This paper is in the following e-collection/theme issue:

Cardiac Arrhythmia, Atrial Fibrillation (115) Machine Learning (2076) Wearable Devices and Sensors (1133) Cardiac Disease Management (156) e-Cardiology and Telemonitoring in Cardiology (221) Novel Sensors and Data Acquisition Methods in Cardiology (225) Mobile Apps for Cardiology (189)

Published on in Vol 9 (2025)

Preprints (earlier versions) of this paper are available at https://preprints.jmir.org/preprint/78075, first published .
Photoplethysmography-Based Machine Learning Approaches for Atrial Fibrillation Burden: Algorithm Development and Validation

Photoplethysmography-Based Machine Learning Approaches for Atrial Fibrillation Burden: Algorithm Development and Validation

Photoplethysmography-Based Machine Learning Approaches for Atrial Fibrillation Burden: Algorithm Development and Validation

Authors of this article:

Hong Wang1, 2 Author Orcid Image ;   Binbin Liu1, 2 Author Orcid Image ;   Hui Zhang1 Author Orcid Image ;   Zheqi Zhang1, 2 Author Orcid Image ;   Zhigeng Jin1 Author Orcid Image ;   Hao Wang3 Author Orcid Image ;   Yu-Tao Guo1 Author Orcid Image
Article Authors Cited by Tweetations Metrics

    1Department of Pulmonary Vessel and Thrombotic Disease, Sixth Medical Center, Chinese PLA General Hospital, 6 Fucheng Road, Haidian District, Beijing, China

    2Graduate School of PLA General Hospital, Beijing, China

    3Department of Cardiology, Second Medical Centre, National Clinical Research Centre for Geriatric Diseases, Chinese PLA General Hospital, Beijing, China

    Corresponding Author:

    Yu-Tao Guo, Prof Dr


    Background: Atrial fibrillation (AF) burden is associated with cardiovascular events such as stroke and heart failure. Recent advancements in photoplethysmography (PPG) technology have provided new insights into noninvasive and convenient AF burden detection.

    Objective: This study aimed to establish an AF burden model based on smartwatch-monitored PPG technology to track the progression of AF.

    Methods: This prospective pilot study (January 2024 to January 2025) at the Chinese PLA General Hospital enrolled patients with paroxysmal AF. Participants underwent simultaneous rhythm monitoring using smartwatch PPG and 24-hour Holter electrocardiogram monitoring (the gold standard). Five PPG-derived AF burden metrics were defined: (1) ratio of AF episode duration to total monitoring time (M1), (2) ratio of AF episode frequency to total measurements (M2), (3) AF episode density (M3), (4) AF episode variability (M4), and (5) proportion of rapid ventricular rate in AF episodes (>120 beats per minute; M5). Smartwatch PPG signals were collected once per minute. Sensitivity, specificity, accuracy, precision, and F1 score were used to evaluate the PPG algorithm’s AF detection capability through comparison with the gold standard (24-hour Holter monitoring). The mean absolute error (MAE) and Spearman rank correlation coefficient (rs) were used to assess the correlation between the PPG-based AF burden metrics and the gold standard.

    Results: A total of 145 participants with paroxysmal AF (n=96, 66.2% male; mean age 63.28, SD 14.23 years) were included. 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%). For M1, the MAE for the model of AF episode duration as a proportion of total monitoring time was 0.0400 (P=.008), with a correlation coefficient (rs) of 0.8788 (P<.001). For M4, the MAE for the AF episode variability model was 3.9967 (P<.001), with a correlation coefficient (rs) of 0.7876 (P<.001). The MAE for the average real variability model was 4.6436 (P<.001), with a correlation coefficient (rs) of 0.8127 (P<.001). The MAE for the average AF change model was 0.3893 (P=.27), with a correlation coefficient (rs) of 0.7246 (P<.001).

    Conclusions: The PPG-based AF burden model demonstrated high concordance with the gold standard of 24-hour Holter monitoring in tracking AF episode duration and variability, providing new perspectives for exploring AF progression dynamics.

    Trial Registration: Chinese Clinical Trial Registry ChiCTR2300075516; https://www.chictr.org.cn/showproj.html?proj=200976

    JMIR Cardio 2025;9:e78075

    doi:10.2196/78075

    Keywords

    atrial fibrillation; photoplethysmography; atrial fibrillation burden; wearable devices; arrhythmia 


    Atrial fibrillation (AF) is one of the most common arrhythmias, associated with an increased risk of cardiovascular adverse events such as stroke and heart failure, and causes significant health, social, and economic burdens [1]. Numerous studies have found a dose-response relationship between AF burden and cardiovascular event risk, making early assessment of AF burden critical to reduce AF complications [2-4]. However, there is currently no standardized definition for AF burden. It generally refers to the percentage of time spent in AF during total monitoring duration. Other studies have alternatively quantified the number of AF episodes by using episode density to define AF burden to explore the relationship between AF burden and cardiovascular adverse events [2,5].

    Currently, devices for monitoring AF burden include cardiovascular implantable electronic devices and noninvasive rhythm recorders (such as 24-hour Holter electrocardiograms [ECGs]), both of which demonstrate high diagnostic accuracy. However, these devices face limitations, including requiring hospital-based monitoring under clinician supervision, high costs, and labor-intensive data interpretation by physicians. Recently, photoplethysmography (PPG) has been developed for AF screening and has shown promising accuracy [6,7]. PPG devices for AF screening are commercially available in various formats, including handheld devices, smartwatches, or wristbands. They use built-in optical sensors to monitor blood volume changes in skin capillary beds and estimate heart rhythm via reflected light wavelengths, offering greater comfort and convenience and providing potential opportunities for out-of-hospital ambulatory AF burden assessment [8]. However, the accuracy of PPG-based AF burden detection algorithms in smartwatches remains uncertain [9].

    Our study aimed to evaluate the accuracy of smartwatch integrated PPG algorithms in assessing AF burden.


    Study Population

    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).

    Ethical Considerations

    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.

    Signal Acquisition and Processing

    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.

    Development and Optimization of the Primary PPG-Based AF Burden Model

    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.

    Table 1. The classification of the training data.
    Training data categoryRecord (1-minute segment)
    Atrial fibrillation1090
    Premature contraction540
    Sinus rhythm2067

    Definition of AF Burden

    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).

    ‎
    Figure 1. A schematic diagram of metric M1. AF: atrial fibrillation; ECG: electrocardiogram; PPG: photoplethysmography; SR: sinus rhythm.

    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).

    ‎
    Figure 2. A schematic diagram of metric M2. AF: atrial fibrillation; ECG: electrocardiogram; PPG: photoplethysmography; SR: sinus rhythm.

    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).

    ‎
    Figure 3. A schematic diagram of metric M3. AF: atrial fibrillation; ECG: electrocardiogram; PPG: photoplethysmography; SR: sinus rhythm.

    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).

    ‎
    Figure 4. A schematic diagram of metric M4. AF: atrial fibrillation; ECG: electrocardiogram; PPG: photoplethysmography; SR: sinus rhythm.

    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).

    ‎
    Figure 5. A schematic diagram of metric M5. AF: atrial fibrillation; bpm: beats per minute; ECG: electrocardiogram; PPG: photoplethysmography; SR: sinus rhythm.
    ‎
    Figure 6. A schematic diagram of metric M5. AF: atrial fibrillation; bpm: beats per minute; ECG: electrocardiogram; PPG: photoplethysmography; SR: sinus rhythm.

    Statistical Analysis

    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 (rs; Table 2)

    Table 2. Clinical significance of each metric.
    Performance metricFormulaClinical significance
    SensitivityTPa/(TP + FNb)Proportion of actual positive cases correctly identified
    SpecificityTNc/(TN + FPd)Proportion of actual negative cases correctly excluded
    Accuracy(TP + TN)/(TP + FP + FN + TN)Proportion of correctly classified cases
    PrecisionTP/(TP + FP)Proportion of TPs among all positive test results
    F1-score2 × TP/(2 × TP + FP + FN)Balances FNs and FPs when their clinical consequences are comparable
    AUC-ROCe—fOverall discriminative ability across all thresholds (0.9=excellent; 0.7‐0.9=moderate)
    MAEg∑i=1n|pi−gi|n,pi:PPGgi:ECGQuantifies the average prediction error; smaller MAE indicates higher clinical utility for risk stratification or treatment timing
    rsh—fMeasures monotonic association; high rs (rs>0.6) suggests that the model captures clinically relevant trends, even if they are nonlinear

    aTP: true positive.

    bFN: false negative.

    cTN: true negative.

    dFP: false positive.

    eAUC-ROC: area under the receiver operating characteristic curve.

    fNot applicable.

    gMAE: mean absolute error.

    hrs: Spearman rank correlation coefficient.

    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).


    Overview

    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).

    ‎
    Figure 7. A flowchart of the study. AF: atrial fibrillation; ECG: electrocardiogram; PPG: photoplethysmography; SR: sinus rhythm.

    Baseline Characteristics

    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.

    Table 3. Baseline characteristics of the participants (N=145).
    CharacteristicValues
    Demographics
     Age (y), mean (SD)63.28 (14.23)
     Male sex, n (%)96 (66.2)
    Medical history, n (%)
     Coronary artery disease31 (21.4)
     Heart failure4 (2.8)
     Hypertension75 (51.7)
     Hyperlipidemia66 (45.5)
     Diabetes mellitus31 (21.4)
     Previous stroke, SEa, or TIAb3 (2.1)
     Vascular disease47 (32.4)
     Renal dysfunction7 (4.8)
     COPDc9 (6.2)
     Hyperthyroidism3 (2.1)
     OSAd15 (10.3)
    Anticoagulants, n (%)
     Warfarin2 (1.4)
     Dabigatran14 (9.7)
     Rivaroxaban21 (14.5)
     Apixaban3 (2.1)
     Edoxaban24 (16.6)
     Low–molecular weight heparin3 (2.1)
    Antiarrhythmic drugs, n (%)
     Propafenone28 (19.3)
     Amiodarone16 (11.0)
     Dronedarone7 (4.8)
     Bisoprolol17 (11.7)
     Metoprolol39 (26.9)

    aSE: systemic arterial embolism.

    bTIA: transient ischemic attack.

    cCOPD: chronic obstructive pulmonary disease.

    dOSA: obstructive sleep apnea.

    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.

    Table 4. The performance metrics of the photoplethysmography-based atrial fibrillation burden model.
    Performance metricPerformance (%; 95% CI)
    Sensitivity90.0 (87.9‐95.1)
    Specificity97.2 (95.9‐98.5)
    Precision92.9 (88.6‐97.3)
    Accuracy93.3 (88.2‐98.5)
    F1-score90.5 (86.3‐94.7)
    AUCa89.5 (89.4‐89.7)

    aAUC: area under the receiver operating characteristic curve.

    ‎
    Figure 8. Receiver operating characteristic curves for photoplethysmography–based atrial fibrillation burden model. AUC: area under the curve; ROC: receiver operating characteristic.
    Table 5. The performance metrics of the photoplethysmography-based atrial fibrillation burden model across varying threshold values.
    Cutoff point for AF burden thresholdAccuracyPrecisionSensitivitySpecificityF1 score
    40%0.910.850.840.940.84
    45%0.920.850.840.940.85
    50%0.920.840.850.940.85
    55%0.920.840.860.940.85
    60%0.920.830.870.940.85
    65%0.920.830.870.940.85
    70%0.920.820.880.930.85
    ‎
    Figure 9. Representative pulse waveform recording from a patient. AF: atrial fibrillation; PPG: photoplethysmography; SR: sinus rhythm.

    Evaluation of AF Burden Quantification Performance by the AF Burden Model

    Metric M1 Evaluation

    For the ratio of AF duration to the total monitoring time, the PPG median was 0.0402 (IQR 0.0086‐0.8671) versus 0.0020 (IQR 0.0086‐0.8671) for 24-hour Holter monitoring, with model performance metrics of MAE=0.0400 (P=.008) and rs=0.8788 (P<.001; Figure 10A). For the ratio of AF episodes of >6 minutes to the total monitoring time, the PPG median was 0.0093 (IQR 0.0000‐0.7330) versus 0.0000 (IQR 0.0000‐1.0000) for 24-hour Holter monitoring, with model performance metrics of MAE=0.0582 (P=.44) and rs=0.9233 (P<.001; Figure 10B). Regarding the ratio of AF episodes of >1 hour to the total monitoring time, the PPG median was 0.0000 (IQR 0.0000‐0.2947) versus 0.0000 (IQR 0.0000‐1.0000) for 24-hour Holter monitoring, with model performance of MAE=0.1204 (P=.04) and rs=0.9293 (P<.001; Figure 10C).

    ‎
    Figure 10. Scatter plot and regression line of AF burden estimated by PPG compared to AF burden estimated by ECG. Each dot represents the AF burden by each method for each patient with AF. Dashed line is the regression line. Solid gray line is line of equality. AF: atrial fibrillation; ECG: electrocardiogram; PPG: photoplethysmography.
    Metric M2 Evaluation

    The ratio of AF episodes to the total measurements showed a PPG median of 0.0033 (IQR 0.0009‐0.0056) versus 0.0001 (IQR 0.0000‐0.0008) for 24-hour Holter monitoring, with model performance metrics of MAE=0.0320 (P<.001) and rs=–0.0807 (P=.33) compared to the gold standard. For the ratio of AF episodes of >6 minutes to the total measurements, the PPG median was 0.0010 (IQR 0.0007‐0.0028) versus 0.0000 (IQR 0.0000‐0.0007) for 24-hour Holter monitoring, with model performance metrics of MAE=0.0015 (P<.001) and rs=0.2360 (P=.004). Regarding the ratio of AF episodes of >1 hour to the total measurements, the PPG median was 0.0004 (IQR 0.0000‐0.0009) versus 0.0000 (IQR 0.0000‐0.0007) for 24-hour Holter monitoring, with model performance of MAE=0.0005 (P<.001) and rs=0.7092 (P<.001; Figure 11).

    ‎
    Figure 11. Scatter plot and regression line of AF burden estimated by PPG compared to AF burden estimated by ECG. Each dot represents the AF burden by each method for each patient with AF. Dashed line is the regression line. Solid gray line is line of equality. AF: atrial fibrillation; ECG: electrocardiogram; PPG: photoplethysmography.
    Metric M3 Evaluation

    The PPG-derived AF density showed a median of 0.2700 (IQR 0.1700-0.4725) compared to 0.0000 (IQR 0.0000-0.5900) for 24-hour Holter monitoring. When validated against the gold standard, the model demonstrated an MAE of 0.1725 (P<.001) with an rs of 0.6576 (P<.001; Figure 12)

    ‎
    Figure 12. Scatter plot and regression line of AF burden estimated by PPG compared to AF burden estimated by ECG. Each dot represents the AF burden by each method for each patient with AF. Dashed line is the regression line. Solid gray line is line of equality. AF: atrial fibrillation; ECG: electrocardiogram; PPG: photoplethysmography.
    Metric M4 Evaluation

    The SD of hourly AF episodes measured through PPG showed a median of 5.3229 (IQR 1.0375-15.9058) versus 0.4390 (IQR 0.0000-12.8558) for Holter monitoring, with MAE=5.6009 (P<.001) and rs=0.8135 (P<.001; Figure 13A). AF variability assessment revealed a median PPG value of 3.3478 (IQR 0.6957-9.2174) compared to 0.1739 (IQR 0.0000-5.2174) for 24-hour Holter monitoring, with MAE=3.9967 (P<.001) and rs=0.7876 (P<.001; Figure 13B). Daytime (6 AM-10 PM) variability showed a median PPG value of 3.6000 (IQR 0.8667-11.4667) versus 0.0000 (IQR 0.0000-5.600) for 24-hour Holter monitoring, with MAE=4.8425 (P<.001) and rs=0.7659 (P<.001; Figure 13C), whereas nighttime (10 PM-6 AM) variability showed a median PPG value of 1.000 (IQR 0.0000-6.0000) versus 0.0000 (IQR 0.0000-0.2857) for 24-hour Holter monitoring, with MAE=2.6171 (P<.001) and rs=0.6712 (P<.001; Figure 13D). Mean real variability measurements yielded a median value of 4.5282 (IQR 0.9031-12.4119) for PPG versus 0.3499 (IQR 0.0000-8.9325) for 24-hour Holter monitoring, with MAE=4.6436 (P<.001) and rs=0.8127 (P<.001; Figure 13E). Mean AF variation assessment yielded a median PPG value of 0.0466 (IQR –0.2987 to 1.3361) versus –0.0142 (IQR –0.1249 to 0.0654) for 24-hour Holter monitoring, with MAE=0.3893 (P=.27) and rs=0.7246 (P<.001; Figure 13F).

    ‎
    Figure 13. Scatter plot and regression line of AF burden estimated by PPG compared to AF burden estimated by ECG. Each dot represents the AF burden by each method for each patient with AF. Dashed line is the regression line. Solid gray line is line of equality. AF: atrial fibrillation; ECG: electrocardiogram; PPG: photoplethysmography. VOM: variation of the mean (VOM=[∣average AF episode at day−average AF episodes at night∣/average 24-h AF episodes] ×100%).
    Metric M5 Evaluation

    For the proportion of AF episodes with a ventricular rate of >120 bpm to total monitored episodes, PPG measurements showed a median of 0.0007 (IQR 0.0007-0.0027) compared to 0.0000 (IQR 0.0000-0.0070) for 24-hour Holter monitoring, with model performance metrics of MAE=0.0151 (P=.76) and rs=0.3435 (P<.001). Similarly, for the proportion of time in AF with a ventricular rate of >120 bpm to total monitoring time, the median PPG value was 0.0007 (IQR 0.0007-0.0027) versus 0.0000 (IQR 0.0000-0.0070) for 24-hour Holter monitoring, yielding an identical model performance (MAE=0.0151 and P=.76; rs=0.3435 and P<.001). Table 6 summarizes the quantification performance of the AF burden model.

    Table 6. Photoplethysmography (PPG)–monitored atrial fibrillation (AF) progression features compared to 24-hour electrocardiogram (ECG) monitoring (N=145).
    PPG, median (IQR)ECG, median (IQR)MAEaP valuersbP value
    Duration feature
    Ratio of AF duration to total monitoring time (%)0.0402 (0.0086 to 0.8671)0.0020 (0.0086 to 0.8671)0.0400.0080.8788<.001
    Ratio of AF episodes lasting over 6 min to total monitoring time (%)0.0093 (0.0000 to 0.7330)0.0000 (0.0000 to 1.0000)0.0582.440.9223<.001
    Ratio of AF episodes lasting over 1 h to total monitoring time (%)0.0000 (0.0000 to 0.2947)0.0000 (0.0000 to 1.0000)0.1204.040.9293<.001
    Number feature
    Ratio of AF episodes to total measurements (%)0.0033 (0.0009 to 0.0056)0.0001 (0.0000 to 0.0008)0.0032<.001–0.0807.33
    Ratio of AF episodes lasting over 6 min to total measurements (%)0.0010 (0.0007 to 0.0028)0.0000 (0.0000 to 0.0007)0.0015<.0010.2360.004
    Ratio of AF episodes lasting over 1 h to total measurements (%)0.0004 (0.0000 to 0.0009)0.0000 (0.0000 to 0.0007)0.0005<.0010.7092<.001
    Aggregation feature
    AF density0.2700 (0.1700 to 0.4725)0.0000 (0.0000 to 0.5900)0.1725.130.6576<.001
    Circadian rhythm feature
    SD per h of AF episodes5.3229 (1.0375 to 15.9058)0.4390 (0.0000 to 12.8558)5.6009<.0010.8135<.001
    Variability of AF episodes3.3478 (0.6957 to 9.2174)0.1739 (0.0000 to 5.2174)3.9967<.0010.7876<.001
    Daytime AF variability3.6000 (0.8667 to 11.4667)0.0000 (0.0000 to 5.6000)4.8425<.0010.7659<.001
    Nighttime AF variability1.0000 (0.0000 to 6.0000)0.0000 (0.0000 to 0.2857)2.6171<.0010.6712<.001
    Average real variabilityc4.5282 (0.9031 to 12.4119)0.3499 (0.0000 to 8.9325)4.6436<.0010.8127<.001
    VOMd of AF episodes (%)0.0466 (–0.2987 to 1.3361)–0.0142 (–0.1249 to 0.0654)0.3893.270.7246<.001
    Heart rate feature
    Ratio of duration of pulse rate of >120 bpme to monitoring time (min)0.0007 (0.0000 to 0.0027)0.0000 (0.0000 to 0.0070)0.0151.760.3435<.001
    Ratio of number of pulse rates of >120 bpm to total measurements0.0007 (0.0000 to 0.0027)0.0000 (0.0000 to 0.0070)0.0151.760.3435<.001

    aMAE: mean absolute error; mae=∑i=1n|pi−gi|npi:PPGgi:ECG.

    bSpearman rank correlation coefficient.

    cAverage real variability: average AF episodes, SD, and mean absolute deviation per hour were calculated over 24 hours; average real variability was then equal to the average SD and mean absolute deviation over 24 hours.

    dVOM: variation of the mean ([∣average AF episodes during the day − average AF episodes at night∣/average 24-hour AF episodes] × 100%; daytime AF variability: 6 AM-10 PM; nighttime AF variability: 10 PM-6 AM).

    ebpm: beats per minute.


    Principal Findings

    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 rs 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 rs 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 rs 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 rs of 0.8788 (P<.001) for the ratio of AF duration to total monitoring time and an MAE of 0.1204 (P=.04) and rs=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 rs, 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 rs=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.

    Limitations

    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.

    Conclusions

    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.

    Acknowledgments

    The authors thank the Huawei Heart Health Research Team for their technical support during the research process. They are grateful to Yu Cao, Hongli Zhang, Yubo You, Teng Xu, Shuai Zhao, and Jie Zhou (Huawei Heart Health Research Team) for the development and optimization of the photoplethysmography-based atrial fibrillation burden model. This study was funded by the National Natural Science Foundation of China (grant 82170309) and the Beijing Natural Science Foundation (grant L232117).

    Data Availability

    The datasets generated or analyzed during this study are available from the corresponding author on reasonable request.

    Authors' Contributions

    Hao Wang and YTG are joint corresponding authors and contributed equally to conceptualization, project administration, funding acquisition, resources, supervision, and writing—review and editing. Hong Wang played a role in the acquisition, analysis, and interpretation of the data. Hong Wang drafted the manuscript. Hong Wang, HZ, ZJ, BL, and ZZ performed data preprocessing. All authors approved the final version of the manuscript.

    Conflicts of Interest

    None declared.

    1. Van Gelder IC, Rienstra M, Bunting KV, et al. 2024 ESC Guidelines for the management of atrial fibrillation developed in collaboration with the European Association for Cardio-Thoracic Surgery (EACTS). Eur Heart J. Sep 29, 2024;45(36):3314-3414. [CrossRef] [Medline]
    2. Chew DS, Li Z, Steinberg BA, et al. Arrhythmic burden and the risk of cardiovascular outcomes in patients with paroxysmal atrial fibrillation and cardiac implanted electronic devices. Circ Arrhythm Electrophysiol. Feb 2022;15(2):e010304. [CrossRef] [Medline]
    3. Wong JA, Conen D, Van Gelder IC, et al. Progression of device-detected subclinical atrial fibrillation and the risk of heart failure. J Am Coll Cardiol. Jun 12, 2018;71(23):2603-2611. [CrossRef] [Medline]
    4. Go AS, Reynolds K, Yang J, et al. Association of burden of atrial fibrillation with risk of ischemic stroke in adults with paroxysmal atrial fibrillation: the KP-RHYTHM study. JAMA Cardiol. Jul 1, 2018;3(7):601-608. [CrossRef] [Medline]
    5. Charitos EI, Stierle U, Ziegler PD, et al. A comprehensive evaluation of rhythm monitoring strategies for the detection of atrial fibrillation recurrence: insights from 647 continuously monitored patients and implications for monitoring after therapeutic interventions. Circulation. Aug 14, 2012;126(7):806-814. [CrossRef] [Medline]
    6. Fan YY, Li YG, Li J, et al. Diagnostic performance of a smart device with photoplethysmography technology for atrial fibrillation detection: pilot study (pre-mAFA II registry). JMIR Mhealth Uhealth. Mar 5, 2019;7(3):e11437. [CrossRef] [Medline]
    7. Guo Y, Wang H, Zhang H, et al. Mobile photoplethysmographic technology to detect atrial fibrillation. J Am Coll Cardiol. Nov 12, 2019;74(19):2365-2375. [CrossRef] [Medline]
    8. Väliaho ES, Lipponen JA, Kuoppa P, et al. Continuous 24-h photoplethysmogram monitoring enables detection of atrial fibrillation. Front Physiol. Jan 4, 2022;12:778775. [CrossRef] [Medline]
    9. January CT, Wann LS, Calkins H, et al. 2019 AHA/ACC/HRS focused update of the 2014 AHA/ACC/HRS guideline for the management of patients with atrial fibrillation: a report of the American College of Cardiology/American Heart Association Task Force on Clinical Practice Guidelines and the Heart Rhythm Society in collaboration with the Society of Thoracic Surgeons. Circulation. Jul 9, 2019;140(2):e125-e151. [CrossRef] [Medline]
    10. Guo Y, Wang H, Zhang H, et al. Photoplethysmography-based machine learning approaches for atrial fibrillation prediction: a report from the Huawei heart study. JACC Asia. Dec 21, 2021;1(3):399-408. [CrossRef] [Medline]
    11. Tiver KD, Quah J, Lahiri A, Ganesan AN, McGavigan AD. Atrial fibrillation burden: an update-the need for a CHA2DS2-VASc-AFBurden score. Europace. May 21, 2021;23(5):665-673. [CrossRef] [Medline]
    12. Joglar JA, Chung MK, Armbruster AL, et al. 2023 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. Jan 2, 2024;149(1):e1-e156. [CrossRef] [Medline]
    13. Peacock J, Stanelle EJ, Johnson LC, et al. Using atrial fibrillation burden trends and machine learning to predict near-term risk of cardiovascular hospitalization. Circ Arrhythm Electrophysiol. Nov 2024;17(11):e012991. [CrossRef] [Medline]
    14. Passman RS. Monitoring for AF: identifying the burden of atrial fibrillation and assessing post-ablation. J Innov Card Rhythm Manag. Jan 15, 2017;8(1):2575-2582. [CrossRef] [Medline]
    15. Singer DE, Atlas SJ, Go AS, et al. Atrial fibrillation burden on a 14-day ECG monitor: findings from the GUARD-AF trial screening arm. JACC Clin Electrophysiol. Jan 2025;11(1):110-119. [CrossRef] [Medline]
    16. Charitos EI, Ziegler PD, Stierle U, et al. Atrial fibrillation burden estimates derived from intermittent rhythm monitoring are unreliable estimates of the true atrial fibrillation burden. Pacing Clin Electrophysiol. Sep 2014;37(9):1210-1218. [CrossRef] [Medline]
    17. Bonini N, Vitolo M, Imberti JF, et al. Mobile health technology in atrial fibrillation. Expert Rev Med Devices. Apr 2022;19(4):327-340. [CrossRef] [Medline]
    18. Chen E, Jiang J, Su R, et al. A new smart wristband equipped with an artificial intelligence algorithm to detect atrial fibrillation. Heart Rhythm. May 2020;17(5 Pt B):847-853. [CrossRef] [Medline]
    19. Tison GH, Sanchez JM, Ballinger B, et al. Passive detection of atrial fibrillation using a commercially available smartwatch. JAMA Cardiol. May 1, 2018;3(5):409-416. [CrossRef] [Medline]
    20. Sibomana O, Hakayuwa CM, Obianke A, Gahire H, Munyantore J, Chilala MM. Diagnostic accuracy of ECG smart chest patches versus PPG smartwatches for atrial fibrillation detection: a systematic review and meta-analysis. BMC Cardiovasc Disord. Feb 25, 2025;25(1):132. [CrossRef] [Medline]
    21. Guo Y, Guo J, Shi X, et al. Mobile health technology-supported atrial fibrillation screening and integrated care: a report from the mAFA-II trial Long-term Extension Cohort. Eur J Intern Med. Dec 2020;82:105-111. [CrossRef] [Medline]
    22. Avram R, Ramsis M, Cristal AD, et al. Validation of an algorithm for continuous monitoring of atrial fibrillation using a consumer smartwatch. Heart Rhythm. Sep 2021;18(9):1482-1490. [CrossRef] [Medline]
    23. Reissenberger P, Serfözö P, Piper D, et al. Determine atrial fibrillation burden with a photoplethysmographic mobile sensor: the atrial fibrillation burden trial: detection and quantification of episodes of atrial fibrillation using a cloud analytics service connected to a wearable with photoplethysmographic sensor. Eur Heart J Digit Health. Oct 6, 2023;4(5):402-410. [CrossRef] [Medline]
    24. Chugh SS, Havmoeller R, Narayanan K, et al. Worldwide epidemiology of atrial fibrillation: a Global Burden of Disease 2010 Study. Circulation. Feb 25, 2014;129(8):837-847. [CrossRef] [Medline]
    25. Körtl T, Stehle T, Riedl D, et al. Atrial fibrillation burden specifically determines human ventricular cellular remodeling. JACC Clin Electrophysiol. Nov 2022;8(11):1357-1366. [CrossRef] [Medline]
    26. Tanawuttiwat T, Lande J, Smeets P, et al. Atrial fibrillation burden and subsequent heart failure events in patients with cardiac resynchronization therapy devices. J Cardiovasc Electrophysiol. Jun 2020;31(6):1519-1526. [CrossRef] [Medline]
    27. Healey JS, Lopes RD, Granger CB, et al. Apixaban for stroke prevention in subclinical atrial fibrillation. N Engl J Med. Jan 11, 2024;390(2):107-117. [CrossRef] [Medline]
    28. Mark DB, Anstrom KJ, Sheng S, et al. Effect of catheter ablation vs medical therapy on quality of life among patients with atrial fibrillation: the CABANA randomized clinical trial. JAMA. Apr 2, 2019;321(13):1275-1285. [CrossRef] [Medline]
    29. Kuniss M, Pavlovic N, Velagic V, et al. Cryoballoon ablation vs. antiarrhythmic drugs: first-line therapy for patients with paroxysmal atrial fibrillation. Europace. Jul 18, 2021;23(7):1033-1041. [CrossRef] [Medline]
    30. Wazni OM, Dandamudi G, Sood N, et al. Cryoballoon ablation as initial therapy for atrial fibrillation. N Engl J Med. Jan 28, 2021;384(4):316-324. [CrossRef] [Medline]
    31. Drexler M, Blum T, Heinroth KM, et al. Heart rate variability as a predictor of successful catheter-guided pulmonary vein isolation for atrial fibrillation. Herz. Mar 2024;49(2):147-154. [CrossRef] [Medline]
    32. Zhu Z, Wang W, Cheng Y, Wang X, Sun J. The predictive value of heart rate variability indices tested in early period after radiofrequency catheter ablation for the recurrence of atrial fibrillation. J Cardiovasc Electrophysiol. Jun 2020;31(6):1350-1355. [CrossRef] [Medline]
    33. Linz D, Andrade JG, Arbelo E, et al. Longer and better lives for patients with atrial fibrillation: the 9th AFNET/EHRA consensus conference. Europace. Mar 30, 2024;26(4):euae070. [CrossRef] [Medline]

    ‎
    AF: atrial fibrillation
    bpm: beats per minute
    ECG: electrocardiogram
    MAE: mean absolute error
    mAFA: Mobile Atrial Fibrillation Application
    pNN50: percentage of successive normal sinus intervals that differ by more than 50 milliseconds
    PPG: photoplethysmography
    RVR: rapid ventricular rate

    Edited by Kam Cheong Wong; submitted 26.May.2025; peer-reviewed by Fan Lin, Giulio Romiti, Po-Tseng Lee; final revised version received 23.Aug.2025; accepted 04.Sep.2025; published 10.Nov.2025.

    Copyright

    © Hong Wang, Binbin Liu, Hui Zhang, Zheqi Zhang, Zhigeng Jin, Hao Wang, Yu-Tao Guo. Originally published in JMIR Cardio (https://cardio.jmir.org), 10.Nov.2025.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Cardio, is properly cited. The complete bibliographic information, a link to the original publication on https://cardio.jmir.org, as well as this copyright and license information must be included.