Diagnostic Utility of Smartwatch Technology for Atrial Fibrillation Detection – A Systematic Analysis

Mehmet Ali Elbey1, Daisy Young2, Sri Harsha Kanuri1, Krishna Akella1, Ghulam Murtaza1, Jalaj Garg3, Donita Atkins4, Sudha Bommana4, Sharan Sharma4, Mohit Turagam5, Jayashree Pillarisetti6, Peter Park4, Rangarao Tummala4, Alap Shah4, Scott Koerber4, Poojita Shivamurthy4, Chandrasekhar Vasamreddy4, Rakesh Gopinathannair4, Dhanunjaya Lakkireddy4

1Arrhythmia Research Fellow, Kansas City Heart Rhythm Institute, Overland Park, Kansas.2Department of Internal Medicine, Stony Brook Southampton Hospital, Southampton, NY.3Division of Cardiology, Cardiac Arrhythmia Service, Medical College of Wisconsin, Milwaukee, Wisconsin.4Division of Cardiac Electrophysiology, Kansas City Heart Rhythm Institute & Research Foundation; Overland Park Regional Medical Center, HCA Midwest Overland Park, Kansas.5Helmsley Electrophysiology Center, Icahn School of Medicine at Mount Sinai, New York, NY.6University of Texas Health Science Center at San Antonio, San Antonio, TX.

Abstract

Background

Smartphone technologies have been recently developed to assess heart rate and rhythm, but their role in accurately detecting atrial fibrillation (AF) remains unknown.

Objective

We sought to perform a meta-analysis using prospective studies comparing Smartwatch technology with current monitoring standards for AF detection (ECG, Holter, Patch Monitor, ILR).

Methods

We performed a comprehensive literature search for prospective studies comparing Smartwatch technology simultaneously with current monitoring standards (ECG, Holter, and Patch monitor) for AF detection since inception to November 25th, 2019. The outcome studied was the accuracy of AF detection. Accuracy was determined with concomitant usage of ECG monitoring, Holter monitoring, loop recorder, or patch monitoring.

Results

A total of 9 observational studies were included comparing smartwatch technology, 3 using single-lead ECG monitoring, and six studies using photoplethysmography with routine AF monitoring strategies. A total of 1559 patients were enrolled (mean age 63.5 years, 39.5% had an AF history). The mean monitoring time was 75.6 days. Smartwatch was non-inferior to composite ECG monitoring strategies (OR 1.06, 95% CI 0.93 – 1.21, p=0.37), composite 12 lead ECG/Holter monitoring (OR 0.90, 95% CI 0.62 – 1.30, p=0.57) and patch monitoring (OR 1.28, 95% CI 0.84 - 1.94, p=0.24) for AF detection. The sensitivity and specificity for AF detection using a smartwatch was 95% and 94%, respectively.

Conclusions

Smartwatch based single-lead ECG and photoplethysmography appear to be reasonable alternatives for AF monitoring.

Key Words : Smartwatch, Atrial Fibrillation, Photoplethysmography.

Dhanunjaya (DJ) Lakkireddy MD, FACC, FHRS Executive Medical Director The Kansas City Heart Rhythm Institute (KCHRI) HCA MidWest 12200, W 106th street, Overland Park Regional Medical Center Overland Park, KS 66215 HCA MidWest, Overland Park, Kansas 66221

Introduction

Atrial fibrillation (AF) is the most common cardiac arrhythmias affecting over 33.5 million people worldwide, increasing morbidity and mortality 1-3. AF is frequently subclinical or paroxysmal, which causes a significant barrier to its expedient diagnosis and treatment. Electrocardiogram (ECG) is often normal between the episodes and reflects only single time-point measurements. Other implantable or recording devices are limited by user activation, external factors and are either expensive or invasive. Photoplethysmography was recently developed and implemented in wearable Smartwatches in 2017, providing a cost-effective and non-invasive means for continuous ECG monitoring. Over the past two years, several large-scale prospective trials have compared Smartwatch technology with current monitoring standards for AF detection, such as ECG, Holter Monitor (HM), Implantable Loop Recorder (ILR), and Patch Monitoring (PM). Since initial development, several studies have become available comparing smartwatch technology with the current standards 4-12. Hence, we performed a meta-analysis comparing the accuracy of Smartwatch-based single-lead ECG and PPG to current monitoring standards for AF detection.

Methods

Search Strategy

The initial search strategy was developed by two authors (KA and GM). Systematic search, without language restriction, using PubMed, EMBASE, SCOPUS, Google Scholar, and ClinicalTrials.gov from inception to November 25th, 2019 using the keywords: “smartwatch” OR “watch” AND “atrial fibrillation” or “watch” AND “atrial fibrillation detection.”

Study Selection

The eligibility criteria our systematic review and meta-analysis included:

1. All prospective studies reporting clinical outcomes comparing Smartwatch technology simultaneously with current monitoring standards (ECG, Holter, and Patch monitor).

2. Human subjects aged ≥ 18 years

3. Studies in the English language.

Case reports, editorial, and systematic reviews were excluded.

Data Extractions

Two investigators independently performed the literature search and screened all titles and full-text versions of all relevant studies that met the study inclusion criteria.

The references of all identified articles were also reviewed for relevant studies meeting the eligibility criteria. The data from the included studies were extracted using a standardized protocol and a data extraction form. Any discrepancies between the two investigators were resolved with a consultation with the senior investigator (DL). The following data were extracted: title, year of publication, type of study, mean age, sample size, baseline technology used, specific watch used, the specific algorithm used for AF detection, comparator modality, the quantity of ECG leads, and time monitored [Table 1]. Quantitative data on AF detection, including a discrete number of AF events, sensitivity, and specificity, were obtained [Table 2]). The Newcastle Ottawa Scale [Table 3] was used to appraise the quality of the included studies 13. We rated the quality of the studies (good, fair, and poor) by awarding stars in each domain. A “good” quality score required 3 or 4 stars in the selection, 1 or 2 stars in comparability, and 2 or 3 stars in outcomes. A “fair” quality score required 2 stars in the selection, 1 or 2 stars in comparability, and 2 or 3 stars in outcomes. A “poor” quality score reflected 0 or 1 star(s) in selection, or 0 stars in comparability, or 0 or 1 star(s) in outcomes.

Table 1. Baseline characteristics of the studies included in our meta-analysis
Study Perez et al Bashar et al Wasserlauf et al Dorr et al Tison et al Bumgarner et al Faranesh et al Rajakariar et al Genicot et al
Year 2019 2019 2019 2019 2018 2018 2019 2019 2018
Type Prospective Prospective Prospective Prospective Prospective Prospective Prospective Prospective Prospective
Mean Age 59 N/A 72.1 76.4 42 68 N/A 67 60
# enrolled 450 37 24 508*, 549♱ 51 93 96 200 100
Baseline tech used Plethysmo-graph Plethysmo-graph Kardiaband Plethysmo-graph Plethysmo-graph SW single lead ECG (Kardia band) Plethysmo-graph SW single lead ECG (iPhone ECG) Plethysmo-graph
Specific watch used Apple Watch Samsung Simband Apple Watch Samsung Gear Fit2 Apple Watch Apple Watch Fitbit SW N/A N/A
Specific algorithm used PPI on plethysmo-graphy PPI on plethysmo-graphy Smart- Rhythm 2.0 (Kardia band) PPI on plethysmo-graphy PPI on plethysmo-graphy Kardiaband N/A Kardiaband PPI on plethysmo-graphy
Comparator 7-day ECG patch - single lead Holter Monitor Reveal LINQ iPhone ECG Ambulatory ECG post-cardioversion Ambulatory ECG post-cardioversion Single lead ECG patch (Cardea SOLO) 12-lead ECG 24-hour Holter Monitor
# of leads in comparator Single Multiple Single Single 12-lead 12-lead Single 12-lead 12-lead
Time monitored 117 days N/A 31,349 hours 1 min each 30 min 30 second strips 7 days N/A 99 days

* for Plethysmograph

Outcomes

The primary outcome evaluated in our study was the accuracy of AF detection. Accuracy was determined with concomitant usage of ECG monitoring, Holter monitoring, loop recorder, or patch monitoring.

Table 2. Quantitative Evaluation of Atrial Fibrillation Events Detected
Study Perez et al Bashar et al Wasserlauf et al Dorr et al Tison et al Bumgarner et al Faranesh et al Rajakariar et al Genicot et al
# enrolled 450 37 24 508*,549♱ 51 93 96 200 100
# with AF 153 10 24 237 51 93 35 38 N/A
% with AF 34 27 100 46.65 100 100 36.46 19 N/A
Sensitivity of SW AF detection N/A 98.18% 97.70% 93.70% 98% 93% N/A 89.5% N/A
Specificity of SW AF detection N/A 98.07% 98.90% 98.20% 90.20% 84 N/A 94.40% N/A

* for Plethysmograph

Table 3. Qualitative Evaluation of Included Studies Using Newcastle-Ottawa Scale
Study Selection(max 4 stars) Comparability(max 2 stars) Outcome(max 3 stars)
Perez et al **** ** ***
Bashar et al *** ** **
Wasserlauf et al **** ** ***
Dorr et al **** ** ***
Tison et al **** ** **
Bumgarner et al **** ** ***
Faranesh et al **** ** ***
Rajakariar et al **** ** **
Genicot et al **** ** ***



Statistical Analysis

Statistical analysis for odds ratio (OR) estimates of each study was calculated using Stata (Version 16.1, StataCorp, College Station, TX 77845). Subsequent meta-analysis was performed using Comprehensive Meta-Analysis (CMA) software (version. 3.3.070, Biostat, Englewood, NJ 07631) with a random-effects model. Results were expressed as an OR with a 95% confidence interval (CI). Higgins I-squared (I2) was used to quantify heterogeneity (I2<50% was defined as low) 14. P < 0.05 was considered statistically significant. Sensitivity analyses were performed for outcomes that demonstrated significant heterogeneity (I2 >50%) to assess the individual contribution to the aggregate. Funnel plots were also used in conjunction with sensitivity analysis to assess for publication bias.

Results

Search Results and study characteristics

A total of 1,045 citations were identified [Figure 1] during the initial search. Nine hundred four records were excluded, and a total of 51 relevant articles were identified. After a detailed evaluation, nine articles ultimately met the inclusion criteria enrolling a total of 1,559 patients [Figure 1]. A total of 1559 patients were enrolled (mean age 63.5 years, 39.5% had an AF history). The mean monitoring time was 75.6 days. [Table 1] summarizes the study characteristics of the included trials.

Figure 1. PRISMA flow diagram



Primary Outcome [Figure 2]

AF Detection

Smartwatch was non-inferior to composite ECG monitoring strategies (OR 1.06, 95% CI 0.93 – 1.21, p = 0.37) [Figure 3A], composite 12 lead ECG/Holter monitoring (OR 0.90, 95% CI 0.62 – 1.30, p = 0.57) [Figure 4A] and patch monitoring (OR 1.28, 95% CI 0.84 - 1.94, p = 0.24) [Figure 5A] for AF detection. Both mean sensitivity and specificity for AF detection using smartwatch was 95% and 94%, respectively. Several studies demonstrated statistically significant differences in AF sensing capability. Perez et al., Faranesh et al., and Tison et al. demonstrated statistically significant oversensing 4, 8, 10, while Wasserlauf et al. and Genicot et al. demonstrated smartwatch comparative undersensing 6, 12.

Figure 2. Diagnostic Utility of Smartwatch Technology for Atrial Fibrillation Detection: The Smartwatch-AF Study



Sensitivity Analysis

Due to significant heterogeneity observed in the primary outcome, a sensitivity analysis was performed by excluding one study at a time to see if any study had a substantial contribution to observed heterogeneity. The heterogeneity found may be associated with individual study differences, institutional variation, and the difference in evaluation among the included studies.

Sensitivity analysis performed on a composite analysis [Figure 3B] demonstrated no significant heterogeneity changes (I2 = 99.2%). In comparison, the reduction in heterogeneity (from I2 = 75.3% to 0%) was observed with the exclusion of Genicot et al., which had a significant proportion of weight in the study for watch monitoring vs. composite 12 lead and Holter monitoring comparison. This is consistent with the single study outlier noted in the corresponding funnel plot [Figure 4B]. Sensitivity analysis was not performed on watch monitoring vs. patch monitoring comparison as only two studies were included in the subgroup analysis (I2 = 99.1%).

Figure 3. A. Smartwatch Monitoring vs. Composite ECG, Holter, Loop Recorder and Patch Monitor; B. Funnel plot demonstrating asymmetry suggestive of publication bias



Figure 4. A. Smartwatch Monitoring vs. Composite 12-Lead ECG and Holter Monitoring; B. Funnel plot demonstrating asymmetry suggestive of publication bias



Publication Bias

In addition to sensitivity analysis, publication bias was assessed visually using funnel plots [Figure 3], demonstrating asymmetrical funnel plot figures. Consequently, despite the overall findings of no difference between watch monitoring and routinely practiced wearable/implantable devices, results should be interpreted with caution (given funnel plot asymmetry and high heterogeneity).

Figure 5. Smartwatch Monitor vs. Patch Monitor



Discussion

Our analysis demonstrated no significant difference in AF detection in Smartwatch monitoring compared to composite ECG monitoring, Holter monitoring, loop recorder, and patch monitoring [Figure 2]. There has been a progressive increase in the incidence and prevalence of AF worldwide, with an increased risk of morbidity and mortality. Atrial fibrillation is known to have a significant impact on health care costs, with the major cost-drivers being the loss of productivity, stroke, and hospitalizations 1. Studies have also shown that increased AF burden directly correlates with thromboembolic stroke risk 15-19. Accordingly, increased awareness of AF symptoms and early clinical diagnosis is imperative to prevent long term morbidity and mortality. Often, given the asymptomatic nature of paroxysmal AF, long term monitoring for detecting these clinically relevant silent AF episodes is essential 4, 20. Although cardiac implantable electronic devices are commonly used for monitoring of silent AF episodes, they are associated with potential disadvantages such as invasiveness of the procedure, procedure-related complications, and long-term patient discomfort 4. There is a growing need to develop non-invasive and wearable technology to enable continuous monitoring of silent arrhythmias in high-risk patients 21.

Photoplethysmography-based technology included in the smartwatches (Apple or Samsung) is regarded as the most accurate method for diagnosing AF 21-24. Photoplethysmography based smart devices; mobile health (Mobile Health) in combination with machine learning, has transformed patient care by precisely and accurately diagnosing AF 25-27. Photoplethysmography in smartwatches consists of an infrared light-emitting diode sensor that detects blood volume changes in the microvasculature 21, 28. The synchronous changes in blood volume in small blood vessels with each heartbeat are transformed into a physiological pulsatile waveform by Photoplethysmography 21, 28. It is regarded as a portable, low-cost, simple, and wearable technology most suitable for monitoring patients in primary care and community-based clinical settings 21. Furthermore, this technology has been formerly used for measuring oxygen saturation, blood pressure, cardiac output, autonomic changes, and peripheral vascular disease 4; with better reliability than previously used technologies such as pulse palpation, modified sphygmomanometers, and non–12-lead ECG for detection of AF episodes 29-31. Smartwatch-based arrhythmia detection (with a photoplethysmography-based AF detection) is a simple, non-invasive technique and a safer alternative to the routinely utilized AF detection tools.

Previous studies

Several studies have been conducted to assess the efficacy of photoplethysmography based smartwatch technologies for detecting AF episodes. The overall sensitivity and specificity for detecting AF episodes using smartphone technology is approximately 90-96% and 85-99%, respectively 32-37. ECG watchband (KardiaBand, AlivaCor, USA), which is connected to the Apple Watch, was first introduced in April 2017 for detecting AF [38]. Kardiaband was based on a proprietary algorithm (rhythm irregularity and absence of P waves) for AF detection 38 and transmitting a 30s segment of single-lead ECG via Bluetooth to the Apple Watch 38. Similarly, Bumgarner et al. compared the efficacy of the Apple Watch with a standard clinical 12 lead ECG in 100 AF patients and demonstrated that the sensitivity and specificity of the Apple Watch for detecting AF are 93% and 84%, respectively 39. Several other studies have used AlivaCor Kardia Mobile approaches and demonstrated that sensitivity and specificity were >95% 40-42. With the help of motion and noise artifacts and premature atrial contraction algorithms, photoplethysmography based smartwatch detected AF with higher sensitivity (98.18%), specificity (97.43%), and accuracy (97.54%) 5. According to Wasserlauf et al., AF-based smartwatches had higher sensitivity for detecting the AF episodes (episode sensitivity 97.5%) and AF duration (duration sensitivity 97.7%) as compared to implantable cardiac monitor (ICM) 6.

The WATCH-AF trial demonstrated relatively high sensitivity (93.7%), specificity (98.2%), overall accuracy (96.1%), positive predictive value (97.8%), and negative predictive value (94.7%) in diagnosing AF 7, findings that also echoes (comparing simultaneously performed 12-lead ECG) in a recently published study by Rajakariar et al. 43. Using single-channel electrocardiogram (ECG), multi-wavelength photoplethysmography, tri-axial accelerometry, the accuracy, sensitivity, and specificity of detecting AF episodes with Samsung Simband watch was 95%, 97%, and 94%, respectively [34]. Photoplethysmography based smartwatch technology combined with deep neural network passively predicted AF in patients undergoing cardioversion better than standard 12-lead ECG with higher sensitivity (94%) and specificity (90.2%) in an ambulatory care setting 8. Furthermore, although both Apple Watch Series 3 and Fitbit were equipped with Photoplethysmography technology, the precision and accuracy for AF detection was higher in Apple Watch Series 3 (75% correlation) as compared to Fitbit (FBT) Charge HR Wireless Activity Wristband (30% correlation) in a phase-II prospective clinical study conducted in Japan 28. The false-positive rate and accuracy of AF detection in healthy volunteers and AF patients using smartwatches with Photoplethysmography based algorithm was approximately 0.2% and 96%, respectively 44, thus demonstrating the efficacy of photoplethysmography based wearable devices accurately differentiating AF from sinus rhythm in at-risk patients 44.

Limitations

There are several limitations to the performed meta-analysis. The limitations did not include a comprehensive text and comparison to literature.

1. Patients with implantable cardiac pacemakers were excluded from the studies.

2. Some studies included patients with prior history of paroxysmal AF, while others excluded these patients, limiting the generalizability of the studies.

3. Study heterogeneity.

4. Variation in algorithms used for different devices.

5. Differences in metrics of assessment among different studies.

Conclusions

While composite 12-lead ECG, Holter monitor, implantable loop recorders, or patch recording are the standards for AF detection, photoplethysmography based smartwatch technology is a simple, efficient, and non-inferior alternative that may expedite detection and treatment of subclinical AF, preventing morbidity and mortality from stroke and cardiovascular disease.

References

  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]
  • [ PubMed ]