The KardiaMobile app, developed by AliveCor [21], is a commonly used device in studies investigating the monitoring of AF. The equipment comprises a finger pad sensor for the left and right hands, which is attached to the mobile phone, and an app that receives the transmitted information. A 2017 study by Tu et al gives a detailed account of the technology used by the device [22]. When fingers are placed on the sensors, recording begins. Using a 19,000 Hz center frequency and a modulation index of 200 Hz/mV, electrical activity is transmitted to the mobile phone via ultrasound signal frequency modulation to obtain the ECG trace. The frequency and modulation index signal that is transmitted to the microphone of the mobile phone are digitized at 44.1 kHz and 24-bit resolution, respectively. The app is then able to produce an ECG trace from the received signal (300 samples/second, 16-bit resolution), which can either be viewed as it is produced or be stored as a PDF file for later viewing. On the basis of the absence of P waves and irregularity of the RR (R wave to R wave) interval, the associated algorithm within the app produces a potential AF diagnosis [22].

The Lau et al study [23] took a group of 109 patients, 39 of which were in AF, and used the AliveCor algorithm to analyze the rhythms, obtaining a sensitivity of 87% and specificity of 97% when compared with interpretation by a cardiologist. The algorithm was then altered by increasing the weighting of absent P waves in the diagnostic process, and recordings were conducted on this same set of patients. The optimized algorithm was found to have a sensitivity of 100% and specificity of 96%. A second set of 204 patients, with 48 having known AF, were then analyzed by the optimized algorithm, which obtained a sensitivity of 98% and specificity of 97% in this second dataset. Other studies have largely found the sensitivity of the AliveCor device to be in the region of 90% to 98%, with specificity values ranging from 29.2% to 99% in various studies [24-29]. However, a study on the use of AliveCor in inpatients in cardiology and geriatric wards only found a sensitivity of 54.5% and 78.9%, respectively, for these groups [26], whereas another study published in 2017 used the AliveCor algorithm as a control and found a sensitivity of 71.4% for AliveCor [30]. A study of 13,222 patients identified 101 participants to have previously undiagnosed AF, with 65.3% of these patients reporting no symptoms before this diagnosis [24]. Within the field of pediatrics, accurate single-lead ECG tracings were also obtained in both healthy children and children with arrhythmic abnormalities [31]. The device is generally well tolerated, and patients report that it is easy to use [27]. Implementation of the AliveCor in community health care settings has also been investigated. Patients generally responded positively to the suggestion of an ECG being conducted by a pharmacist while they were attending the pharmacy for other purposes [32]. However, some patients were unwilling to have this conducted for the fear of discovering something was wrong. In a primary care setting, the use of AliveCor was well received by patients, but the staff had some reservations about being involved because of a self-perceived lack of knowledge [33]. Patients also report that they felt more confident in managing their AF by using the app [27,34]. The AliveCor device [18,35-37] has also been incorporated with patient education to increase the knowledge of the patients about AF [38]. Mobile apps have been extended to monitor paroxysmal AF for intermittent anticoagulation [39]. This evidently raises public health issues of systematic and opportunistic screening. In a recent evaluation of the European Heart Rhythm Association Consensus on screening AF, opportunistic screening was recommended for individuals aged above 65 years and younger patients at high risk of stroke [40]. As such, it is important to understand local screening guidelines before systematic or opportunistic screenings commence as large number of false positives yields unnecessary, expensive, and potentially harmful follow-ups that might affect patients’ mental health.

The detection of AF by analysis of finger and facial PPG signals using the CardiioRhythm mobile phone app has been tested in 2 recent studies. In the study by Chan et al [30], finger PPG waveforms of 20-seconds duration were captured 3 times using an iPhone 4S equipped with the CardiioRhythm app and compared with the AliveCor single-lead ECG as referenced. Finger PPG waveforms were captured using the light emitting diode flash of the iPhone and the camera, which detected reflected light to measure the arterial pulsation. These waveforms were recorded at 30 Hz, with the duration of each recording being 17.1 seconds, and were filtered with a bandpass filter using a range of 0.7 Hz to 4.0 Hz. AF was diagnosed from this if 2 of the 3 PPG recordings were irregular. This was defined as a lack of repeating pattern, using a Support Vector Machine to classify the waveforms as similar or nonsimilar to other waveforms. Analysis of the diagnoses put forward by the CardiioRhythm app in this study showed a sensitivity of 92.9% and specificity of 97.7% in detecting AF compared with a sensitivity of 71.4% for the AliveCor device that was obtained by this study. This study found the specificities for the 2 devices to be comparable. The first prospective, international, 2-center, clinical validation study (DETECT AF PRO) based on similar finger PPG technology was also conducted recently to demonstrate the feasibility of such apps alone with a 5-min sensitivity and specificity of 91.5% and 99.6%, respectively [41].

In another study by Yan et al, the CardiioRhythm app was used to analyze facial PPG signals in 217 patients recorded using the front camera of an iPhone 6S without physical contact (Figure 1) [42]. Overall, 3 successive 20-second recordings were acquired per patient. Pulse irregularity in 1 or more PPG readings or 3 uninterpretable PPG readings were considered a positive AF screening result. The CardiioRhythm facial PPG app demonstrated a sensitivity of 94.7% and specificity of 95.8%. The positive and negative predictive values were 92.2% and 97.1%, respectively.

A recent study conducted by Mortelmans et al [43] used the PPG app FibriCheck. The patient places the left index finger over the flashlight and camera, holding their finger horizontally and keeping the finger in place for 1 min. To ensure an accurate reading, the screen becomes red when appropriate contact has been made, and the practitioner manually commences the measurement. The amount of light that is reflected onto the camera is then captured and used to calculate the variation in local arteriole blood volume pulse variation. The rhythm of the pulse is then identified based on the RR interval. The app in the mobile phone then judges the quality of the signal based on the detection of the pulse. If the detection was masked with noise, or beats were absent, these recordings were not included in the app’s analysis. When measurements were detected as a good signal, these data were interpreted by the AF algorithm.

Concurrently with the PPG measurement, a single-lead ECG was taken and analyzed by the FibriCheck app. Selection of the data took place based on quality, and good quality measurements were analyzed by the AF algorithm using variability between RR intervals to detect irregularities. The FibriCheck app was used on an iPhone 5S. With 242 participants, this study found a sensitivity of 98%, specificity of 88%, and accuracy of 93% for the FibriCheck algorithm when the data were obtained via PPG. Interpretation by the app of the single-lead ECG performed better, yielding a sensitivity of 98%, specificity of 90%, and accuracy of 94% when compared with standard ECG. When comparing the 2 methods by false-positive results, when analyzed by FibriCheck, 8 false positives were produced from a trace obtained by PPG measurement and 11 false positives were produced from a trace obtained by a single-lead ECG.

Some studies have looked at the use of the inertial measurement unit, using modern microelectromechanical (MEMS) sensors already present in mobile phones, to detect AF [44]. The patient is advised to lie supine and place the mobile phone on their chest. The accelerometer of the mobile phone, which detects the orientation of the phone, can detect cardiogenic movements of the chest. The movements detected are those caused by the opening of the aortic valve, and so, the interval between each successive valve opening is recorded and considered to determine the presence of AF. This yielded a sensitivity of 93.8% and specificity of 100% in the detection of AF. A similar study used the MEMS sensors of mobile phones to obtain measurements, which when analyzed by machine learning methods found a sensitivity of 98.5% and specificity of 95.2% in its best performing method [45].

The feasibility of producing a wearable sensor that records HR and transmits information to a mobile phone app was examined [46]. The way this app identifies AF is similar to the FibriCheck app. As an example, a 42-year-old male admitted with new-onset AF of undetermined duration was deemed appropriate for electrocardioversion because of his FitBit Charge HR device recording the time at which increased HR was observed, allowing the onset of AF to be identified [47]. In a study by Yan et al, the diagnostic accuracy of the CardiioDeepRhythm app, a deep convolutional neural network for detecting AF from the PPG signal acquired using an off-the-shelf wrist-worn device (Empatica E4, Milan, Italy), was tested in 51 in-hospital patients reporting a sensitivity and specificity of 93% and 94%, respectively [48]. These findings demonstrate the promising value of PPG sensors for ambulatory AF monitoring. Krivoshei et al [49] conducted a study investigating the use of PPG in AF, with a view to implementing the technology in a smartwatch; the protocol for this trial was published as an abstract during a conference [50]. Table 1 summarizes the sensitivities and specificities of the apps in AF detection. Despite some studies conducting direct comparisons between different technological modalities [30], it is still done with reference to the gold standard of a conventional 12-lead ECG within a short period. This is especially important because none of the new app technologies have been properly integrated into the wider health care network as compared with the 12-lead ECG, which is a validated instrument commonly available within primary and secondary care settings.

There has been considerable innovation in the use of mobile phone apps for the purpose of HR monitoring in adults. In particular, numerous studies have investigated the potential of 2 specific methods—seismocardiogram (SCG) and PPG—in producing accurate HR measurements. Landreani et al, for instance, described the application of SCG and ballistocardiogram (BCG) signals for HR monitoring apps in mobile phone–embedded accelerometers [51]. This is achieved by determining the RR interval of sufficient amplitude and subsequently detecting the fiducial peak from the SCG and BCG signals. In PPG, the accuracy of this method was also demonstrated in a study by Alaleef et al, who found a 99.7% accuracy and maximum absolute error of 0.4 beats/min [52]. The accuracy and feasibility of PPG signals from mobile phone apps also varied between different types of PPG apps. In Parpinel et al’s study, higher feasibilities and accuracies were found for contact PPG–based apps compared with the noncontact PPG–based ones [53]. Koenig et al [54] found that their algorithm using PPG to assess HR had a consistent accuracy when compared with ECG (correlation index R>.99).

Apart from HR, HRV measurements in mobile phone apps have also been extensively researched. Sometimes coined as the RR interval, HRV is the measure of the variation in time intervals between heartbeats, usually with reference to ventricular contraction. Classically, HRV is measured using the ECG using mainly the time-domain and frequency-domain approaches introduced in 1996 [69,70].

With the use of PPG, Bolkovsky et al used both Android and iPhone mobile phones to obtain RR intervals and subsequently deduce HRV via complex HRV algorithms. Although the results were statistically same as the ECG gold standard, the main issue was the insufficient sampling rate in both phones (20 Hz in Android and 30 Hz in iPhone), which was below the suggested rate of 250 Hz [71]. mobile phone PPG has also been advocated by Plews et al, who reported almost perfect correlations of PPG with the ECGs (R=.99), with an acceptable technical error of estimates and trivial differences in standardized differences [72]. The accuracy and reliability of mobile phone PPGs have also been validated in other studies; in some cases, mobile phone PPGs were found to have comparable accuracies with commonly used HRV computer software programs (R=.92) [73]. Other methods of HRV measurement have also been studied, such as seismocardiography. Although sampling frequency from mobile phone devices accounted for a significant source of error, RR series measurements differed by less than 10 ms in some studies, suggesting the comparable accuracy of this measuring technique [74].

Although many studies support mobile phone PPG, common errors in using this low-cost technology include frequent noise and artifacts in measurement. To reduce the impact of noise, however, Huang et al proposed the use of a continuous wavelet transform denoising technique to extract the pulse signal and subsequently deduce the RR intervals in the denoised signal. The experiments showed low mean absolute errors of only 3.53 ms, highlighting the efficacy of this proposed method in reducing mobile phone PPG errors. The use of this technique for error reduction should thus be investigated further [75].

Although methods have been proposed to reduce the impact of noise and artifacts in HRV measurement, other studies have investigated the effect of mobile phone models on the SD of beat-to-beat error measurement (SDE) of HRV indices. In 1 study, 2 different mobile phone models (Samsung S5 and Motorola) showed significant device influence in the supine posture measurement, with the Motorola model having a higher SDE than the Samsung S5 [76].

Despite being effective in accessing HR and HRV, the applications of PPG monitoring are limited by multiple confounders such as finger pressure, skin tone, light intensities, and user movement leading to artefactual measurements [95,96]. As such, this will have an impact on the feasibility and reliability of mobile phone–based PPG within clinical practice. A minimum sampling rate is necessary for clinically accurate measurements—30 Hz for HR and 200 Hz for HRV measurements [97,98]. However, the frame rate of mobile phones usually operates around 30 Hz, which is a major limitation identified by Bolkhovsky et al [71]. It was proposed that using cubic interpolation, the second derivative, and the zero-crossing algorithm instead of minima detection would overcome this limitation and allow for better HR detection [99-102]. A filter is required to remove artifacts without compromising the original signal when conducting a time-domain analysis to evaluate small variations occurring in normal-to-normal intervals [103]. Examples of filters are independent component analysis using accelerometer data to remove artifacts [104] or employing a fourth-order bandpass filter [99,101]. A heating problem was also noticed by Garcia-Agundez et al when testing their algorithm, which resulted in complaints about the discomfort of holding the mobile phone [102]. Furthermore, although the SCG accelerometers could be used without supporting ECG signals, HR detection was only possible if patients were motionless and supine, leading to considerable variability in data collection because of different measurement positions. There were also difficulties encountered in setting appropriate R-peak thresholds and in comparing the data with simultaneous ECG signals (the gold standard technique) [105,106]. Another area to consider is the use of PP (P wave to P wave) intervals and PRV instead of the RR interval and HRV. Although most studies use RR and HRV approaches with supporting literature that time domain, frequency domain, and Poincare plot HRV parameters computed using RR interval and PP interval methods showed no significant differences, the PP variability was found to be accurate (0.1 ms) compared with RR variability [107]. Furthermore, most studies usually compare between normal sinus rhythm and AF. However, a simple ectopic will be able to derange specificity to unacceptable levels. In this respect, more research needs to done to evaluate the impact of short-term arrhythmic abnormalities on PPG-based and single-lead ECG–based AF detection modalities.

Handheld single-lead ECG recorders, which correspond to the standard lead I, do not cater to potential positional variability. For example, in cases where the heart is positioned more vertically, the amplitude of the QRS complexes may be reduced and comparable with the amplitude of artifacts [15]. Furthermore, a single recorded lead I ECG also does not allow for differentiation between types of narrow and wide complex QRS tachycardia [15]. Although small studies have been conducted to show that it is plausible to use a single-lead ECG to diagnose the ST-segment elevation acute coronary syndrome [108], it is not sufficient to warrant routine clinical use. In some situations, new materials and sensors (eg, biopatches) have allowed these ECGs to be recorded from atypical places (eg, mastoid area) [109] and under different environmental conditions such as after immersion in water or in areas with magnetic fields between 1.5 and 3 T (during magnetic resonance imaging) [110-113]. Novel biopatches have also allowed other parameters such as respiratory rate, body position, temperature, and quality of sleep or physical activity to be monitored, which aids in excluding ST depression caused by increased physical activity or changes in body position [19,114,115]. However, when this is coupled with fewer leads, not only will there be suboptimal signal-to-noise ratio but signals generated might also be different from those produced by standard ECG leads [15]. Furthermore, the utility of biopatches comes with a drawback of having a short intersensor distance, which could potentially compromise the quality of the P wave recorded [15].

Despite being a promising concept, there are currently clear limitations to the use of mobile phone apps for HR, HRV, and AF monitoring. Looking beyond the technological drawbacks, there is little evidence as to how mobile phone monitoring will be matched with health care infrastructural changes to allow for such data to sync with the electronic medical record [116,117]. Furthermore, not all patients will have the same device or mobile phone, which may result in different recorded results. Special clinical attention also has to be given to incidental findings of ECG abnormalities such as ventricular and supraventricular tachycardia in asymptomatic low-risk patients [15]. The impact of app use on health care services is also not to be underestimated. The necessity for a health care professional to confirm arrhythmias detected by these apps will result in a significant burden on services. Given the transient nature of some arrhythmias, an app with a high specificity is likely to be more beneficial to avoid over investigation and treatment of patients, reducing the impact on existing services.

A recent large-scale community AF screening program on more than 10,000 patients aged above of 65 years was conducted, proving the feasibility of integrating this technological within-health care services to identify with new AF [118]. The main drawback illustrated by this study is the lack of downstream management pathways. This raises the medico-legal aspect of introducing mobile phone apps. First, there must be an understanding as to how these apps will support decision making or purport to intervene in clinical decisions. This, in turn, will lay the foundation for a robust government framework to evaluate its effect on clinical outcomes and potential unintended consequences [119,120]. A deeper engagement in this respect was raised in January 2016 when 2 class action lawsuits were filed against such manufacturers, calling into question the reliability and accuracy of these devices and apps [121]. Perhaps the issue is in classifying these devices and apps as certified medical devices. Despite the intuitive need for medical devices to be of high quality, local regulations for medical devices have to be followed. Such regulations range from the Medical Devices Framework concept of intended purpose to the risk-based case-by-case approach employed by the US Food and Drug Administration [122].

Looking specifically at the European regulations, the Medical Device Directive was replaced by the new Medical Device Regulation. Although the new regulatory framework still revolves around intended purpose, the bar for medical device classification has been lowered because of a broader definition assigned to the term medical purpose. In short, despite the undisputed potential of such apps in the detection of arrhythmias, adhering to appropriate laws and regulations remains a significant hurdle to address [123].