Diagnostic validation of smart wearable device embedded with single-lead electrocardiogram for arrhythmia detection

Abstract

Objective

To validate a single-lead electrocardiogram algorithm for identifying atrial fibrillation, atrial premature beats, ventricular premature beats, and sinus rhythm.

Methods

A total of 656 subjects aged 19 to 94 years were enrolled. Participants were simultaneously tested with a wristwatch (Huawei Watch GT2 Pro, Huawei Technologies Co., Ltd, Shenzhen, China) and a 12-lead electrocardiogram for 3 minutes. A total of 1926 electrocardiogram signals from 628 subjects (282 men and 346 women) aged 19 to 94 years (median 64 years) were analyzed using an algorithm.

Results

The numbers of subjects with atrial fibrillation, atrial premature beats, ventricular premature beats, and sinus rhythm were 129, 141, 107, and 251, respectively, and together they had a total of 1926 electrocardiogram signals. For the three-class classification system, the recall, precision, and F1 score were 97.6%, 96.5%, 97.0% for sinus rhythm; 96.7%, 96.9%, 96.8% for atrial fibrillation; and 92.8%, 94.2%, 93.5% for ectopic beats, respectively. The macro-F1 score of the three-class classification system was 95.8%. For the four-class classification system, the recall, precision, and F1 score were 97.6%, 96.5%, 97.0% for sinus rhythm; 96.7%, 96.9%, 96.8% for atrial fibrillation; 90.5%, 89.4%, 89.9% for atrial premature beats; and 86.1%, 89.6%, 87.8% for ventricular premature beats, respectively. The macro-F1 score of the four-class classification system was 92.9%.

Conclusions

The single-lead electrocardiogram algorithm embedded into smart wearables demonstrated good performance in detecting atrial fibrillation, atrial/ventricular premature beats, and sinus rhythm, and thus would facilitate atrial fibrillation screening and management.

Introduction

Atrial fibrillation (AF) is one of the major causes of stroke, heart failure, and peripheral vascular embolism worldwide, which greatly increases medical costs. ¹ Once AF is diagnosed, patients may need to undergo anticoagulant therapy to prevent thromboembolic events and improve prognosis based on CHA₂DS₂-VASc and HAS-BLED. Nonetheless, AF is often asymptomatic, resulting in possible delays in diagnosis. Some patients are even diagnosed with AF only when they have concurrent stroke.^2,3 Thus, asymptomatic AF has become a major public health problem, and improving the AF detection rate and ensuring timely individualized intervention are of great significance for improving the prognosis of patients, reducing the incidence of complications, and reducing medical costs.

The early diagnosis of AF, especially silent or subclinical asymptomatic AF, has currently gained widespread interest. Since 2012, guidelines for the diagnosis and treatment of AF recommended opportunistic screening for AF in people aged over 65 years, with a combination of pulse palpation and electrocardiogram (ECG). ⁴ The 2016 ESC/EACTS Guidelines for the Management of AF suggested pulse palpation and a 12-lead ECG be applied for the opportunistic screening of AF. However, pulse palpation is not specific for ventricular arrhythmias, and some AF episodes will be missed with a 12-lead ECG given the paroxysmal nature of AF. ⁵

Subsequently, various mobile technologies (mHealth), including single/multilead electrocardiograms (ECGs), ⁶ photoplethysmography (PPG),^7,8 oscillometry, and video recordings ⁹ were developed to detect AF. Smartwatch-based ECG monitors allow the individual to record a single-lead ECG by placing a finger on the stem of the watch for 30s. The noninvasive feature of smartwatch-based ECG monitor makes it comfortable to wear for long time periods in daily life, which enables the device to continuous surveillance of intermittent AF episodes. ¹⁰ Currently, smartwatch-based single-lead ECG usually interprets the ECG signals as either “sinus rhythm,” “possible AF,” or “unreadable.” ¹¹ The sensitivities and specificities of algorithms embedded on the single-lead ECG devices detecting AF were around 95%. ¹² However, most algorithms used by single-lead ECG device generalized with a reduction in F1 score to a maximum F1 of 83.1%, when heart rhythm was classified as sinus rhythm, AF, other rhythm, and noise, ¹³ even a combination of state-of-the-art algorithms increased the F1 score to just 90%. ¹⁴ Because a single-lead ECG can only record the ECG along a single vector, these devices may misdiagnosis cardiac electrical activity due to a low amplitude, such as P-waves, especially when discriminating cardiac arrhythmias other than AF. Also, atrial/ventricular premature beats may cause irregular rhythm, so there is a disadvantage of the dependence on regularity of rhythm to define AF. ¹⁵ Thus, the accurate identification of premature ectopic, AF, and sinus rhythm is crucial because accurate diagnosis of AF would warrant shared decision-making for stroke prevention based on underlying cardiovascular risk factors of patients.

But there are few studies that have assessed the accuracy of these mobile technologies in distinguishing atrial and ventricular premature beats from AF. In this study, we validated a single-lead ECG algorithm to detect AF, atrial premature beats, ventricular premature beats, and sinus rhythm, through a comparison to a standard 12-lead ECG to provide additional useful information.

Methods

Subjects

Between June 2020 and April 2021, 656 subjects aged 19 to 94 years (median 65 years) were enrolled from Chinese PLA General Hospital and First Affiliated Hospital of Tsinghua University (Beijing Huaxin Hospital).

Subjects aged over 18 years who were willing to sign a written informed consent form were included in the study. Exclusion criteria included patients unable to use wearables, those with mental or memory problems, or those with a pacemaker or implantable cardioverter defibrillator.

All subjects provided their informed written consent to participate in this study. This study was conducted according to the World Medical Association Declaration of Helsinki and was reviewed and approved by the institutional review board of Chinese PLA General Hospital (S2017-105-02) and the clinical research ethics committee of Beijing Huaxin Hospital (Clinical 2021-03).

Signal acquisition, processing, and diagnosis

Subjects were advised to lie down in a supine position and breathe spontaneously while being simultaneously measured using a wristwatch (Huawei Watch GT2 Pro) and 12-lead ECG (Figure 1). Single-lead ECG recordings were taken from the subjects under the supervision of trained study personnel. After the trained study personnel correctly connected the 12-lead ECG electrodes to the subjects, the subjects were fitted with wristwatches (Huawei Watch GT2 Pro). All subjects wore wristwatches on their left wrists and used their right hands to touch the wristwatches. The trained personnel simultaneously pressed the start button on the 12-lead ECG machine, and the 12-lead ECG machine and wristwatch then collected ECG signals for 3 minutes.

Figure 1.

A prototype for arrhythmia detection using wristwatch (Huawei Watch GT2 pro) and 12-lead ECG.

Because the wristwatch was fitted on the subject's left wrist, the waveform of the single-lead ECG was considered equivalent to that of the limb Lead I of the 12-lead ECG. Furthermore, we compared the waveform from Lead I of the 12-lead ECG with that of the single-lead ECG. The signal quality of the single-lead ECG was automatically evaluated by the algorithm module to determine whether the signals needed to be further analyzed for cardiac rhythm.

The algorithm of single-lead ECG algorithm used was developed by Huawei. The algorithm's ECG signal filtering, feature extraction, signal quality control, feature calculation, and the output are presented in Figure 2. First, the ECG signal was filtered to remove baseline drift and high-frequency signal interference, after which feature points were extracted to identify QRS waves and extract more feature points. Accelerometer (ACC) signals were used to determine the quality of the ECG signals, and poor-quality ECG signals were removed. Other features, such as the intervals between R waves and R waves (RRI), were calculated according to the selected feature points, and then feature classification was carried out using either the three- or four-class classification model (Figure 3). A rhythm-based approach was adopted to differentiate the arrhythmias, with the algorithm running in the watch throughout the comparison between current and preceding intervals.

Figure 2.

A flowchart of the study.

Figure 3.

The work flowing of single-lead ECG algorithm. AF, atrial fibrillation.

The wristwatch with the pre-installed software captured single-lead ECG data, which were then transmitted via Bluetooth to a mobile phone application (the Huawei health application) installed on a Huawei mobile phone. Then, the ECG data were transmitted to cloud servers via the mobile phone network. The Huawei health application can store a large number of ECG recordings on a mobile phone. Finally, the 12-lead ECG data and the signal from the single-lead ECG recording of the wristwatch were transmitted to Huawei cloud servers and compared via an algorithm. The data of ECG recordings were accessible to authorized users through cloud servers.

Rhythm analysis

The 12-lead ECG served as the gold standard for measuring cardiac rhythms, and this status was evaluated by two independent cardiologists. These two independent cardiologists, who were blinded to the analysis of the results, reviewed the 12-lead ECG results. If their diagnoses were consistent, the 12-lead ECG result would be regarded as the gold standard; if they were inconsistent, then the 12-lead ECG result would be excluded. Single-lead ECG signals that were affected by the subject's arm and/or finger movement or by improper wearing were also excluded. The analysis results of the algorithm included sinus rhythm, AF, atrial premature beats, and ventricular premature beats. Figure 2 shows the flowchart of the study.

Statistical analysis

The recall, precision, and F1 score were utilized to validate the diagnostic ability of single-lead ECG algorithm for the arrhythmias compared to the 12-lead ECG. The macro-F1 score was calculated to evaluate the comprehensive accuracy of multiple classifications.

The diagnostic ability of the three-class classification system was defined as the ability to detect AF, ectopic beats (atrial premature beats and ventricular premature beats), and sinus rhythm, while the four-class classification system referred to the detection of AF, atrial premature beats, ventricular premature beats, and sinus rhythm.

The F1 score of the three-classification system was calculated as:

$$\mathrm{F}1\mathrm{scor}{\mathrm{e}}_i={\displaystyle \frac{2\times A_{ii}}{{\sum }_j{A}_{ji}+{\sum }_j{A}_{ij}}\times 100\text{\%}.}$$ (1)

The macro-F1 score of three-classification system was calculated as follows:

$$\mathrm{macro}\text{-}\mathrm{F}1\mathrm{score}={\displaystyle \frac{{\sum }_i\mathrm{F}1\mathrm{scor}{\mathrm{e}}_i}{3}\times 100\text{\%}.}$$ (2)

In the above calculations, $i=0,1,2,3;j=0,1,2,3;$ 0 represents sinus rhythm, 1 represents AF, and 2 represents the ectopic heartbeats.

The confusion matrix of the three-classification system results is presented in Table 1.

Table 1.

Confusion matrix of the three-class classification system.

		Predicted cardiac rhythm by algorithm
		0 (sinus rhythm)	1 (ectopic beats)	2 (AF)
Cardiac rhythm	0 (sinus rhythm)	A00	A01	A02
	1 (ectopic beats)	A10	A11	A12
	2 (AF)	A20	A21	A22

Note: AF, atrial fibrillation.

The F1 score of the four-classification system was calculated as follows:

$$\mathrm{F}1\mathrm{scor}{\mathrm{e}}_i={\displaystyle \frac{2\times A_{ii}}{{\sum }_j{A}_{ji}+{\sum }_j{A}_{ij}}\times 100\text{\%};}$$ (3)

The macro-F1 score of the four-classification system was calculated as follows:

$$\mathrm{macro}\text{-}\mathrm{F}1\mathrm{score}={\displaystyle \frac{{\sum }_i\mathrm{F}1\mathrm{scor}{\mathrm{e}}_i}{4}\times 100\text{\%};}$$ (4)

In the above calculations, $i=0,1,2,3;j=0,1,2,3;$ 0 represents sinus rhythm, 1 represents AF, 2 represents atrial premature beats, and 3 represents ventricular premature beats.

The confusion matrix of the four-class classification system results is presented in Table 2.

Table 2.

Confusion matrix of the four-class classification system.

		Predicted cardiac rhythm by algorithm
		0 (sinus rhythm)	1 (atrial premature beats)	2 (ventricular premature beats)	3 (AF)
Cardiac rhythm	0 (sinus rhythm)	A00	A01	A02	A03
	1 (atrial premature beats)	A10	A11	A12	A13
	2 (ventricular premature beats)	A20	A21	A22	A23
	3 (AF)	A30	A31	A32	A33

Note: AF, atrial fibrillation.

Categorical variables were expressed as numbers and percentages and compared with Chi-squared or Fisher's exact tests. A value of P < 0.05 was regarded as significant. Analyses were performed with IBM SPSS Statistics (IBM Corp, Somers, NY) version 22 and Open-Epi (Open Source Epidemiological Statistics for Public Health) version 3.01.

Results

A total of 28 subjects were excluded due to single-lead ECG signals affected by the subject's arm and/or finger movement or by improper wearing. No subjects were excluded due to inconsistent 12-lead ECG diagnoses between the two independent cardiologists. Finally, 1926 ECG signals from 628 subjects (282 men and 346 women) aged 19 to 94 years (median 64 years) were analyzed using the algorithm. The numbers of subjects with AF, atrial premature, ventricular premature, and sinus rhythm were 129, 141, 107, and 251, respectively. Among the 628 subjects (mean age: 63.77 ± 13.90 years; age range: 19–94 years), there were 282 men (mean age: 62.74 ± 13.69 years; age range: 20–93 years) and 346 women (mean age: 64.61 ± 14.01 years; age range: 19–94 years).

There were 129 subjects with AF, 248 subjects with ectopic beats (141 with atrial premature beats and 107 with ventricular premature beats), and 251 subjects with sinus rhythm. A total of 1926 valid ECG samples were collected. The baseline characteristics of subjects with sinus rhythm, ventricular premature beats, atrial premature beats, and AF can be found in Table 3.

Table 3.

Baseline characteristics of subjects with sinus rhythm, ventricular premature beats, atrial premature beats, and AF.

Variable	Sinus rhythm (n = 251)	Ventricular premature beats (n = 107)	Atrial premature beats (n = 141)	AF (n = 129)	P
Age (years)	57.96 ± 11.18	67.12 ± 14.88	62.96 ± 13.51	67.57 ± 13.91	<0.05
Men (n, %)	115 (45.82)	50 (46.73)	57 (40.43)	60 (46.51)	>0.05
Height (m)	168.04 ± 7.71	168.68 ± 17.72	165.41 ± 8.61	168.13 ± 8.16	<0.05
Weight (kg)	72.21 ± 13.68	72.81 ± 18.83	69.09 ± 20.54	75.48 ± 22.56	<0.05

Note: AF, atrial fibrillation.

Diagnostic ability of the three-class classification system

The results of confusion matrix of the three-class classification system's confusion matrix are presented in Table 4. For sinus rhythm subjects, the recall was 97.6%, precision was 96.5%, and F1 score was 97.0%. For AF subjects, the recall was 96.7%, precision was 96.9%, and F1 score was 96.8%. For ectopic beats subjects, the recall was 92.8%, precision was 94.2%, and F1 score was 93.5%. The macro-F1 score of the three-class classification system was 95.8%. The specificity of sinus rhythm, AF, and ectopic beats were 93.8%, 97.6%, and 96.8%, respectively.

Table 4.

Results confusion matrix of the three-class classification system's confusion matrix.

		Cardiac rhythm predicted by algorithm
		0 (sinus rhythm)	1 (ectopic beats)	2 (AF)
Cardiac rhythm	0 (sinus rhythm)	868	6	0
	1 (ectopic beats)	24	561	5
	2 (AF)	3	33	418

Note: AF, atrial fibrillation.

Diagnostic ability of the four-class classification system

The results of confusion matrix of the four-class classification system's confusion matrix are presented in Table 4. For sinus rhythm subjects, the recall was 97.6%, precision was 96.5%, and F1 score was 97.0%. For AF subjects, the recall was 96.7%, precision was 96.9%, and F1 score was 96.8%. For premature atrial contraction subjects, the recall was 90.5%, precision was 89.4%, and F1 score was 89.9%. For premature ventricular contraction subjects, the recall was 86.1%, precision was 89.6%, and F1 score was 87.8%. The macro-F1 score of the four-class classification system was 92.9%. The specificity of sinus rhythm, AF, atrial premature beats and ventricular premature beats were 91.1%, 95.7%, 96.0%, and 95.2%, respectively. Figure 4 shows the waveforms of the 12-lead ECG and the single-lead ECG from the wristwatch (Table 5).

Figure 4.

The representative waveforms of the 12-lead ECG and the single-lead ECG from the wristwatch. A patient is simultaneously tested with a wristwatch and 12-ead ECG. AF, atrial fibrillation.

Table 5.

Results of the four-class classification system’s confusion matrix.

		Cardiac rhythm predicted by algorithm
		0 (sinus rhythm)	1 (atrial premature beats)	2 (ventricular premature beats)	3 (AF)
Cardiac rhythm	0 (sinus rhythm)	868	2	4	0
	1 (atrial premature beats)	6	237	25	3
	2 (ventricular premature beats)	18	3	296	2
	3 (AF)	3	19	14	418

Note: AF, atrial fibrillation.

Out of 454 AF ECG signals, 3, 19, and 14 signals were misdiagnosed by single-lead algorithm as sinus rhythm, atrial premature beats and ventricular beats, respectively, by the single-lead algorithm. Out of 319 ventricular premature beat signals, 18, 3, and 2 signals were misdiagnosed as sinus rhythm, atrial premature beats and AF, respectively, by the single-lead algorithm Out of 271 atrial premature beat signals, 6, 25, and 3 signals were misdiagnosed by single-lead algorithm as sinus rhythm, ventricular premature beats, and AF, respectively, by the single-lead algorithm. Out of 874 sinus rhythm signals, 2, and 4 signals were misdiagnosed by single-lead algorithm as atrial premature beats and ventricular premature beats respectively, by the single-lead algorithm. No sinus rhythm signals were misdiagnosed as AF by the single-lead algorithm. Subjects whose AF was misdiagnosed as atrial premature beats are shown in Figure 5.

Figure 5.

Subject with AF misdiagnosed as atrial premature beats. 4A atrial premature beats; 4B AF. AF: atrial fibrillation.

Discussion

In this study, our main finding was that the single-lead ECG algorithm embedded into smart wearables identified AF, atrial premature beats, ventricular premature beats, and sinus rhythm with high accuracy, comparable to that of a standard 12-lead ECG.

The advancement of wearable technology and the rising popularity of mobile healthcare have enabled remote, continuous and wearable ECG monitoring. Based on a 12-lead ECG, a method for heart beat detection and its associated classification algorithms were developed that include enough information to achieve high accuracy. Achieving real-time analysis of wearable ECG signals with low complexity and high accuracy, in mobile devices with a single-lead ECG is still challenging.

The F1 scores of the present single-lead ECG-based algorithm for sinus rhythm, ectopic beats, and AF were 99.37%, 90.6%, and 90.85%, respectively. De Chazal et al. developed a heartbeat classification algorithm based on ECG morphology and heartbeat interval features for atrial premature beats and ventricular premature beats with the sensitivities of 75.9% and 77.7%, and the positive predictive rates for atrial premature beats and ventricular premature beats were 38.5% and 81.9%, respectively. ¹⁶ Recently, the application of deep learning in, single-lead ECG classification achieved accuracies of 93.63% and 95.57% for the diagnosis of ventricular ectopic beats and supraventricular ectopic beats, respectively. ¹⁷ Another study on single-lead long-term ECGs used a convolutional neural network-based model, and found that its F1 score for ventricular premature beat detection was 92.6%, while that for atrial premature beat detection was 72.2%. ¹⁸

Moreover, the reported diagnostic ability of single-lead ECGs for AF was different. A systematic review conducted in 2021 including 43 studies demonstrated that the single-lead ECG manufactured by AliveCor could achieve sensitivities and specificities of 66.7% to 98.5% and 99.4% to 99.0%, respectively, for AF detection, respectively. ¹⁹ However, in another study, 33.7% Kardia Band recordings were interpreted as “unclassified” by the automated algorithm due to a heart rate <50 or >100 beats/min, noisy tracing, or tracing <30 seconds, ²⁰ indeed, the increased motion artifact may have reduced the diagnostic accuracy of the automated algorithm. Another important factor to consider is the algorithm’s ability to differentiation between AF and ectopic beats.

In our prior study, we developed a PPG-based machine learning model for AF prediction. Although the positive predictive value and negative predictive value were both over 96%, compared with the 72-hour Holter ECG, the false-positive rate of the model predicting AF was 5.58 (95% confidence interval: 5.00–6.22). Among these false positive, 85% were atrial bigeminy, trigeminy, and atrial flutter, indicating that further improvement of the differentiation between atrial arrhythmias and AF is necessary. ²¹

The present algorithm improved this discrimination ability. The macro-F1 score of the three-class classification system was 95.8% with respect to the sinus rhythm, ectopic beats, and AF. Previous studies on the differentiation between AF, ectopic beats and sinus rhythm using wearables are summarized in Table 6. Most of these studies were based on two-channel ambulatory ECGs or a 6-lead ECG. Three studies (Casas et al., ²² Zhipeng Cai et al., ²³ , and Yang et al. ²⁴ ) focused on the classification of normal sinus rhythm and ectopic beats, while the other two studies (Bacevicius et al. ²⁵ and Zhang et al. ²⁶ ) attempted to differentiate AF from sinus rhythm and ectopic beats. The F1 scores of the algorithms and methods used to classify sinus rhythm and ectopic beats were all over 90.0%, except for the F1 score in Yang's study, ²⁴ where it was 72.2% for classifying premature atrial complex. The sensitivities, specificities, and accuracies of the algorithms and methods used to classify AF, sinus rhythm and ectopic beats were all over 94%.^25,26 Our algorithm, which is based on a single-lead ECG showed an ability to distinguish between sinus rhythm, ectopic beats, and AF similar to that of a 12-lead ECG. Moreover, the macro-F1 score of our algorithm based on single-lead ECG for the four-class classification system of sinus rhythm, AF, atrial premature beats, and ventricular premature beats was 92.9%, which was also comparable to the results of previous studies those were based on two-channel ambulatory ECGs or a 6-lead ECG.

Table 6.

Diagnostic ability of algorithms based on wearables for ectopic beats and AF.

Study	N	Databases	Wearables	Algorithm/method	Classifications	PVCs	PACs	AF
Casas et al. 2018 22	108,653 ECG	MIT-BIH database	Two-channel ambulatory ECG recordings	Bayesian classification algorithms	Normal, PVC, and others	F1 score >0.95
Zhipeng Cai et al. 2020 23	1832 10 s ECG segments from single ECG leads from CPSC2018 database; 7903 10s ECG of two-channel ambulatory ECG from MIT-BIH database 1630 10 s ECG segments from single ECG leads of wearable 12-lead ECG	China Physiological Signal Challenge (2018) database; MIT-BIH database; Wearable ECG database	Two-channel ambulatory ECG in MIT-BIH database; Wearable 12-lead ECG	Rule-based premature beat recognition method	Normal, PACs and PVCs	F1 measures for normal beats, PACs, and PVCs were 99.37, 90.6, and 90.85% in training data (93.61% across all beats)
Bacevicius et al. 2022 25	344 participants including 121 patients with AF，95 patients with stable SR and 128 patients with SR and frequent premature contractions		6-lead ECG	PPG-based algorithm; ECG-based algorithm	AF			6-lead ECG: sensitivity 99%, specificity 99% PPG-based algorithm: sensitivity 94%, specificity 97%
Yang et al. 2021 24	47 subjects of MIT-BIH arrhythmia database 42 075 PVCs and 17 511 PACs of China Physiological Signal Challenge (2020)	MIT-BIH arrhythmia database China Physiological Signal Challenge (2020) database	Two-channel ambulatory ECG in MIT-BIH database; single-lead ECGs in the CPSC dataset	Convolutional neural network model	PVCs and PACs	F1 score: s 92.6%	F1 score:72.2%
Zhang et al. 2022 26	47 subjects from the MIT-BIH arrhythmia database 25 long-term ECG recordings of subjects with AF from the MIT-BIH atrial fibrillation database	MIT-BIH arrhythmia database MIT-BIH atrial fibrillation database China Physiological Signal Challenge 2021	Two-channel ambulatory ECG in MIT-BIH database; 12-lead Holter or 3-lead wearable ECG monitoring devices from the China Physiological Signal Challenge 2021	Preprocessing, suspicious AF segment screening, and premature beat screening method.	AF and premature beats		The proposed method eliminated single premature beat segments with 99.5% accuracy, and other premature beats with more than 95% of ECG segments	The accuracy of 96.9%

Note: PVCs: premature ventricular complexes; PACs: premature atrial contractions; PPG: photoplethysmography; AF: atrial fibrillation; ECG: electrocardiogram.

Limitations

There were several limitations of this study. (i) We did not validate other arrhythmias with a single-lead ECG algorithm, for example, atrial flutter, which needed to be further evaluated; (ii) We conducted analyses of the recall, precision, and F1 scores to validate the diagnostic ability of a single-lead ECG algorithm for arrhythmia and compared to a 12-lead ECG, based on the 1926 ECG signals from a relatively small number of participants (a total of 628 subjects were recruited). In the future work, we will test the diagnostic ability of the algorithm in a large cohort of participants to determine the false negative rate for AF diagnosis, that is, determine how many participants with AF would be misdiagnosed by the algorithm as having ectopic beats, and how many would be misdiagnosed as sinus rhythm. (iii) Though the wearables with an embedded ECG algorithm demonstrated good diagnostic ability for AF, atrial premature beats, ventricular premature beats, and sinus rhythm in the hospital, this diagnostic ability also needs to be confirmed for AF screening in a real-world clinical setting; and (iv) how this single-lead ECG algorithm can be applied in practice, for example, together with mAFA, to facilitate patient-centered management, needs to be investigated in the future.

Conclusion

The single-lead ECG algorithm embedded into smart wearables demonstrated good performance in detecting AF, atrial/ventricular premature beats, and sinus rhythm, which would facilitate AF screening and management.