Author + information
- Received August 27, 2018
- Revision received October 1, 2018
- Accepted October 11, 2018
- Published online February 18, 2019.
- Marcus Dörr, MDa,b,∗ (, )
- Vivien Nohturfft, MSca,
- Noé Brasier, MScc,
- Emil Bosshard, BScc,
- Aleksandar Djurdjevic, MScc,
- Stefan Gross, PhDa,b,
- Christina J. Raichle, MDc,
- Mattias Rhinisperger, BScc,
- Raphael Stöckli, BScc and
- Jens Eckstein, MD, PhDc,d
- aDepartment of Internal Medicine B, University Medicine Greifswald, Greifswald, Germany
- bGerman Centre for Cardiovascular Research (DZHK), partner site Greifswald, Greifswald, Germany
- cChief Medical Information Officer (CMIO) Office, University Hospital Basel, Basel, Switzerland
- dDepartment of Internal Medicine, University Hospital Basel, Basel, Switzerland
- ↵∗Address for correspondence:
Dr. Marcus Dörr, Department of Internal Medicine B, University Medicine Greifswald, Ferdinand-Sauerbruch-Strasse, 17475 Greifswald, Germany.
Objectives The WATCH AF (SmartWATCHes for Detection of Atrial Fibrillation) trial compared the diagnostic accuracy to detect atrial fibrillation (AF) by a smartwatch-based algorithm using photoplethysmographic (PPG) signals with cardiologists’ diagnosis by electrocardiography (ECG).
Background Timely detection of AF is crucial for stroke prevention.
Methods In this prospective, 2-center, case-control trial, a PPG pulse wave recording using a commercially available smartwatch was obtained along with Internet-enabled mobile ECG in 672 hospitalized subjects. PPG recordings were analyzed by a novel automated algorithm. Cardiologists’ diagnoses were available for 650 subjects, although 142 (21.8%) datasets were not suitable for PPG analysis, among them 101 (15.1%) that were also not interpretable by the automated Internet-enabled mobile ECG algorithm, resulting in a sample size of 508 subjects (mean age 76.4 years, 225 women, 237 with AF) for the main analyses.
Results For the PPG algorithm, we found a sensitivity of 93.7% (95% confidence interval [CI]: 89.8% to 96.4%), a specificity of 98.2% (95% CI: 95.8% to 99.4%), and 96.1% accuracy (95% CI: 94.0% to 97.5%) to detect AF.
Conclusions The results of the WATCH AF trial suggest that detection of AF using a commercially available smartwatch is in principle feasible, with very high diagnostic accuracy. Applicability of the tested algorithm is currently limited by a high dropout rate as a result of insufficient signal quality. Thus, achieving sufficient signal quality remains challenging, but real-time signal quality checks are expected to improve signal quality. Whether smartwatches may be useful complementary tools for convenient long-term AF screening in selected at-risk patients must be evaluated in larger population-based samples. (SmartWATCHes for Detection of Atrial Fibrillation [WATCH AF]:; NCT02956343)
Atrial fibrillation (AF) is the most common cardiac arrhythmia. Its prevalence is even expected to double within the next 40 years in adults 55 years old or older (1,2). AF has a tremendous economic impact because of its strong association with increased morbidity, particularly an up to 5-fold higher risk for stroke, and with higher mortality (3–5).
A diagnostic challenge arises from the issue that AF is often asymptomatic and frequently remains undetected until the first thromboembolic event occurs (2,6). Higher detection rates and earlier identification of unknown AF could contribute to reduce the risk of stroke and other consequences by initiating appropriate therapeutic measures. Enhanced and prolonged screening efforts have been demonstrated to lead to markedly increased detection rates. Thus, the use of external rhythm recording devices, such as Holter electrocardiography (ECG) monitors or external loop recorders (7–14), was associated with a more than 5-fold higher incidence of AF after 30-day screening in patients with cryptogenic stroke in the EMBRACE (30-Day Cardiac Event Monitor Belt for Recording Atrial Fibrillation After a Cerebral Ischemic Event) trial (14). Invasive cardiac monitoring by insertable leadless cardiac monitors is related to an even higher AF detection rate and therefore is considered the gold standard in high-risk patients (7–10,12,13). In the CRYSTAL-AF trial (Study of Continuous Cardiac Monitoring to Assess Atrial Fibrillation After Cryptogenic Stroke), this resulted in a 6-fold higher AF incidence after 36 months compared with Holter ECG after cryptogenic stroke (12,15). However, because of inconvenience, lack of reimbursement, or other reasons, these methods are not applicable to all subjects who have suspected AF. Reliable, cost-effective, convenient, and easy-to-apply tools for extended noninvasive AF detection would be very helpful the fill the remaining gap. Very promising are photoplethysmography (PPG)–based smartphone applications that enable rhythm analysis using the custom built-in camera, and they have been shown to identify AF with high accuracy (13,16–23). This method has recently been added to the AF screening recommendations of the European Heart Rhythm Association along with other technologies (11). The newest generation of smartwatches is equipped with PPG sensors that can be used for the same purpose. Using such technology could potentially raise rhythm screening to the next level as compared with previous approaches with smartphones by enabling convenient long-term screening or extended noninvasive rhythm analysis in individuals who are not suitable for other screening methods (24).
The aim of the prospective, 2-center, case-controlled WATCH AF (SmartWATCHes for Detection of Atrial Fibrillation) trial was to compare the diagnostic performance of a novel PPG-based algorithm to detect AF using a commercially available smartwatch with cardiologists’ diagnosis that was based on Internet-enabled mobile ECG (iECG) as a proof of principle.
Study design and participants
WATCH AF was an investigator-initiated, prospective, case-control trial, conducted at 2 university centers (University Medicine Greifswald, Germany and University Hospital Basel, Switzerland). Independent data monitoring was provided by the Clinical Trial Unit in Basel. The study was conducted in compliance with the Declaration of Helsinki. Ethics approval was obtained from the local ethics committees (BB 141/16, EKNZ BASEC 2016-01175). All participants provided written informed consent. Inpatients 18 years old or older, without pacemakers or implanted defibrillators, were eligible for the study. On the basis of electronic patients’ records, candidates were screened for a recent history of AF. If this diagnosis was present, they were recruited after giving informed consent. Age- and sex-matched patients without a history of AF in their medical records were recruited as potential matches for the sinus rhythm (SR) group. It was considered that some of the patients with a history of AF were likely to be in SR at the time of recruitment. Because the iECG recordings were analyzed in a blinded fashion, the final allocation of the patients to the respective groups was possible only after recruitment was closed and all data were analyzed. This design allowed a prospective study character, as well as evaluation of test criteria.
Data acquisition and measurements
Data on patients’ characteristics, including medical histories and medication, were obtained from electronic patients’ records and personal interviews.
Study measurements were performed in a quiet surrounding with the subject in a comfortable sitting position. The following 5-min recordings were obtained in 1 session (Online Figure 1):
1. A PPG recording on 1 wrist using the built-in sensors of a smartwatch (Gear Fit 2, Samsung Electronics Co., Ltd., Seoul, South Korea) with a study version of the Heartbeats application (Preventicus GmbH, Jena, Germany). This app was used to record PPG signals without any feedback for the study personnel regarding rhythm analysis results. Raw data files were saved on a paired smartphone (Samsung S5 mini, Samsung Electronics Co.) and transferred to a central server.
2. A single-lead iECG using the AliveCor Kardia system (AliveCor Inc., Mountain View, California) paired with an iPhone 4s (Apple Inc., Cupertino, California) with the left index and middle fingers on the left electrode and the right index and middle fingers on the right electrode of the device. All iECG recordings were saved in PDF format and served as references in our study. This method has been recommended for AF screening by the European Heart Rhythm Association and has been approved by the U.S. Food and Drug Administration for AF screening (11).
3. A second PPG recording on the other wrist using a Wavelet wrist band (Wavelet Health, Mountain View, California) paired with a Wi-Fi–enabled iPad mini (Apple Inc.). Raw data files were automatically sent to a company server but finally could not be analyzed because of rejection by the company.
Processing of PPG recordings
Pseudonymized smartwatch-derived PPG files were analyzed by Preventicus, blinded for all other data, including cardiologists’ and automated iECG diagnosis. As predefined, cohesive 1-, 3-, and 5-min segments were extracted from each patient’s recording on the basis of the best signal-to-noise ratio. These segments were subsequently automatically evaluated using the Heartbeats PPG algorithm (version 20171120), which was developed for analysis of PPG signals recorded by PPG sensors of smartwatches or wearables (16). In brief, this algorithm differentiates between SR and AF by using a complex nonlinear combination analysis comprising beat-to-beat changes of pulse wave time intervals (16,17). Rhythm analysis is preceded by an automated data quality check on the basis of signal-to-noise ratio and accelerometer data. Disturbed or noisy PPG data segments (e.g., secondary to motion artifacts) are thereby automatically excluded by the algorithm, and recordings with a proportion of noisy or disturbed data segments of 10% or more are automatically rejected from the analysis. Accordingly, a 5-min analysis is automatically rejected if more than 30 s are noisy or disturbed. If fewer than 30 s are noisy or disturbed, those data segments are detected and removed automatically from analysis. For the 3-min segment analysis, the algorithm screens the 5-min files for a 3-min signal segment with <10% noisy signal. For the 1-min analysis, the algorithm does the same for a segment of 1-min duration (23).
Interpretation of iECG
All iECG recordings were analyzed by 2 blinded cardiologists and were categorized as “AF,” “SR,” or “nonusable” (e.g., insufficient quality, atrial flutter). In case of uncertainties, a third cardiologist was consulted, blinded for initial diagnoses. Cardiologists’ iECG diagnoses were used as reference for comparison with the PPG algorithm and the automated iECG interpretation, respectively.
Results of the 1-, 3- and 5-min PPG segment analysis, the cardiologists’ iECG diagnosis, and the automated iECG diagnosis were returned to the University Hospital Basel to be merged for statistical analyses.
Continuous values are expressed as mean and standard deviation; categorical values are expressed as absolute numbers and percentage. Differences between continuous variables were analyzed by the Wilcoxon rank sum test. Categorical variables were compared by the chi-square test or the Fisher exact test, as appropriate.
The central aim was to test the accuracy of the novel PPG-based algorithm to detect AF in comparison with the reference method of cardiologists’ diagnoses based on iECG. As a secondary analysis, the accuracy of the automated iECG interpretation algorithm to detect AF was compared with cardiologists’ interpretation of iECG.
The main analysis was based on all subjects who had a diagnosis provided by the PPG algorithm for the 1-min recordings. Additionally, the 3- and 5-min recordings were analyzed to test whether longer recording periods would result in improved diagnostic accuracy.
For comparison of each method (PPG, automated iECG interpretation, and cardiologists’ iECG diagnosis), different measures of diagnostic accuracy were calculated. First, we calculated sensitivity and specificity using contingency tables. Second, positive predicted values (PPVs), negative predicted values (NPVs), and accuracy were calculated. In addition, positive likelihood ratios (LR+ = sensitivity / [1 − specificity])] and negative likelihood ratios (LR− = [1 − sensitivity] / [specificity]) were calculated. The correct classification rate (CCR) as a percentage was also calculated with the formula: CCR = (number of correct classified AF cases + number of correct classified SR cases) / main sample size × 100, using an extended sample of all subjects with cardiologists’ rhythm diagnosis (n = 650).
Because of the case-control design of our study, the sample size in the positive (AF) and the negative (SR) groups do not necessarily reflect real disease prevalence. Consequently, calculations of PPV, NPV, and accuracy could result in incorrect results (25). Therefore, we recalculated these measures for a lower-risk scenario with an AF prevalence of 5%, which can be expected in an older general population setting (26), as well as for prevalence rates of 10% and 15%, corresponding to a CHA2DS2-VASc (congestive heart failure, hypertension, age ≥75 years, diabetes mellitus, prior stroke or TIA or thromboembolism, vascular disease, age 65 to 74 years, sex category [female]) score of 2 and 3, respectively (27), by using the 1-min sensitivity and specificity based on the Bayes theorem:
All statistical analyses were performed using Stata software version 14.1 (StataCorp, College Station, Texas) and SPSS software version 25.0 (IBM Corp., Armonk, New York).
A total of 672 patients were enrolled at the 2 centers between December 2016 and November 2017. Data from 650 patients with rhythm information on the basis of cardiologists’ diagnosis were eligible for analysis, after excluding 22 subjects for various reasons (Figure 1). For the main analysis, 142 subjects were excluded for insufficient 1-min PPG recording quality, thus resulting in a sample size of 508, among them 271 with SR and 237 with AF (Figure 1). The 2 groups were well balanced with respect to baseline characteristics, except for a higher proportion of subjects with a history of heart failure and pulmonary embolism, as well as a higher CHA2DS2-VASc score, a higher rate of vitamin K antagonist or direct oral anticoagulant agent use, and a lower rate of antiplatelet use in the AF group (Table 1).
For the 3-min and 5-min subanalyses, 203 and 281 subjects with poor PPG recording quality were excluded, resulting in sample sizes of 447 (204 with AF) and 369 (167 with AF), respectively (Online Figure 2).
Diagnostic accuracy of the PPG algorithm compared with cardiologists’ diagnosis
For the main analysis, we found a sensitivity of 93.7% (95% confidence interval [CI]: 89.8% to 96.4%) and a specificity of 98.2% (95% CI: 95.8% to 99.4%) for AF detection (Table 2). AF detection by PPG was related to excellent PPVs and NPVs of 97.8% (95% CI: 94.9% to 99.3%) and 94.7% (95% CI: 91.4 to 97.0%), respectively, as well as an excellent overall accuracy of 96.1% in this sample (Table 2). Similarly, excellent values for LR+ (52.1, 95% CI: 23.5 to 160.7) and LR− (0.06, 95% CI: 0.1 to 0.04) were found (Table 2). False-positive results were observed in 1.0% (n = 5 of 508) and false-negative results in 3.0% (n = 15 of 508). Numbers and percentages of correctly and falsely classified cases for the main analysis are also illustrated in Online Figure 3. Retrospective analyses revealed that the false-positive results were explainable by interference signals (n = 2), or the algorithm delivered borderline results (n = 3), whereas the respective recordings of the false-negative results have wrongly been classified as multiple extrasystoles (n = 6), or the algorithm delivered borderline results (n = 4) or clear SR results (n = 5).
Additional analyses were performed on the basis of an extended sample of all recordings for which a cardiologist’s diagnosis was available (n = 650), (Figure 1). Here, the automated PPG algorithm did not provide a diagnosis in 142 cases (21.8%) on the basis of 1-min analyses, thus resulting in a CCR of 74.8%.
Calculation of all diagnostic measures for the PPG algorithm by using 3-min and 5-min recordings did not lead to any relevant differences as compared with the 1-min analyses, with the exception of lower CCR values resulting from higher proportions of missing diagnoses (Table 2).
Diagnostic accuracy of the automated iECG interpretation algorithm
Although this was not the main focus of our study, we also calculated diagnostic measures for the automated iECG interpretation algorithm for detection of AF in comparison with cardiologists’ diagnosis. Among all 650 iECG recordings with cardiologists’ diagnosis, the algorithm did not provide a result in 101 cases (15.5%), thereby resulting in a sample size of 549 subjects (280 with AF) for this subanalysis.
Sensitivity and specificity to detect AF were 99.5% (95% CI: 97.5% to 99.9%) and 97.4% (95% CI: 94.7% to 98.9), respectively, with excellent PPV, NPV, and accuracy values (97.6%, 95% CI: 95.0% to 99.0%; 99.2%; 95% CI: 97.3% to 99.9%; and 98%, 95% CI: 96.9% to 99.2%). The LR+ and LR− were 38.2 (95% CI: 3.9 to 90.8) and 0.007 (95% CI: 0.03 to 0.001), respectively.
Among the extended sample with available cardiologist’s diagnosis (n = 650), the automated iECG algorithm did not provide a diagnosis for 101 (15.5%) of the subjects, which resulted in a CCR of 83.1%.
Prevalence-adjusted diagnostic measures for automated PPG and iECG analysis
Because PPV, NPV, and accuracy depend on the underlying disease prevalence, we recalculated these measures for both algorithms on the basis of a lower-risk scenario with an AF prevalence rate of 5%, which can be expected in an older general population setting (26), as well as for prevalence rates of 10% and 14%, corresponding to a CHA2DS2-VASc score of 2 and 3, respectively (27). Both methods were characterized by moderate to acceptable PPVs for the different AF prevalence scenarios (Table 3). NPVs and accuracy values remained very high for all tested prevalence scenarios (Table 3).
WATCH AF is the largest prospective trial comparing the diagnostic accuracy of a smartwatch-derived PPG signal algorithm to detect AF with an ECG diagnosis. By using a commercially available smartwatch, detection of AF was feasible, with very high diagnostic accuracy on the basis of cohesive 1-min PPG recordings in subjects with sufficient signal quality (Figure 2). The diagnostic accuracy did not improve much when longer recordings of 3 or 5 min were analyzed, although the percentage of files with a suitable signal quality decreased. However, a further important finding of this study is that the applicability of such algorithms in practice is currently limited by a high dropout rate resulting from insufficient signal quality.
As a proof of principle, we could show that smartwatch-derived PPG analyses enable AF detection with an accuracy (sensitivity, 93.7%; specificity, 98.2%; overall accuracy, 96.1%) that is at least on the same level as that of other methods. Smartphone-based PPG algorithms showed a sensitivity between 90% and 96% and a specificity between 85% and 99% (16–19,22,23). The novel PPG algorithm outperforms other methods that have been proposed for AF screening, such as pulse palpation (specificity, 72%), modified sphygmomanometers (specificity, between 90% and 97%), or non–12-lead ECG (specificity between 77% and 97%) (13,28,29).
Only 1 comparable trial, a substudy of the eHeart study, evaluated the applicability of a smartwatch for detection of AF on the basis of PPG recordings (20). This study included a large number of subjects, but a completely different study design and algorithm (a deep neural network) were used. The study comprised 3 stages: (I) algorithm development using a remote cohort of 9,750 subjects; (II) external validation in an in-person cardioversion cohort of 51 patients; and (III) an exploratory analysis in 1,617 ambulatory patients (4% with AF). The tested algorithm exhibited a C-statistic of 0.97 (95% CI: 0.94 to 1.00) to detect AF against 12-lead ECG diagnosis in the validation cohort (sensitivity, 98.0%; specificity, 90.2%). In stage III, however, the accuracy to detect AF was only moderate (sensitivity, 67.7%; specificity, 67.6%). Other major limitations of this study were the high number of dropouts and that AF was self-reported in stages I and III. The diagnostic measures of our WATCH AF trial are comparable with that reported for stage II of this study. However, AF diagnosis was confirmed by ECG in all cases in WATCH AF, but only in 51 subjects of the eHeart substudy.
A smaller pilot study used a research watch-based wearable device (Simband, Samsung Electronics Co.) combining a single-channel ECG, a PPG sensor, and a triaxial accelerometer (19). Although very high accuracy to detect AF was reported, the value of this study was limited by the low sample size of 36 for algorithm development (12 with AF) and an even smaller sample size of 10 subjects (3 with AF) for validation.
A recent study by Bumgarner et al. also used smartwatches for AF detection but with a different technology and approach (21). The Kardia Band (AliveCor Inc.), an accessory for the Apple Watch (Apple Inc.) that enables recording of 1-lead ECG for 30 s, was used, coupled with an application for automated rhythm analysis. In total, 100 consecutive patients with AF presenting for cardioversion were enrolled. Of 169 Kardia Band recordings, 57 (33%) were not interpretable, and the results of the remaining 112 recordings were compared with physicians’ diagnosis using 12-lead ECG. The tested algorithm demonstrated a 93% sensitivity, an 84% specificity, and a K coefficient of 0.77 for AF detection. With respect to the low sensitivity, this tool is currently not suitable for screening purposes. In addition, 2 major limitations must be considered: the high proportion of nonusable recordings; and the need for active involvement of the patient, a feature that does not allow for “continuous” rhythm screening.
The very high specificity of the PPG-based automated AF detection algorithm tested in WATCH AF could make it an attractive method for large-scale screening, given the low number of false-positive results. High false-positive rates would limit its use as a mass screening tool because of unnecessary follow-up examinations. However, further technical improvements are necessary. Thus, among the 650 subjects who also had a cardiologist’s rhythm diagnosis, the algorithm did not provide a rhythm diagnosis in 21.8% because of insufficient signal quality, thereby showing that achieving a good signal quality from the wrist is still challenging. Notably, the rate of missing diagnosis was also high for automated iECG interpretation (15.5%). Thus, both methods tested depend on sufficient signal quality to achieve high enough sensitivity and specificity.
Once technical issues such as signal quality or limited battery duration have been resolved, another important task will be to identify the setting in which smartwatch-based AF screening could be used. Smartwatch-based tools may not be understood as a competing method in settings where other devices are the current gold standard. Such novel screening tools, however, could potentially fill the gap between external and implanted device use in patients when these methods are not applicable for medical reasons or because of lacking reimbursement. Our data do not argue for the application of smartwatch-based tools in the unselected general population, given the moderate PPV when they were used in samples with lower AF prevalence rates. Such a tool could have its strengths in patients with a high pre-test probability for AF, for example, in individuals with a CHA2DS2-VASc score of 2 or more where an AF prevalence of at least 10% can be expected (27). However, we did not test the algorithm in such a setting but only recalculated the diagnostic measures on the basis of a mathematical theorem. Therefore, these additional data have to be interpreted very carefully. Ultimately, prospective trials in population-based at-risk samples have to assess the real diagnostic value in this setting. Notably, the limitation of lower predictive values in the presence of low disease prevalence is a well-known statistical issue for diagnosis in general (25,30).
Limitations that are known for all PPG-based devices include potential influences of extrasystoles, reduced PPG pulse wave tissue penetration in subjects with intense skin pigmentation (31), or worsening of the signal-to-noise ratio in patients with tremor. A significant proportion of recordings had to be excluded because of insufficient signal quality, at least partly explainable by the trial version of the algorithm used that did not include automated signal quality check. It cannot be excluded that the diagnostic performance of the tested tool may even be more limited in an ambulatory setting as compared with our study conditions where movement artifacts could be more frequent. In this trial, we only used 1 smartwatch model (Gear Fit 2, Samsung Electronics Co.). A potential advantage of the algorithm is that it is not limited to a specific operating system. The PPG-based algorithm was compared against iECG and not against the gold standard 12-lead ECG or Holter ECG. This design was chosen because iECG has been recommended in the European Heart Rhythm Association guidelines for AF screening. Therefore, it appeared to be relevant to compare these 2 novel technologies directly. According to current guidelines, final confirmation of AF still remains an ECG-based diagnosis (13). Thus, PPG-based AF detection devices may still be understood as AF screening tools with a need for a confirmatory ECG for suspected AF.
Moreover, the study group was not limited to patients older than 65 years of age because the tested screening tools are also used by younger individuals in real life, thus making validation data for this group necessary. Our study included individuals with or without a history of AF by using a case-control design and therefore did not reflect real-life conditions. The ability to identify a new diagnosis of the disease must be tested in further studies.
In the near future, smart devices could facilitate large-scale screening for AF and have a substantial clinical impact by enabling AF detection earlier in subjects with a high risk for AF, in those with recurrent AF after ablation or pharmacological cardioversion, or in those with cryptogenic stroke. Specifically, these tools could be used when other devices such as Holter ECG monitors or implanted loop recorders may not be applicable or may not be reimbursed. An ideal tool for this purpose could be a more advanced device combining a PPG-based algorithm that enables continuous rhythm screening with technology that alerts the patient in case of suspected AF and enables immediate registration of diagnostic ECG.
Detection of AF with a commercially available smartwatch is in principle feasible, with a very high diagnostic accuracy. Its applicability is still limited by a high proportion of uninterpretable recordings. Achieving a sufficient signal quality from the wrist is still a challenge that has to be resolved by technological improvements and more advanced algorithms. In light of the increasing use of smartwatches, their consistent availability, and the possibility of a user-independent continuous application, this technology could raise smart device–based rhythm screening to the next level compared with previous approaches with smartphones and become a widely applicable, convenient, and potentially cost-effective tool for large-scale AF screening or extended rhythm monitoring in high-risk individuals. Further investigation involving broad population screening is the next step.
COMPETENCY IN MEDICAL KNOWLEDGE: Detection of AF on the basis of an automated algorithm using a commercially available smartwatch alone is in principle feasible, with very high diagnostic accuracy, given sufficient signal quality.
TRANSLATIONAL OUTLOOK: In the future, use of smartwatches may have substantial clinical impact by enabling convenient and continuous large-scale rhythm monitoring in subjects at risk for AF who may not use other devices such as Holter ECG monitors or implantable loop recorders. Randomized population screening trials are needed to validate the clinical utility of this tool for detection of arrhythmias.
The authors would like to thank Ada Winterhalder and Ulrike Schmitt for their support.
The trial was supported by unrestricted research grants from Preventicus GmbH and the University Hospital Basel. The funders had no influence on the design or conduct of the trial and were not involved in data collection or statistical analysis, in the writing of the manuscript, or in the decision to submit it for publication. Dr. Dörr holds 0.5% virtual shares in Preventicus. Dr. Eckstein holds 0.5% virtual shares in Preventicus; and has received a travel grant from Preventicus. All other authors have reported that they have no relationships relevant to the contents of this paper to disclose.
All authors attest they are in compliance with human studies committees and animal welfare regulations of the authors’ institutions and Food and Drug Administration guidelines, including patient consent where appropriate. For more information, visit the JACC: Clinical Electrophysiology author instructions page.
- Abbreviations and Acronyms
- atrial fibrillation
- correct classification rate
- congestive heart failure, hypertension, age ≥75 years, diabetes mellitus, prior stroke or TIA or thromboembolism, vascular disease, age 65 to 74 years, sex category (female)
- confidence interval
- Internet-enabled mobile electrocardiography
- negative likelihood ratios
- positive likelihood ratios
- negative predicted values
- positive predicted values
- sinus rhythm
- Received August 27, 2018.
- Revision received October 1, 2018.
- Accepted October 11, 2018.
- 2019 American College of Cardiology Foundation
- Chugh S.S.,
- Havmoeller R.,
- Narayanan K.,
- et al.
- Wolf P.A.,
- Abbott R.D.,
- Kannel W.B.
- Kernan W.N.,
- Ovbiagele B.,
- Black H.R.,
- et al.
- Grond M.,
- Jauss M.,
- Hamann G.,
- et al.
- Kishore A.,
- Vail A.,
- Majid A.,
- et al.
- Albers G.W.,
- Bernstein R.A.,
- Brachmann J.,
- et al.
- Mairesse G.H.,
- Moran P.,
- Van Gelder I.C.,
- et al.
- Lau C.P.,
- Siu C.W.,
- Yiu K.H.,
- Lee K.L.,
- Chan Y.H.,
- Tse H.F.
- Freedman B.,
- Camm J.,
- Calkins H.,
- et al.
- Koenig N.,
- Seeck A.,
- Eckstein J.,
- et al.
- Krivoshei L.,
- Weber S.,
- Burkard T.,
- et al.
- Chan P.H.,
- Wong C.K.,
- Poh Y.C.,
- et al.
- Nemati S.,
- Ghassemi M.M.,
- Ambai V.,
- et al.
- Tison G.H.,
- Sanchez J.M.,
- Ballinger B.,
- et al.
- Bumgarner J.M.,
- Lambert C.T.,
- Hussein A.A.,
- et al.
- Brasier N.,
- Raichle C.J.,
- Dorr M.,
- et al.
- Altman DGm Bland J.M.
- Berge T.,
- Lyngbakken M.N.,
- Ihle-Hansen H.,
- et al.
- Tischer T.,
- Schneider R.,
- Lauschke J.,
- et al.
- Hobbs F.D.,
- Fitzmaurice D.A.,
- Mant J.,
- et al.