explanation techniques, if the system is to be used for clinical decision support.

Second, because AT/AFL episodes and paced rhythms were excluded from the primary training and validation framework according to the predefined task definition, the present model was not designed or evaluated as a general classifier of supraventricular tachyarrhythmias or paced ECG recordings. In particular, pacemakers may artificially regulate RRIs and thereby alter the RRI irregularity patterns on which the present algorithm depends. Accordingly, its applicability to AT/AFL detection and to recordings containing paced rhythms remains limited and should be addressed in future studies using dedicated datasets and complementary features beyond RRI irregularity alone.

Third, although internal validation was performed using institutional data and external validation was conducted using the AFDB, the model’s generalizability across different clinical settings, patient populations, recording conditions, and ECG devices remains uncertain. In particular, the institutional and external validation datasets were enriched for AF relative to a general screening population. As PPV and NPV are mathematically dependent on disease prevalence, the favorable predictive values observed in this study may not directly translate to lower-prevalence screening settings; in particular, the PPV may be lower when the algorithm is applied to populations in which the true prevalence of AF is low. Noise characteristics can also vary substantially among wearable devices and datasets, and this may partly explain the markedly lower noise sensitivity observed in the external validation. Future studies should determine whether the same model, or a suitably adapted version, can be applied across diverse hardware platforms, external noise conditions, and lower-prevalence screening populations. In addition, because windows containing both AF and non-AF labels were excluded from training to avoid ambiguous supervisory signals, the reported classification performance may be somewhat optimistic relative to more heterogeneous real-world recordings containing frequent rhythm transitions. The proportion of such excluded candidate training windows was less than 0.1% in the present dataset. Consistent with this low transition rate, the patient-level AF burden distribution was bimodal, with many patients showing either no AF or very high AF burden and relatively few patients showing intermediate AF burden. Therefore, patients with highly active PAF characterized by frequent transitions between AF and non-AF rhythms were likely underrepresented. Moreover, because window-level reference labels were assigned using a majority vote rule, short AF episodes occupying less than half of a given window could be underrepresented, particularly in longer windows such as the 6-minute setting. Thus, the present framework should be interpreted as a window-level AF screening and burden estimation approach rather than a definitive episode-level detector for very brief AF events.

Fourth, the proposed method depends on automatic R-peak detection because the model input was derived from RRI data. Potential sources of R-peak detection error include noise and artifacts in Holter or wearable ECG recordings, as well as limitations in the acquisition system. Although the device incorporated analog band limiting intended for antialiasing before 250 Hz ADC sampling, the available specifications did not indicate a sharp low-pass filter below 100 Hz; therefore, residual aliasing of high-frequency components cannot be completely excluded. To assess the practical reliability of the R-peak detection step, we performed a supplementary manual quality assessment using 300 randomly sampled 30-second ECG segments; however, this was not an exhaustive beat-level validation of all 2850 hours of institutional recordings. In the external validation, manually annotated noise segments in the AFDB may not have fully reproduced the sustained mechanical artifacts encountered in garment-type wearable ECG recordings, and waveform-level visual noise labels may not always correspond to degradation of RRI-based feature representations. These factors may have contributed to the low noise sensitivity observed in the external validation. In addition, because the RRI-derived histogram was limited to a predefined heart rate range of 30 to 400 beats per minute, RRIs longer than 2.0 seconds were not explicitly represented as prolonged-pause features. Thus, clinically significant pauses, such as those observed in tachy-brady syndrome, would require separate pause-detection logic or waveform-based review.

Fifth, although we performed a rudimentary semiquantitative review of ectopic burden in patients with 0% AF burden, this assessment was based on sampled non-AF ECG segments and was not a formal beat-by-beat adjudication of premature atrial contractions or premature ventricular contractions across the entire non-AF class. The review suggested that the non-AF recordings were not limited to exceptionally clean sinus rhythm, but the precise prevalence and distribution of ectopic beats in all non-AF windows remain unknown. As the present model relies entirely on RRI irregularity, frequent ectopy may mimic AF-like irregularity, and may affect the interpretation of the model’s true specificity.

This study developed and evaluated an RRI-based AF screening approach in which RRI time series data extracted from single-lead garment-type Holter ECGs were transformed into 2D images for DL-based classification. A 2D-CNN model trained on these visualized RRI features demonstrated high classification accuracy in both internal and external validations, suggesting potential utility for AF screening in prolonged monitoring settings using recordings without paced rhythms; recordings containing paced rhythms should currently be considered outside the intended use of this RRI-based algorithm. The 3-class framework, which explicitly separated noise from AF and non-AF rhythms, may be useful for handling real-world garment-type recordings that contain uninterpretable segments. However, broader validation across more diverse patient populations, recording environments, and device types will be necessary before routine clinical deployment can be considered. Importantly, the real-world performance of this framework will depend on the reliability of R-peak detection and on device- and dataset-specific noise characteristics.

Future work should further evaluate real-time feasibility and explore integration with waveform-based approaches for other supraventricular arrhythmias, such as AT/AFL.