Table 3.

Summary of the usage of wearable devices in cardiovascular care.

Cardiovascular care	Reference	Wearable device	Technology	Accuracy validity
Arrhythmia detection∗	Hannun et al[61]	Zio patch (iRhythm Technologies, USA)	Deep neural network	ROC of 0.97 to classify 12 rhythm classes on a test dataset consisting of 328 ECG records from 328 patients.
	mSToPS Trial 2018[62]	Zio patch (iRhythm Technologies, USA)	Cox proportional model	ECG monitoring immediately resulted in a higher rate of AF diagnosis than after 4 months (3.9% vs. 0.9%). Compared with non-monitored controls, participants who received monitoring had a higher rate of AF diagnosis and greater initiation of anticoagulants at 1 year.
	Apple heart study 2019[63]	Apple Watch	Deep neural network	Of the 86 participants who received irregular pulse notifications, 72 showed AF on concurrent ECG patch strips, with a PPV for the irregular pulse notification of 0.84.
	Health eHeart Study 2018[64]	Apple Watch	Deep neural network	The DNN exhibited a C statistic of 0.97 to detect AF in the external validation cohort of 51 patients undergoing cardioversion. In an exploratory analysis relying on self-reported persistent AF in ambulatory participants, the C statistic was 0.72.
	Bumgarner et al[65]	Apple Watch combination with the AliveCor KardiaBand		Compared with ECG, the Kardia Band interpreted AF with 93% sensitivity, 84% specificity, and a K coefficient of 0.77. Among 113 cases where Kardia Band and physician readings of the same recording were interpretable, the agreement was excellent (K coefficient = 0.88).
	Huawei Heart Study 2019[66]	Huawei Watch GT, Honor Watch, and Honor Band	PPG algorithm	Of 424 participants who received a “suspected AF” notification, 227 individuals were confirmed as having AF, with the PPV of PPG signals being 91.6%.
	Chen et al[67]	Amazfit Health Band 1S (Huami Technology, Anhui, China)	Deep learning (SEResNet)	The sensitivity, specificity, and accuracy of wristband PPG readings to detect AF were 88.00%, 96.41%, and 93.27%, respectively, and those of wristband ECG readings were 87.33%, 99.20%, and 94.76%, respectively.
BP measurement†	Watanabe et al[70]	Cuff-less BP estimation	PPG algorithm	The MAD between the BP value of CLB and the cuff-wearing sphygmomanometer was <8 mmHg (6.1 for SBP).
	Moon et al[71]	InBodyWATCH	Neural network	The ME was 2.2 ± 6.1 mmHg for SBP and −0.2 ± 4.2 mmHg for DBP; these were not significant (P = 0.472 for SBP and P = 0.880 for DBP). The estimated SBP/DBP ratios obtained from the InBodyWATCH within ± 5 mmHg of manual SBP/DBP were 71.4%/ 83.8%; within ± 10 mmHg, they were 86.7%/98.1%; and within ± 15 mmHg, they were 97.1%/99.0%.
	Chandrasekhar et al[72]	A smartphone	Stepwise regression	The smartphone-based device yielded bias and precision errors of 3.3 and 8.8 mmHg for SBP and −5.6 and 7.7 mmHg for DBP over a 40- to 50-mmHg range of BP.
	Van Helmond et al[73]	Everlast watch, BodiMetrics	Statistics	The average differences between the Everlast smartwatch and hospital-grade automated sphygmomanometer were systolic BP of 16.9 mmHg ± 13.5 mmHg and diastolic BP of 8.3 ± 6.1 mmHg. The average difference between the BodiMetrics performance monitor and hospital-grade automated sphygmomanometer was systolic BP of 5.3 ± 4.7 mmHg.
Diabetes, high cholesterol, etc.‡	Ballinger et al[74]	Fitbit, Apple Watch, and Wear OS	Deep learning	Popular wearable devices showed high accuracy at detecting multiple medical conditions, including diabetes (0.8451), high cholesterol (0.7441), high BP (0.8086), and sleep apnea (0.8298).
Diabetes detection‡	Avram et al[75]	Azumio Smartphone app	Deep neural network	The network achieved an area under the curve for prevalent diabetes of 0.766 in the primary cohort and 0.740 in the contemporary cohort.
Hyperkalaemia diagnosis‡	Galloway et al[76]	AliveCor ECG	Deep neural network	The DNN was trained to detect hyperkalemia using only ECG leads I and II. The sensitivity by duration was 94%, and the specificity was 74%.
Myocardial infarction‡	Sopic et al[77]	SmartCardia INYU	Machine learning	The classifier that uses all available features (n = 72) reaches a geometric mean accuracy of 83.26% (Sensitivity = 87.95%, Specificity = 78.82%).
Arterial stiffness measurement‡	Miao et al[78]	SIAT 3-in-1	Machine learning	Using Omron arterial stiffness equipment as the reference, the proposed model achieved the best accuracy of 0.89, 0.2136, and 6.2432 in the correlation coefficient, ME, and standard difference for vascular age estimation.
Cardiovascular risk assessment‡	Women's Health Study[79]	ActiGraph GT3X+	Statistics (proportional hazards regression)	A strong inverse association between overall volume of PA and all-cause mortality was observed. The magnitude of risk reduction (about 60%–70%) was far greater than that estimated from meta-analyses of studies using self-reported PA (about 20%–30%).
	Akbulut et al[80]	CVDiMo	Machine learning	The use of PA results and stress levels deduced from the emotional state analysis resulted in better risk estimation. The highest accuracy of classifying the short-term health status was 96%.

^∗ECG patches can classify multiple rhythms with acceptable accuracy, while PPG-based devices showed acceptable accuracy in detecting AF.

^†Wearable devices showed acceptable accuracy in controlled settings, the validation in large-scale population is rare.

^‡Research on the use of wearable devices in detecting or predicting CVDs is rare; preliminary studies showed the potential of wearable devices in cardiovascular healthcare. AF: Atrial fibrillation; BP: Blood pressure; CVD: Cardiovascular disease; CVDiMo: Cardiovascular Disease Monitoring; CLB: Cuff-less BP estimation; DBP: Diastolic blood pressure; DNN: Deep neural network; ECG: Electrocardiogram; MAD: Mean absolute difference; ME: Mean error; PA: Physical activity; PPG: Photoplethysmography; PPV: Positive predictive value; ROC: Receiver operating characteristic curve; SBP: Systolic blood pressure.