Authors: Steven A. Lubitz (1Cardiac Arrhythmia Service & Cardiovascular Research Center, Massachusetts General Hospital;; 2Harvard Medical School, Boston, MA;), Michael V. McConnell (3Fitbit LLC (Google LLC), San Francisco;; 4Division of Cardiovascular Medicine, Stanford University School of Medicine, Stanford, CA;), Caitlin Selvaggi (5Biostatistics Center, Massachusetts General Hospital;), Aparna Krishnamoorthy (5Biostatistics Center, Massachusetts General Hospital;), Steven J. Atlas (2Harvard Medical School, Boston, MA;; 6Division of General Internal Medicine, Massachusetts General Hospital;), David D. McManus (7Division of Cardiovascular Medicine, Department of Medicine, University of Massachusetts Chan Medical School, Worcester, MA;), Sherry Pagoto (8Department of Allied Health Sciences, University of Connecticut, Storrs, CT), Daniel E. Singer (2Harvard Medical School, Boston, MA;; 6Division of General Internal Medicine, Massachusetts General Hospital;), Alexandros Pantelopoulos (3Fitbit LLC (Google LLC), San Francisco;), Andrea S. Foulkes (2Harvard Medical School, Boston, MA;; 5Biostatistics Center, Massachusetts General Hospital;), Anthony Z. Faranesh (3Fitbit LLC (Google LLC), San Francisco;)
Categories: Article, atrial fibrillation, wearable, screening, ambulatory electrocardiography
Source: Circulation. Arrhythmia and electrophysiology
Authors: Steven A. Lubitz, Michael V. McConnell, Caitlin Selvaggi, Aparna Krishnamoorthy, Steven J. Atlas, David D. McManus, Sherry Pagoto, Daniel E. Singer, Alexandros Pantelopoulos, Andrea S. Foulkes, Anthony Z. Faranesh
Wrist-worn wearables can detect irregular heart rhythms using photoplethysmography, but electrocardiograms are required to confirm atrial fibrillation (AFib). We sought to determine the frequency of a recurrent irregular heart rhythm detection (IHRD; ≥30 minutes of an irregular rhythm), estimate the potential diagnostic yield of different electrocardiographic monitoring strategies for confirming AFib, and identify predictors of recurrent IHRDs.
The Fitbit Heart Study enrolled wrist-worn photoplethysmography device users without diagnosed AFib. Of 455,699 participants, 1,057 who wore and returned a 1-week electrocardiogram patch monitor after receiving an IHRD were analyzed. Baseline clinical data, device-derived metrics, IHRDs during follow-up, and electrocardiographic patch data were used for analysis.
570 (53.9%) participants were aged 40–64 years, 422 (39.9%) were ≥65 years, and 510 (48.2%) female. Median follow-up after electrocardiogram patch initiation was 80 days (interquartile range 45–122 days). The frequency of another IHRD was 57.2% (95%CI 53.1%–60.9%) at three months. After an initial IHRD, the estimated diagnostic yield for AFib with a 10-second electrocardiogram was 7.6% (95%CI 6.2%–9.0%), twice daily 30-second electrocardiograms over 1-week 19.0% (95%CI 16.7%–21.2%), 24-hour monitor 17.4% (95%CI 15.5%–19.3%), 1-week monitor 32.2% (95%CI 29.4%–35%.0), 2-week monitor 46.8% (95%CI 42.7%–50.8%), and 4-week monitor 60.8% (95%CI 56.5%–65.1%). Risk of a recurrent IHRD was greater with older age (P<0.001), male sex (P=0.001), vascular disease (P=0.03), longer initial runs of consecutive IHRDs at detection (P=0.02), and less nightly sleep (P=0.03).
Irregular heart rhythms are common after initial detection using a wrist-worn wearable device. Longer electrocardiographic monitoring periods increase the likelihood of confirming AFib.
clinicaltrials.gov NCT04380415
Undiagnosed atrial fibrillation (AFib) is often episodic^1^ and can be morbid.^2,3^ Wearable devices with photoplethysmography (PPG) sensors can detect undiagnosed AFib using software algorithms that assess pulse waveforms for irregularity, and generate irregular heart rhythm detections (IHRDs). Confirmation of AFib using electrocardiography (ECGs) may enable stroke prevention efforts and other interventions that minimize morbidity.
In two recent studies utilizing wrist-worn wearable devices equipped with PPG sensors, 30% of participants that received an IHRD had AFib confirmed by a subsequent 1-week ECG patch.^4,5^ Longer durations of ECG monitoring after an IHRD may result in a higher rate of ECG confirmation of AFib, but the exact relationship between duration of ECG monitoring and the likelihood of confirming AFib after an IHRD is not well known.
The primary objective of the study was to assess the frequency of a recurrent IHRD among wearable users and estimate the potential diagnostic yield of different durations of ECG monitoring for confirmation of AFib. We further explored factors associated with a recurrent IHRD.
Individual level data from the study will not be made available due to participant confidentiality and privacy, as well as company policy regarding user data.
We used data collected from participants in the Fitbit Heart Study, the design and primary results of which have been reported previously.^5,6^ In brief, the Fitbit Heart Study was a single-arm remote clinical trial that demonstrated that an IHRD notification had a 98% positive predictive value for concurrent AFib confirmed on a 1-week ECG patch. Fitbit device users aged at least 22 years were invited to enroll between May 6 and October 1, 2020. Participants were eligible if they were U.S. residents and used a compatible Fitbit device (Ionic, Charge 3, Charge 4, Versa, Versa Lite, Versa 2, Versa 3, Sense, Inspire HR, Inspire 2) with a paired Fitbit account and an Android or Apple iOS smartphone with installed Fitbit app. Participants were required to confirm that they had not been diagnosed with AFib, used oral anticoagulants, or had a pacemaker or defibrillator. All participants provided informed consent via smartphone or web. The protocol was approved by the Advarra Institutional Review Board (Columbia, MD). Of the 455,699 individuals enrolled, 4,728 had an IHRD, of whom 1,057 were included in the ECG analysis set, which is the subject of this report. The median time from IHRD to ECG patch monitoring was 18.6 days (interquartile range [IQR] 10.9–30.7 days).
A novel algorithm developed by Fitbit ran centrally on a server using routinely collected PPG data after wearable device syncing via smartphone. The algorithm analyzed PPG data continuously during periods in which participants were stationary (as determined by device accelerometers) to minimize motion artifact and false positive AFib detections. The PPG data were analyzed as 5-minute pulse tachograms acquired every 2.5 minutes (i.e., overlapping by 50%). If 11 consecutive tachograms were classified as irregular, then an IHRD was generated. The algorithm therefore required at least 30 minutes of a sensed irregular heart rhythm to generate an IHRD. The algorithm previously was reported to have a PPV of 98.2% for AF confirmed concurrently on an ECG patch among individuals who had a prior IHRD, as well as specificity of 98.4% and sensitivity of 67.6% for an IHRD during a 1-week ECG patch monitoring period.^5^ The algorithm was the same across devices. Longitudinal PPG data were collected centrally for 30 days prior to and then throughout the trial, enabling analysis of physiologic characteristics (see below) and recurrent IHRDs throughout the study period.
At enrollment, participants were asked to report their age, sex, race, ethnicity, height, weight, smoking status, alcohol intake, and medical comorbidities including histories of congestive heart failure, hypertension, stroke, vascular disease, diabetes, sleep apnea, and family history of atrial fibrillation. The CHA2DS2-VASc score was calculated by assigning two points each for age ≥ 75 years or a prior stroke, and one point each for a history of congestive heart failure, hypertension, diabetes, age 65–74 years, or female sex.^7^ We computed the body mass index (BMI) as the weight (in kilograms) divided by the height (in meters) squared.
Data gathered from the wearable devices included continuous heart rate, resting heart rate, activity, and sleep. In brief, optical photoplethysmography sensors on the Fitbit device detect changes in blood volume over time in the capillaries, from which the pulse rate is derived. The detected pulse rate was processed and averaged to produce an estimate of heart rate. Resting heart rate was determined from the heart rate signal during periods of sleep and stillness during the day, as determined by the on-device accelerometer. Activity intensity was calculated from the heart rate estimate during exercise. The percentage of theoretical maximum heart rate was used to calculate the activity intensity, which is personalized based on individual age and resting heart rate. Duration of moderate to vigorous physical activity, defined to be the equivalent of ≥ 3METS, was recorded. Periods of sleep were determined by the on-device accelerometer. Software quantified sleep duration and classified sleep stages using accelerometer-detected motion, heart rate variability, and respiration rate. Fitbit software algorithms that estimated heart rate, activity duration, and sleep duration are considered proprietary. Since the IHRD algorithm does not analyze for irregularity during periods of activity, we also ascertained the algorithm analyzable time (i.e., time when the participant was at rest). During episodes of IHRDs, we further ascertained the episode duration and heart rate.
Participants with an IHRD were notified via the Fitbit app and instructed to schedule a telehealth visit with a PlushCare (San Francisco, CA) physician via the app. The telehealth physician collected medical history, assessed for symptoms and adverse events, and confirmed eligibility. Participants were then mailed a single-lead ECG patch monitor (ePatch, BioTelemetry, Inc., Malvern PA) and instructed to apply and wear the monitor for one week before returning it via mail using pre-paid packaging.
The primary objective of the study was to assess the frequency of a recurrent IHRD among wearable users and to estimate the potential diagnostic yield of different durations of ECG monitoring for confirmation of AFib. We further explored factors associated with a recurrent IHRD. First, we estimated the frequency of recurrent IHRDs after initiation of ECG patch monitoring to serve as a proxy for a recurrent arrhythmia that could be detected with long-term continuous rhythm monitoring. Second, we estimated the potential yield of diagnosing AFib under various clinically relevant hypothetical ECG monitoring scenarios. Third, we modeled factors associated with recurrent IHRDs among all individual and those with non-persistent AF. The methodology for each part of the analysis is provided below and summarized in Figure 1.
We assessed the proportion of participants with a first recurrent IHRD during the study period among participants who wore and returned an ECG patch monitor using the Kaplan-Meier survival function, expressed as a cumulative incidence. Person-time began at the start of ECG patch monitoring and the end (event time) was defined as the minimum of either a recurrent IHRD or the end of the study period which occurred on October 26, 2020 (the last day participant data was run through the IHRD algorithm). We stratified analyses by age and sex. The start of ECG patch monitoring was selected as the start of person-time for follow-up for a first recurrent IHRD to provide a clinically relevant starting point to assess the yield of various durations of confirmatory ECG monitoring (see further methods below).
We projected the yield of different confirmatory ECG monitoring durations and strategies following an initial IHRD that the authors deemed clinically relevant. Selected scenarios
To estimate the impact of confirmatory ECG monitoring, we sampled probabilities of AFib detection during the 1-week ECG patch monitoring period (scenarios 1–4), or extrapolated the probability of a first recurrent IHRD with adjustment for the incomplete sensitivity of an IHRD for AFib confirmed on an ECG (scenarios 5 and 6).
For scenarios (1), (2), and (3), we used the timing of AFib episodes from the ECG patch reports as the reference. For each sampling strategy, we computed the probability of a positive test given the total possible number of test opportunities during the ECG monitoring period. For example, for the 10-second ECG between 8am and 5pm, the number of possible tests is the number of seconds (minus 10 seconds) between 8am and 5pm, per day which overlap with the patch monitoring period. The number of possible positive tests were computed, and the ratio of the positive tests to the number of test opportunities was taken as the likelihood of detection. A similar approach was used for the 24-hour Holter monitor scenario. For the twice daily 30 second recordings, the cumulative probability of no detection across all days was first computed, and the probability of detection was calculated as 1 – Pr(no detection across all days). For scenario (4) we used the frequency of AFib observed using the one-week ECG patch monitor for the trial. For scenario (5) and (6) we used the cumulative incidence of recurrent IHRDs at 2 and 4 weeks following the start of ECG patch monitoring, respectively. We then multiplied the point-estimate by the ratio of ECG-confirmed AFib to frequency of observed IHRDs during the one-week ECG patch monitor (340/241=1.4) to infer the projected ECG-confirmed frequency of AFib, to account for the incomplete sensitivity of PPG-based IHRDs for detecting recurrent AFib events.^5^
The sample size for scenarios (1) - (4) was 1,057, the number of subjects included in the ECG analysis set from the trial. The sample size for scenarios (5) and (6) was 1,009, the number of participants who had complete data for the survival analysis. For scenarios (1) – (3), the standard error (SE) was computed as SE=SD/N, where SD is the standard deviation of the detection probabilities over the participants. For scenarios (4) – (6) the standard error of the probability of IHRD recurrence, p, was computed from the variance of the sample proportion SE=p⋅1-p/N. For scenarios (5) and (6), the SE was multiplied by 1.4 to estimate the SE of the probability of ECG confirmed AFib recurrence. The 95% confidence intervals were computed as +/− 1.96*SE.
We identified risk factors associated with recurrent IHRDs by fitting proportional hazards models in which we regressed the time to a recurrent IHRD on selected factors. Factors included participant age, sex, race/ethnicity (categorized as White, Black or African American, Hispanic, or other/prefer not to say), body mass index (categorized as normal [<25 kg/m^2^], overweight [25-<30 kg/m^2^], or obese [≥30 kg/m^2^]), smoking status (categorized as daily, less than daily, or none), alcohol intake (categorized as daily, less than daily, or none), congestive heart failure, hypertension, stroke, vascular disease, diabetes, sleep apnea, and first degree family history of AFib.
We hypothesized that participant characteristics such as heart rate, activity, and sleep – all of which are ascertained by the Fitbit device – may be associated with recurrent IHRDs. We therefore defined time-averaged daily variables for these characteristics during the 7-day exposure window prior to the initial IHRD. We also hypothesized that some heart rate factors temporally related to the initial IHRD would be associated with recurrent IHRDs, and therefore derived several IHRD-associated characteristics from the Fitbit device during the 24-hour window following the initial IHRD detection but prior to ECG monitoring including the maximum IHRD episode (defined as the longest duration of continuous irregular tachograms), number of IHRD episodes, and cumulative total time for all IHRDs within the 24-hour window. For analyses examining factors associated with recurrent IRHDs, we fit multivariable models adjusted for all variables. Models were adjusted for the analyzable time (in hours/day) during the 1-week ECG monitoring period. Individuals with missing data, or with BMI less than 12 kg/m^2^ were excluded.
We also explored associations between risk factors and AFib burden and longest AFib episode on the 1-week ECG patch monitor, among individuals with AFib detected and without persistent AFib throughout the monitoring period using multiple linear regression. AFib burden was defined as the percent time monitored that was spent in AFib. AFib was defined as at least 30 seconds of arrhythmia.
All hypothesis tests were two-sided. R version 4.3.1 was used for statistical analysis. Simulations of different monitoring strategies were done in Python, version 3.
Of the 1,057 participants eligible for analysis (Figure 2), 570 (53.9%) were aged 40–64 years and 422 (39.9%) were aged ≥65 years, 510 (48.2%) were female, and 906 (85.7%) were non-Hispanic White (Table 1). The median duration of follow-up following initiation of ECG patch monitoring was 80 days (interquartile range 45–122 days). The median hours of Fitbit device wear time per day was 23 (IQR, 23–24). The fraction of days at risk with ≥ 18 hours of Fitbit device wear time was 79% (IQR, 65%–100%).
The cumulative incidence of a recurrent IHRD after initiation of ECG patch monitoring was 25.0% (95% CI 22.3%-27.7%) at 1 week, 33.4% (95% CI 30.3%-36.2%) at 2 weeks, 43.4% (95% CI 40.1%-46.5%) at 4 weeks, and 57.2% (95% CI 53.1%-60.9%) at three months (Figure 3).
The diagnostic yield for AFib varied according to the duration of time monitored. Figure 4 shows the estimated diagnostic yield for AFib according to different monitoring scenarios. The estimated diagnostic yield for AFib with a 10-second electrocardiogram after an initial IHRD was 7.6% (95% CI 6.2% - 9.0%), twice daily 30-second electrocardiograms over 1-week 19.0% (95% CI 16.7% - 21.2%), 24-hour continuous monitoring 17.4% (95% CI 15.5% - 19.3%), 1-week continuous monitoring 32.2% (95% CI 29.4% - 35%.0), 2-week continuous monitoring 46.8% (95% CI 42.7% - 50.8%), and 4-week continuous monitoring 60.8% (95% CI 56.5% - 65.1%).
Factors associated with a recurrent IHRD are displayed in Table 2. In multivariable adjusted models, older age, male sex, vascular disease, and a longer run of consecutive IHRDs on the initial day of detection were associated with increased likelihood of a recurrent IHRD. More nightly sleep was associated with a reduced likelihood of a recurrent IHRD.
Among 340 participants with AFib detected on the ECG patch monitor, 285 had non-persistent AFib, of which 268 had complete data for analysis (Supplemental Figure 1). The median AFib burden as determined by the ECG patch monitor was 6.9% (IQR 1.9%-28.8%) overall, and 4.8% (IQR 1.5%-12.7%) among those with non-persistent AFib. Characteristics of participants with non-persistent AFib are displayed in Supplemental Table 1. Factors associated with AFib burden and longest episode duration among 268 individuals with non-persistent AFib are displayed in Supplemental Table 2 and Supplemental Table 3. In multivariable adjusted models, vascular disease, and longer run of consecutive IHRDs on the initial day of detection were each associated with both greater AFib burden and longer episodes of AFib detected on the ECG patch monitor.
In this analysis of 1,057 participants with wearable-based IHRDs who wore and returned subsequent 1-week ECG patch monitors, we observed that the frequency of recurrent IHRDs was high at nearly 60% over three months. We project that prolonged ECG monitoring of either two or four weeks would substantially increase the likelihood of AFib confirmation beyond the 1-week monitors used in the Fitbit Heart Study. Predictors of recurrent IHRDs included older age, male sex, vascular disease, and longer duration of an IHRD on the day of the initial detection, whereas greater average duration of sleep during the week prior to the initial IHRD was associated with a reduced likelihood of another IHRD.
In both the Apple Heart Study and Fitbit Heart Study, which assessed wrist-worn wearable PPG devices for the detection of irregular heart rhythms, 1-week ECG patch monitors were used.^4,5^ In the Apple Heart Study, AFib was confirmed by ECG patch in 34% of participants with a prior irregular rhythm notification who wore and returned the patch.^4^ The positive predictive value of the irregular rhythm notification feature was 84%. These findings are comparable to those of the Fitbit Heart Study, in which AFib was confirmed in 32% of participants with an IHRD, and the IHRD algorithm had a positive predictive value of 98%. To date no specific recommendations regarding the duration or approach to monitoring following a consumer-based wrist-worn wearable have been provided. Our findings address the recurrence rates of IHRDs and estimate the effects of varying durations of ECG monitoring for confirmation of AFib among individuals with smartwatch detected irregular rhythms. In a prior simulation analysis of 590 older participants with implantable loop recorders to screen for AFib, detection of recurrent episodes was highly dependent on the duration and frequency of screening.^8^
Our findings have several implications. First, IHRDs are likely to recur in individuals that receive an initial detection. As such, IHRDs are unlikely to reflect isolated occurrences of AFib. Recurrent IHRDs in individuals without a history of AFib warrant clinical evaluation. It is possible that the high positive predictive value of consumer-grade algorithms for detecting irregular rhythms, such as the Fitbit algorithm, may select for individuals with highly frequent paroxysmal AF that may be clinically relevant. Further work is required to systematically assess characteristics of AFib detected across consumer-grade PPG-based devices in real-world populations, such as episode frequency, duration, and density.^9^
Second, longer ECG monitoring is likely to confirm AFib in a greater proportion of individuals. Considering that most individuals from the Fitbit Heart Study with an IHRD had a recurrence during the relatively short follow-up period, the absence of AFib on a single confirmatory 1-week ECG patch monitor may underestimate the frequency of AFib, perhaps due to the paroxysmal nature of the arrhythmia. Single 30-second “spot checks” may be a practical form of ECG confirmation of AFib, though have substantially lower yield than a 1-week ECG patch monitor, and using such ECG strips for clinical decision making such as initiating anticoagulation requires additional study. Moreover, PPG algorithms do not necessarily alert instantaneously when an irregular pulse is detected, notifications may occur during sleep, and many devices do not have ECG capabilities, rendering immediate on-device rhythm confirmation impractical in some circumstances. Longer forms of continuous ECG monitoring, such as commonly used 2- and 4-week noninvasive monitors, are expected to have a substantially higher diagnostic yield for AFib than a 1-week ECG patch and as such are intuitively favored to confirm AFib. However, the efficacy of longer monitoring requires prospective validation, risks of incidental detection of rhythm abnormalities warrant evaluation, and the cost effectiveness of using such approaches as the initial AFib confirmatory strategy following an IHRD are unknown.
The choice regarding approach and duration of monitoring may be individualized. For example, patients at highest risk of adverse outcomes from AFib, including those at high risk of stroke such as those with a history of vascular disease, may warrant longer or more frequent confirmatory ECG monitoring. Patients with palpitations or other symptoms may be able to leverage consumer-based ECG capabilities to capture the cardiac rhythm when they are symptomatic. Practical factors including availability of specific monitoring technology, patient tolerability to adhesives, and monitor cost may also guide the selection of approaches in specific circumstances.
Third, several specific characteristics were associated with risk for recurrent IHRDs. We observed that older age, male sex, and vascular disease were associated with IHRD recurrences – notably these are established risk factors for incident AFib.^10^ Greater sleep during the week prior to an IHRD, a potentially modifiable risk factor, was associated with a slightly lower risk of a recurrent IHRD. Moreover, characteristics of the initial detection, including a longer period of an irregular pulse were also associated with increased burden of AFib. AFib burden has been linked to increased morbidity.^11^ In aggregate, these exploratory and hypothesis-generating findings imply that it may be possible to use a variety of features available at the time of an initial IHRD to identify individuals at greatest risk for recurrent arrhythmias or those with a higher burden of AFib. Future studies examining whether deliberate sleep modification can alter risk for recurrent IHRDs or AFib are warranted.
Our results should be interpreted in the context of the study design. The Fitbit Heart Study included existing Fitbit smartwatch and fitness tracker users which comprise a predominantly non-Hispanic White population and represent a select group of individuals who agreed to participate in a clinical trial. Prospective evaluation in diverse racial and ethnic backgrounds may enhance the generalizability of our findings. Projected yields of AFib confirmation with different forms of ECG monitoring were modeled based on IHRDs, and therefore misclassification of cardiac rhythm may limit conclusions. However we submit that the high positive predictive value of an IHRD among subjects with a prior irregular rhythm detection (~98%) justifies the use of IHRDs as a valid surrogate for AFib in our analyses.^5^ It is possible that our analyses underestimate the frequency of recurrent AFib due to shorter episodes, considering that at least 30 minutes of an irregular rhythm is needed to trigger an IHRD. We did not simulate the projected diagnostic yield for different confirmatory ECG monitoring approaches for other consumer wearable devices. The Fitbit IHRD algorithm is a specific algorithm and may selectively identify participants with a high probability of intermediate-term arrhythmia recurrence and with a high burden of AFib; validation of our observations of high IHRD recurrence rates in other devices is warranted. Moreover, participants in the Fitbit Heart Study do not represent patients observed in a clinical practice, and as such the absolute risk estimates reported in this study likely differ from patients in practice. The study relied on self-reported historical characteristics from participants, which may introduce bias; nevertheless we submit that this extends the real-world generalizability of the findings. The scenarios modeled in our analysis were primarily based on an initial confirmatory strategy and did not address long-term strategies for AFib confirmation, such as periodic ECG patch monitors or implantable loop recorders, which may represent suitable alternatives to the confirmatory scenarios modeled in our analysis. We also note that to date, firm prospective data demonstrating that screening for AF reduces cardiovascular morbidity are lacking.
In an analysis of 1,057 participants with IHRDs from the Fitbit Heart Study, most had a recurrent IHRD, suggesting that initial IHRDs do not reflect isolated episodes of AFib. Longer periods of initial confirmatory ECG monitoring are likely to increase the diagnostic yield for AFib detection. Fixed clinical factors, as well as sleep, were associated with variable risks of recurrent IHRDs over an intermediate time horizon.
Tables S1–S3
Figure S1