Between January 1, 2024, and January 1, 2025, consecutive patients diagnosed with paroxysmal AF were recruited from the Chinese PLA General Hospital. Inclusion criteria were patients aged ≥18 years who provided written informed consent. Exclusion criteria were inability to use wearable devices, cognitive impairment, or presence of implanted cardiac devices (pacemakers or implantable cardioverter-defibrillators).

This study complied with the World Medical Association’s Declaration of Helsinki and was approved by the Institutional Review Board of the Chinese People’s Liberation Army General Hospital (HZKY-PJ-2023-23). It was also registered with the Chinese Clinical Trial Registry (ChiCTR2300075516). All participants signed the informed consent form before participating in this study. This study strictly adhered to privacy protection protocols in accordance with the Declaration of Helsinki. No financial compensation was provided to participants.

This study involved collecting baseline clinical data and PPG signals. Clinical data included demographics, comorbidities, and medications. PPG signals were obtained as follows: after attaching 24-hour Holter ECG electrodes, participants wore smartwatches (Huawei Watch GT2 Pro; Huawei Technologies Co., Ltd.). Simultaneous recordings of cardiac rhythm from both devices were initiated.

The PPG-based AF burden algorithm was developed using 3698 PPG data segments from the previous Mobile Atrial Fibrillation Application (mAFA) study as the training and validation sets [10]. Each data segment had a duration of 1 minute. The classification of the training data is shown in Table 1. The training data segments were each divided into varying durations (8, 16, 24, 32, 40, and 48 seconds) with random start times determined by a random number generator. During model development, we maintained a balanced representation of AF and non-AF episodes across all duration categories, enforced strict subject-level segregation to prevent any overlap between the training and validation datasets, and carefully preserved comparable AF to non-AF ratios in both sets. The raw PPG signal from the sensor module was processed using a bandpass Butterworth digital filter to eliminate low-frequency baseline drift and high-frequency noise, thereby obtaining clear and effective pulsatile PPG waveforms. To overcome limitations of single-time-AF detection algorithms, which exhibit low motion tolerance due to strict signal quality requirements and are inadequate for AF burden assessment, this study proposes a multiscale fusion AF burden detection algorithm featuring continuous PPG acquisition with minute-by-minute interpretation (defining AF burden as >40% of AF beats per minute), high-quality signal extraction from motion corrupted segments, an adaptive length machine learning model for variable duration PPG signals, and context-aware fusion calibration leveraging AF episode continuity to refine single time predictions, thereby achieving accurate AF burden quantification. In this study, we enrolled 145 patients as a test cohort to validate the detection accuracy of the AF burden algorithm compared to 24-hour Holter monitoring.

Metric M1 is the proportion of AF episode duration detected through PPG monitoring relative to the total monitoring time, quantifying temporal AF dynamics (Figure 1).

Metric M2 is the proportion of AF episodes detected through PPG monitoring relative to the total number of monitoring epochs, assessing frequency-based AF variations (Figure 2).

Metric M3 is the area between the actual AF burden progression curve (derived from PPG monitoring) and the theoretical uniform AF burden development curve, evaluating AF episode clustering (Figure 3).

Metric M4 is a composite metric integrating 3 variability parameters derived from PPG monitoring: AF episode variability (SD of hourly AF counts normalized by 24-hour mean AF frequency), mean real variability (average of SD and mean absolute deviation of hourly AF counts), and mean AF variation (day-night difference in AF counts normalized by overall mean AF frequency). This model quantifies hourly AF burden fluctuations (Figure 4).

Metric M5 quantifies rapid ventricular rate (RVR) AF episodes, defined as the proportion of AF episodes with heart rates of >120 beats per minute (bpm) relative to total monitored AF episodes or the cumulative duration of RVR AF episodes relative to the total monitoring time. This dual parameter model enables characterizing high rate AF burden (Figure 5 and Figure 6).

Data with a normal distribution were presented as means and SDs. Data with a nonnormal distribution were presented as medians and IQRs and were analyzed using the Mann-Whitney U test. Categorical variables were expressed as percentages. Algorithm performance (sensitivity, specificity, accuracy, precision, F₁ score, and area under the receiver operating characteristic curve) was validated against Holter ECG values to assess its capability for continuous AF detection. Correlation between the PPG-based AF burden model and 24-hour Holter ECG monitoring was assessed using the mean absolute error (MAE) and Spearman rank correlation coefficient (r_s; Table 2)

A 2-sided P value of <.05 was considered statistically significant. The 95% CIs were calculated using the Wilson score method without continuity correction. Analyses were conducted using SPSS Statistics (version 22; IBM Corp) and OpenEpi (version 3.01).

From January 1, 2024, to January 1, 2025, a total of 148 participants were initially enrolled in the study. Of these 148 participants, after excluding 2 (1.4%) with atrial flutter and 1 (0.7%) with atrioventricular block, 145 (98%) were ultimately included in the final analysis (n=96, 66.2% male; mean age 63.28, SD 14.23; range 19 to 90 years). On the basis of expert-reviewed 24-hour Holter ECG monitoring, 75 patients exhibited AF episodes during the monitoring period, including 35 (47%) cases of paroxysmal AF and 40 (53%) cases of persistent AF lasting 24 hours (Figure 7).

The study population comprised patients with paroxysmal AF with a mean age of 63.28 (SD 14.23) years, including 66.2% (96/145) male individuals. Baseline characteristics are detailed in Table 3.

The AF burden algorithm demonstrated high performance, with a daily model output rate of 86.6% (95% CI 85.2%-88%). The confusion matrix yielded true positive, false negative, false positive, and true negative counts for PPG signals of 106,975, 11,941, 8787, and 308,082, respectively. Compared to the gold standard, the PPG-based AF burden model demonstrated a sensitivity of 91.5% (95% CI 87.9%-95.1%), specificity of 97.2% (95% CI 95.9%-98.5%), precision of 92.9% (95% CI 88.6%-97.3%), accuracy of 93.3% (95% CI 88.2%-98.5%), and F₁-score of 90.5% (95% CI 86.3%-94.7%). The AF burden model exhibited strong discriminatory power in the test cohort (area under the curve=89.5%, 95% CI 89.4%-89.7%; Table 4). The receiver operating characteristic curve is shown in Figure 8. The performance metrics of the PPG-based AF burden model across varying threshold values are shown in Table 5.

Figure 9 shows the strong concordance between the PPG and Holter ECG waveforms.

This study confirmed that the PPG-based AF burden model demonstrated strong agreement in tracking AF burden variations, with metric M1 showing optimal performance: for the ratio of AF duration to the total monitoring time, it achieved an MAE of 0.0400 (P=.008) and r_s of 0.8788 (P<.001); for the ratio of ≥6-minute AF episodes to the total monitoring time, it achieved an MAE of 0.0582 (P=.44) and r_s of 0.9233 (P<.001); and for the ratio of ≥1-hour AF episodes to the total monitoring time, it achieved an MAE of 0.1204 (P=.04) and r_s of 0.9293 (P<.001).

The current clinical classification of AF largely relies on symptomatic presentation and ECG findings, categorizing AF into paroxysmal, persistent, long-standing persistent, and permanent types. The management of AF involves CHA₂DS₂-VASc (congestive heart failure; hypertension; age of ≥75 years; diabetes mellitus; previous stroke, transient ischemic attack, or thromboembolism; vascular disease; age of 65-74 years; and sex category)–guided thrombotic risk stratification for anticoagulation decisions. However, current classifications inadequately capture AF progression dynamics [9,11]. For instance, in real-world settings, it remains unclear whether frequent, short-duration AF episodes confer a higher thrombotic risk than infrequent but prolonged episodes. Currently, no universally accepted AF burden definition exists, although it is commonly quantified as the percentage of time in AF during monitoring [12]. Alternative definitions incorporate AF episode frequency, density, and temporal variability [2,5,13], reflecting methodological heterogeneity. While implantable loop recorders and 24-hour Holter monitoring are widely used for AF burden assessment, implantable loop recorders are invasive, costly, and may overestimate AF burden by misclassifying atrial arrhythmias as AF [14]. Conversely, 24-hour Holter monitoring underestimates burden due to short duration and tolerability limitations, restricting its utility in daily practice [15]. This study leveraged PPG integrated smartwatches to enable multidimensional ambulatory AF monitoring, innovatively characterizing AF burden across 5 distinct dimensions: temporal duration, episode frequency, density, variability, and RVR. This approach offers novel insights for establishing a PPG-based definition of AF burden.

In recent years, wearable PPG devices have gained traction for AF screening owing to affordability and usability [8,16]. Previous studies have validated PPG’s diagnostic accuracy in AF detection [17-20]. Our team’s previous research (N=187,912) confirmed the feasibility and utility of PPG-based wristbands and smartwatches in population level AF screening, facilitating early AF identification and intervention [7]. In a large cohort (N=1,187,381; mean follow-up 255 days), 93.6% of PPG suspected AF cases were verified, demonstrating the reliability and accuracy of this approach [21]. Despite PPG’s utility in AF screening, its burden monitoring potential is underexplored. Väliaho et al [8] evaluated PPG-based algorithms for continuous AF burden assessment in 173 participants, finding 30-minute intervals as optimal (F₁ score=0.95; sensitivity=94.9%; specificity=98.6%) when comparing 10-, 20-, 30-, and 60-minute reporting intervals. Avram et al [22] evaluated a Samsung smartwatch algorithm (5-minute intervals: sensitivity=87.8%, 95% CI 83.6%-91%; specificity=97.4%, 95% CI 97.1%-97.7%; area under the curve=93.3%), showing strong AF burden correlation (R²=0.986). These findings suggest PPG’s potential for AF burden quantification, although detection precision for brief AF episodes requires refinement. In this study, we developed a PPG-based AF burden algorithm integrated into smartwatch devices. When evaluated via 1-minute analysis epochs against standard 24-hour Holter monitoring as the reference, the algorithm demonstrated a sensitivity of 91.5% (95% CI 87.9%-95.1%), specificity of 97.2% (95% CI 95.9%-98.5%), and overall accuracy of 95.7% (95% CI 94%-97.4%). Previous studies have reported PPG signal limitations from motion artifacts and other interfering factors, leading to substantial data exclusion. Reissenberger et al [23] reported that approximately 50.7% of PPG signals were classified as noise, necessitating exclusion. To address these challenges, we implemented algorithm optimizations through a novel multiscale fusion approach for AF burden detection, achieving a daily model output rate of 0.86 (95% CI 0.852-0.880) and markedly improving data usability. In this study, metric M1 demonstrated optimal performance in tracking AF duration, consistent with previous findings, showing an MAE of 0.0400 (P=.008) and r_s of 0.8788 (P<.001) for the ratio of AF duration to total monitoring time and an MAE of 0.1204 (P=.04) and r_s=0.9293 (P<.001) for the ratio of episodes of >1 hour to total monitoring time. While current studies typically use invasive devices or 24-hour Holter monitoring to define AF burden through episode frequency and density, our study pioneered PPG-based AF burden models integrated into smartwatches. Compared with standard 24-hour Holter monitoring, the AF burden model exhibited higher MAE values, along with a weaker r_s, in assessing metrics M2 and M5. This may be attributed to PPG signal interference caused by poor pulse waveform quality, motion artifacts, and noise, which could reduce the models’ accuracy in evaluating AF episode frequency. However, metric M3 demonstrated good correlation and a weaker MAE, with an MAE of 0.1725 (P=.13) and r_s=0.6576 (P<.001), although further optimization is still needed in the future. Notably, metric M4 demonstrated excellent agreement with the gold standard in tracking AF variability, providing novel insights for assessing temporal AF fluctuations.

AF’s rising prevalence and mortality exacerbate global health burdens, with its association to adverse cardiovascular events including stroke, heart failure, and death potentially linked to AF burden [1,24]. The Catheter Ablation Versus Standard Conventional Treatment in Patients With Left Ventricular Dysfunction and Atrial Fibrillation trial demonstrated that an AF burden threshold of 50% was associated with significant functional and structural changes in cardiomyocytes [25]. There is accumulating evidence suggesting that AF burden is correlated with adverse cardiovascular events, including stroke and heart failure [2,4,26]. A large-scale study (N=39,710) defined AF burden as the ratio of daily AF duration percentage to total monitoring time [13] and the longest single AF episode duration, revealing dose-dependent associations between increasing baseline AF burden and adverse outcomes at the 1- and 3-year follow-ups. A high AF burden may be a significant risk factor for mortality. However, the causal role of AF burden in stroke events and the optimal threshold (ranging from 1 minute to 24 hours) remain unclear. Moreover, the relationship between AF episode frequency, AF density, AF episode variability, and the proportion of RVR AF and adverse cardiovascular events remains unexplored. However, in this study, metrics M1 and M4 demonstrated strong correlation compared to 24-hour Holter monitoring. This provides a novel methodological foundation for future research into the association between PPG-based AF burden detection and adverse cardiovascular events.

While there is evidence suggesting that anticoagulation therapy may be considered based on atrial high rate episode burden [27], the optimal treatment strategy according to specific AF burden thresholds requires further investigation. Our study’s metric M1 for temporal AF burden monitoring showed strong correlation with 24-hour Holter monitoring, potentially offering a novel methodology for future large-scale clinical trials investigating anticoagulant guidance by AF burden quantification. Current AF guidelines recommend active rhythm control to alleviate symptoms in patients with AF, with antiarrhythmic drug therapy and more effective catheter ablation techniques demonstrating the capability to prevent AF recurrence, improve symptoms, and reduce AF burden [28-30]. Drexler et al [31] retrospectively demonstrated that early postprocedural low root mean square of successive differences values following pulmonary vein isolation independently predicted AF recurrence (hazard ratio=0.50; P<.001), whereas Zhu et al [32] prospectively identified postablation root mean square of successive differences and percentage of successive normal sinus intervals that differ by more than 50 milliseconds as independent predictors of recurrence in 102 patients with paroxysmal AF. Collectively, recent randomized trials support the beneficial impact of dynamic AF burden monitoring and reduction on clinical outcomes [33]. Our study’s metric M4 exhibited strong correlation with 24-hour Holter monitoring in assessing AF variability, providing a potential technical foundation for future rhythm management strategies guided by AF burden quantification and prediction of postablation recurrence risk.

There is emerging evidence suggesting that AF burden may contribute to adverse cardiovascular outcomes, including stroke and heart failure, necessitating refined assessment methods for optimized risk prediction. While real-world evaluation of AF burden remains uncertain, incorporating this parameter into clinical decision making could enable more precise risk stratification and therapeutic selection. Our study developed a novel smartwatch-based PPG algorithm showing strong concordance with 24-hour Holter monitoring in long-term AF burden assessment. This innovative model incorporates 5 dimensions—temporal AF duration, episode density, frequency, variability, and proportion with rapid ventricular response—enabling multidimensional AF evaluation. This approach provides new insights for investigating PPG-derived AF burden’s relationship with cardiovascular outcomes, potentially informing early intervention strategies, anticoagulation decisions, and rhythm management protocols.

This study has several limitations that should be acknowledged. First, the relatively modest sample size warrants validation in larger, real-world cohorts to ensure generalizability. In addition, PPG technology remains significantly susceptible to motion artifacts and noise interference, particularly during daily activities and hand movements, which may compromise accurate quantification of AF episodes. This technical limitation presents a notable challenge in assessing RVR AF. However, ongoing advancements in wearable device technology and detection algorithms are expected to mitigate these limitations. Second, while current PPG applications are limited to AF screening and cannot provide definitive diagnosis, our preliminary research demonstrated the diagnostic feasibility of integrated single-lead electrocardiograph technology. Future investigations should prioritize large-scale, multicenter randomized controlled trials to systematically evaluate the clinical utility of combining PPG with iECG technology for comprehensive AF burden monitoring.

A PPG-based AF burden model incorporating dimensions such as duration, frequency, density, variability, and RVR demonstrated good consistency compared to 24-hour Holter monitoring, with optimal tracking performance observed in the duration and variability dimensions. This approach provides technical support for the at-home multidimensional quantification of AF occurrence and progression, offering new insights for defining PPG-based AF burden. However, further optimization and validation are required for the real-world application of this PPG-based AF burden model.