Wearable Photoplethysmography Devices

The wearables market has expanded greatly in recent years, with wrist-worn devices now widely used. Smart wearables provide opportunity to monitor health and fitness in daily life. Wearables such as fitness bands and smartwatches routinely monitor the photoplethysmogram (PPG) signal, an optical measure of the arterial pulse wave which is strongly influenced by the heart and blood vessels. This Chapter presents a comprehensive overview of the state-of-the-art of wearable photoplethysmography devices. It summarises: (i) key considerations in the design of wearable PPG devices; (ii) the physiological parameters that can be estimated from wearable PPG signals; (iii) commercially available devices; and (iv) potential applications in health and fitness monitoring.


INTRODUCTION
The growing use of smart wearables which measure the photoplethysmogram (PPG) provides a wealth of opportunities to monitor health and fitness in daily life. Wearables are becoming more widely used, ranging from fitness bands to smart rings. It has been estimated that by 2022 there will be over 1 billion wearables globally [1], and annual spending on wearables will exceed $80 billion [2]. The growth in wearables is expected to be increased further by their potential for remote monitoring during the COVID-19 pandemic [3]- [5]. Initially, most wearables were used for electrocardiogram (ECG) or activity monitoring, whereas nowadays many smart wearables include a PPG sensor. Early examples of PPG-based wearables include the Mio Alpha (Mio Global, Canada) and Schosche myRhyhm (Schosche Industries, CA, USA) devices [6], both of which appeared in early 2013.
The PPG signal is an optical measure of arterial blood volume. In wearable devices it is typically measured by shining a light on to the skin, and measuring the amount of light reflected back from the skin. The resulting signal is dominated by a pulse wave due to the change in blood volume with each heartbeat (see Fig. 1). Many consumer devices use the PPG for heart rate (HR) monitoring. As well as the heart, the PPG is also influenced by the vascular, respiratory and autonomic nervous systems. Consequently, a range of physiological parameters can be estimated from the PPG, which in the future could facilitate extensive, unobtrusive health monitoring.

Overview
This chapter provides a comprehensive overview of the state-of-the-art of wearable PPG-based devices, and highlights areas for future research to realise their full potential. It covers the following aspects of wearable photoplethysmography devices:  Source: [8] under CC BY 4.0. Data from the PPG Diary Pilot Study 1 dataset [9].

HARDWARE CONFIGURATIONS FOR WEARABLE PHOTOPLETHYSMOGRAPHY DEVICES
Wearable photopethysmography devices are now widely available in a variety of hardware configurations. This Section outlines important factors in hardware design. These factors should be considered alongside the intended application, as the hardware configuration can influence the utility of the acquired PPG signal for estimating physiological parameters. Herein, wearables are defined as devices which attach to (and can be detached from) the body, acquire physiological measurements, and are physically disconnected from the external environment [10].

Measurement site
Wearable PPG devices can acquire PPG signals at a range of anatomical sites. The choice of PPG measurement site can influence both the utility of the acquired PPG signals, and the user acceptability of the wearable device. The shape of PPG pulse waves differs between sites [11], [12], as shown in Fig. 2 (a). This may affect the utility of the pulse wave for physiological parameter estimation using pulse wave analysis. The pulse arrival time (PAT, the time delay between ventricular contraction and PPG pulse wave arrival) is greater at sites further from the heart such as fingers or toes [11], which may influence the utility of PAT measurements. Finally, the form of devices naturally differs between measurement sites, ranging from wristbands to armbands, earbuds to glasses, and therefore the choice of site may influence user acceptability. Fig. 3 shows several examples of PPG-based wearables worn as a headband, armband, and wrist watches.  Source: [13] under CC BY 4.0. Data from the Vortal dataset [14].
The following measurement sites are suitable for wearable devices: 2.1.1. Finger: Smart rings are now commercially available which acquire PPG signals at the finger [16]. The PPG sensor can be placed on the underside of the finger in order to obtain a reflectance PPG signal from as close as possible to the main arteries in the finger. Pulse oximeters typically acquire transmission PPG signals at the finger, and much research has been conducted using PPG signals acquired at the finger [17], particularly as the PPG signals in the widely used MIMIC Database [18] are mostly measured at the finger.
2.1.2. Wrist: Fitness bands and smartwatches acquire PPG signals at the wrist. The PPG sensor is typically housed within the same unit as the display, and typically acquires reflectance signals at the upper wrist. The major arteries of the wrist are located on the lower wrist, centimetres from the typical measurement site, indicating that the PPG signal at the upper wrist likely originates from the microvasculature rather than the major arteries. Indeed, PPG signals acquired close to the major arteries of the wrist have been found to have much higher signal-to-noise ratios than those acquired at the upper wrist [19]. the shape and amplitude of PPG pulse waves has been observed to differ between the finger and wrist [20]. Nonetheless, wrist PPG devices mounted on the upper wrist can provide accurate HR measurements [21].
2.1.3. Arm: Armbands can be used to acquire reflectance PPG signals at the upper arm [22], [23], although this approach is not yet widely used in consumer devices. The arm may be less susceptible to motion artifact than more peripheral sites such as the finger or wrist [24].
2.1.4. Ear: The ear has the potential advantage of being less prone to motion artifact [25]. Wearable PPG devices at the ear can take several forms. Firstly, earring sensors can be used to obtain PPG signals from the earlobe [25]. Secondly, sensors have been applied to the skin immediately behind the ear using a wrap around 'ear cup' system [26]. Thirdly, earbud sensors can be used to obtain PPG signals at the inner ear [27], [28], with the potential advantages that this site is less prone to vasoconstriction that peripheral sites [29], and that noise can be removed through simultaneous acquisition of PPGs using an earbud in each ear [27].
2.1.5. Chest: Chest-worn devices, potentially attached via a chestband or adhesive patch, can acquire PPG signals [30]. The chest is an ideal location for acquiring electrocardiogram and seismocardiogram signals, indicative of heart activity.
2.1.6. Face: Smart glasses have been designed to acquire PPG signals at either the nose bridge [31] or the temple [32]. It may be beneficial to acquire PPG signals at the temple during exercise as HRs derived from PPG signals at the temple can be more accurate than those acquired at the wrist [33].

Sensor design
There are several options to consider when designing a PPG sensor for a wearable device, all of which can influence the utility of the measured PPG signal. This is demonstrated in Fig. 2 (b), which shows PPG pulse waves acquired simultaneously from adjacent fingers on the same hand using two different devices: the pulse wave shape differs, indicating that the hardware configuration could influence parameters extracted from the pulse wave shape. These hardware options are now considered in turn.
2.2.1. Transmission and reflectance photoplethysmography: Wearable PPG devices use either transmission or reflectance photoplethysmography. In transmission photoplethysmography light is shone onto an extremity (e.g. finger, toe or earlobe), and the amount of light transmitted through the extremity is measured using a photodetector on the opposite side. Consequently, transmission photoplethysmography can only be used at limited anatomical sites where a photodetector can be placed opposite the light emitting diode (LED), such as the finger, toe or earlobe. In reflectance photoplethysmography light is shone onto the skin, and the amount of light reflected back is measured using a photodetector positioned close to the emitting LED. Reflectance photoplethysmography can be used at additional sites such as the wrist, arm or chest. It has been observed that reflectance photoplethysmography provides a higher signal-to-noise ratio at the fingertip than transmission photoplethysmography, indicating that the reflectance mode should be used in wearables [34].
Further details of transmission and reflectance photoplethysmography are provided elsewhere in this book: see Chapter 2 "Light Tissue Interaction in PPG" and Chapter 3 "PPG Technology" for details.
2.2.2. PPG wavelength: The wavelength of light used to acquire signals affects the resulting PPG measurement [35]. Wearable devices typically use infrared (longest wavelength), red, or green (shortest wavelength) light. The longer the wavelength of transmitted light, the greater the depth to which the light penetrates into the body [36], so green light penetrates less deeply than red and infrared light [37]. Consequently, red or infrared light is typically used for transmission photoplethysmography. In contrast, green light has been found to provide a higher signalto-noise ratio than red or infrared light in reflectance photoplethysmography [34], [38], and to be more robust to changes in temperature [39]. Indeed, Apple Watches equipped with LEDs of multiple wavelengths (green, red and infrared) can switch between using infrared light for HR monitoring at rest, and green light during exercise [40]. The wavelength of light can also influence the shape of the acquired PPG waveform (see Fig. 1 of [41]). The wavelength of light has been investigated as a factor in ensuring that PPG sensors perform well across different skin types [38]. Reassuringly, research to date has found the performance of certain PPG-based wearables not to be affected by skin type [42], [43].
2.2.3. Multiple PPG signals: Recently, PPG sensors have been designed to acquire multiple PPG signals at a single site in order to reduce the influence of noise or provide additional physiological information. Multiple PPG signals acquired at a single site can be used to assess signal quality by assessing the level of similarity between the signals. Multiple PPG signals can also be used for noise reduction by: identifying the highest quality signal and discarding the others; extracting a composite signal as a sum of the individual signals weighted according to their quality [19]; or using multichannel decomposition to extract significant signal components by exploiting information from all of the signals [44]. Multiple PPG signals of different wavelengths can be used to obtain PPG signals corresponding to different levels of the vasculature, such as the capillaries, arterioles and arteries, which can be used to estimate parameters such as the arteriolar pulse transit time (PTT) [41]. In addition, it has been proposed that an infrared PPG signal can serve as a reference motion signal in order to reduce the influence of tissue deformation on green PPG signals [45].
2.2.4. Sampling frequency: Wearable devices typically sample the PPG at between 50 and 100 Hz [46], [47]. However, studies have shown that HR can be estimated from PPG signals sampled at 9 Hz [48], pulse rate variability can be accurately assessed from PPG signals sampled at 25 Hz [49], respiratory rate can be accurately estimated using a sampling frequency of [16][17][18], [48], and generally PPG features can be accurately measured at sampling frequencies of at least 60 Hz [38]. Therefore, a relatively low sampling frequency may be acceptable for many applications.

PPG bandwidth:
The bandwidth of PPG signals affects their potential utility for parameter estimation. It is beneficial to strongly filter PPG signals prior to HR estimation, using a narrow bandwidth corresponding to the range of plausible HRs such as 0.4-2.25 Hz [50]. This narrow range ensures that most irrelevant content is eliminated, such as low frequency variations due to respiration, and high frequency noise. Having a narrow range can assist with HR and SpO 2 estimation [51]. However, the frequency content of PPG signals outside of this narrow range is required for estimating other physiological parameters. A higher low-pass cut-off is required to preserve high frequency content so that fiducial points on PPG pulse waves can be accurately identified, which can be set as low as 5 Hz for detecting pulse troughs for inter-beat-interval calculation [52], but must be much higher (e.g. 20 Hz [53]) to locate fiducial points accurately enough to measure pulse wave features. Similarly, a lower high-pass cut-off of approximately 0.05 Hz may be required to ensure pulse wave features are faithfully reproduced [54], and that all respiratory content is preserved [55]. Consequently, wearable devices may need to apply different filters to the signal for different purposes.
2.2.6. Measurement frequency and battery life: PPG sensors consume substantial power which can reduce the battery life of wearables [56]. Therefore, rather than continuously monitoring PPG signals, it may be beneficial to turn on PPG sensors either intermittently, or only during periods of low activity. One approach to maintain PPG monitoring whilst reducing power consumption is to only activate the PPG sensor during periods of low activity, since PPG signals are typically of low quality during periods of high activity [8]. This can be achieved by using an accelerometer, which consumes much less power than a PPG sensor, to measure activity levels. Indeed, some wearables now adopt this approach [57]. An alternative approach is to turn off the PPG sensor after a high quality signal segment has been obtained, and to only turn it on again after a certain delay [58]. Such strategies will help ensure that PPG-based wearables are acceptable to users, without requiring too frequent battery charging.
2.2.7. LED driving schemes and power consumption: A key factor in determining the power consumption of a PPG sensor is the proportion of time for which the LED is illuminated [59]. Typically, the LED is duty-cycled, meaning the PPG signal is uniformly sampled by illuminating the LED at a regular sampling frequency [60]. Both the sampling frequency and the duration of time for which the LED is illuminated when taking each sample are a compromise between power consumption and signal fidelity (which influences the accuracy of derived parameters [59]). Alternative approaches have been proposed which use non-uniform sampling to reduce power consumption [60]. Firstly, compressive sampling can be used to reduce the number of samples per pulse wave and still maintain accurate analyses by exploiting the sparseness of the PPG pulse wave [61]. Secondly, windowing can be used to only sample the PPG at points of interest during the pulse wave, such as around the expected time of pulse wave peaks for inter-beat-interval analysis [60]. The appropriate scheme for driving an LED should be chosen with the intended analyses in mind.

Additional signals
Many wearable devices acquire additional signals which can be used with the PPG to provide improved physiological monitoring. Fig. 4 shows four additional signals which are commonly acquired by wearables, and are now described.
2.3.1. Electrocardiogram: The electrocardiogram ( Fig. 4(a)) is a measure of the heart's electrical activity which can be measured by smart wearables either continuously or when activated by the user. The ECG can be monitored continuously between two skin electrodes measuring the potential difference caused by myocardial activity. The signal-to-noise ratio of the ECG tends to be higher when the separation of the electrodes in relation to the heart is greater. For instance, the signal-to-noise ratio is greater when the ECG is measured at the chest, or at two arms (providing a view across the heart), and significantly lower when measured with both electrodes on a single arm [65]. The ECG can also be acquired intermittently when activated by the user of a device (such as a smartwatch) which has one electrode constantly in contact with the skin (typically on the underside of the watch), and a second electrode which the user touches with their opposite hand [66]. An ECG can be recorded for as long as the user has their opposite hand on the device. The inclusion of ECG technology into PPG-based wearables could enhance the utility of the PPG-based wearables. For example, if an irregular heart rhythm is detected then the user could be prompted to record an ECG which could be used to confirm a diagnosis (such as atrial fibrillation). In addition, intermittent pulse arrival time (PAT) measurements could be obtained from simultaneous ECG and PPG measurements, which could be used for blood pressure (BP) estimation.
2.3.2. Accelerometry: Accelerometers measure static and dynamic acceleration. Due to their power efficiency and low price they are already used in the majority of wearables for step counting and recognising some activities (such as walking). Accelerometry signals ( Fig. 4(b)) can also be used to improve parameter estimation from the PPG. Firstly, accelerometry signals can be used to reduce noise in PPG signals by cancelling noise which is common to both accelerometry and PPG signals [67]. Secondly, accelerometry can be used to identify periods when activity levels are too high to estimate parameters reliably from PPG signals. Thirdly, accelerometry could be used to contextualise PPG-derived parameters according to the activity in which they were recorded, as accelerometry can be used to infer body position (e.g. lying or standing), and could be used to identify a wide range of activities of daily living [68].

Seismocardiography:
The seismocardiogram (Fig. 4(c)) is a measure of surface vibrations acquired by an accelerometer in contact with the skin. It can be particularly informative when measured at the chest, where it is influenced by heart activity and can be used to identify the times of ventricular ejection [69]. Pulse transit time can be calculated from the PPG and the seismocardiogram [70], with potential utility for BP monitoring.
2.3.4. Gyroscope: Gyroscopes measure angular velocities around orthogonal axes and thus are suitable for capturing rotational movements. This feature can be used for adaptive motion artifact cancellation from PPG signals [71]. In addition, gyrocardiography was recently proposed as a noninvasive monitoring method for the assessment of cardiac mechanics [72].
It has also been proposed that skin conductance measurements could be used to assess the quality of contact between the skin and a wearable sensor [73]. Some wearables include a galvanic skin response sensor, which whilst commonly used to assess emotions, could also be used to assess skin contact, allowing PPG signals acquired during poor skin contact to be rejected.

PHYSIOLOGICAL PARAMETERS
Several physiological parameters can be estimated from the PPG, and potentially many of these could be integrated into wearables. At present, HR is the parameter most commonly estimated from the PPG by wearables. Other parameters such as BP and respiratory rate (RR) are provided by some wearables, whilst the many of the parameters proposed in the literature have not yet been incorporated into wearables. Techniques for estimating key physiological parameters from the PPG are now described.
Further details relating to some of the parameters are provided elsewhere in this book: see primarily Chapter 4 "PPG Signal Analysis and Synthesis", as well as Chapter 5 "PPG in Oxygenation", Chapter 9 "PPG in Autonomic Function", and Chapter 12 "PPG in Noninvasive Cuff-less BP Monitoring".

Heart rate (HR)
The PPG signal is dominated by pulse waves generated by the ejection of blood from the heart during each heartbeat. Consequently, most PPG-based wearables provide HR measurements. HR is typically estimated from the PPG in four steps [74]: 1) The PPG signal is band-pass filtered to eliminate frequency content outside of the range of plausible HRs.
2) Motion artifact is removed from the PPG, potentially with the aid of accelerometry signals.
3) An initial HR estimate is obtained by analysing the frequency spectrum of this processed PPG signal. 4) A tracking algorithm is used to track HR estimates over time, helping to eliminate erroneous HRs. It is relatively straightforward to estimate HR from a high quality PPG signal in the absence of movement. However, the task is altogether more challenging during exercise which can cause substantial movement artifact in the PPG [75]. This is demonstrated in Fig. 5: at rest (upper raw) the dominant peak in the PPG spectrum coincides with the dominant peak in the ECG spectrum, indicating the true HR. HR estimation is more challenging in this example during walking (middle row): whilst there is a spectral peak corresponding to the true HR at ≈1.4 Hz, care must be taken not to identify the peaks at lower frequencies as the HR. The interested reader is referred to [76] for further details of the signal processing techniques used to remove motion artifact from PPG signals and track HRs during motion. Frequency (Hz) Fig. 5. PPG signals and corresponding frequency spectra: PPG signals acquired during rest, walking and exercise, shown alongside heart beats identified from a simultaneous chest ECG signal, and the corresponding frequency spectra of the PPG and ECG signals.
A recent meta-analysis of studies of the performance of wrist-worn devices for HR monitoring found summary mean absolute errors of 2.15 (95% confidence interval 1.84 to 2.46) bpm during rest, compared to 7.70 (6.32 to 9.07) during treadmill activities [21]. Similar reductions in performance with exercise were reported in [42], and differences in performance between brands were reported in [42], [78]. Future assessments of performance will benefit from recent recommendations for determining the validity of devices [79].
A key parameter derived from HR measurements is the resting HR. Resting HR varies greatly between individuals, and is a marker of cardiovascular risk [80], [81]. Resting HR varies with age [82], circadian rhythms [83], and across seasons [81]. In the future wearables could be used to track an individual's resting HR over time, with changes in resting HR potentially providing valuable insight into their health [84].

Identifying an irregular pulse
PPG-based wearables provide opportunity to detect an irregular pulse in daily life. Irregular heart beats are typically identified from a PPG signal segment as follows [57]: 1) Individual pulse waves corresponding to individual heart beats are identified.
2) Inter-beat-intervals are calculated from consecutive pulse waves, using a fiducial point such as the systolic peak.
3) The irregularity of inter-beat-intervals is assessed statistically, using a technique such as a Poincaré Plot [85]. 4) If the level of irregularity exceeds a threshold then the segment is deemed to contain an irregular pulse. This can be performed using segments of 1 minute duration [57], although it is possible to use this approach with shorter segments to identify possible arrhythmias [86]. Irregular heart beats can occur in healthy individuals without being a cause for concern. However, an irregular pulse can also be caused by arrhythmias such as atrial fibrillation, in which case it is clinically useful to be aware of it. Consequently, it may be helpful to only raise an alert of an irregular pulse if it is observed consistently over multiple recordings (as performed by the Apple Watch [57]). Techniques to identify atrial fibrillation from the PPG have recently been reviewed in [87], [88].

Arterial oxygen saturation (SpO 2 )
Arterial blood oxygen saturation is the proportion of haemoglobin in the blood which is carrying oxygen. It is widely used in clinical practice, and denoted as SpO 2 when measured using pulse oximetry. SpO 2 can be estimated from two simultaneous PPG signals of different wavelengths as follows [89]: 1) Identify individual pulse waves in the two PPG signals.
2) Calculate the normalised AC component for each signal (defined as the total intensity of light at the systolic peak divided by the total intensity at the pulse onset). 3) Calculate the ratio of the normalised AC components for each signal. 4) Estimate SpO 2 from an empirical relationship between SpO 2 and the ratio of normalised AC components (see also [90]). This approach can be aided by the use of a heart-rate tuned comb filter [91]. However, it is difficult to obtain reliable estimates of SpO 2 in the presence of motion artifact. An alternative approach is used in Masimo pulse oximetry which is designed to be more robust during motion and poor tissue perfusion [92], although it is more computationally complex [91].
The accurate estimation of SpO 2 requires consideration of several aspects of wearable design. Firstly, PPG wavelengths should be carefully chosen by considering the absorption spectra of oxygenated and deoxygenated haemoglobin. Secondly, when using reflectance mode photoplethysmography, the spacing between the LED and photodetector in the PPG sensor should be positioned far enough apart to obtain sufficient signal quality [93]. Thirdly, it is important that good skin contact is maintained during SpO 2 measurements, noting that contact pressure might be influenced by body position [94]. There is potentially great benefit in monitoring SpO 2 , particularly to identify obstructive sleep apnea and as a potential indicator of COVID-19 [95].

Respiratory rate (RR)
RR, the number of breaths per minute, is used in clinical practice as a marker of illness. Much research has focused on developing algorithms to estimate RR from the subtle modulations to the PPG signal caused by breathing [55]. The three main modulations of the PPG due to breathing are baseline wander (BW), amplitude modulation (AM), and frequency modulation (FM) [77], [96], as illustrated in Fig. 6. Respiratory rate can be estimated from the PPG by (see [77] for further details): 1) Identifying individual pulse waves. 2) Extracting pulse wave features indicative of a respiratory modulation, such as the amplitude of pulse waves (indicative of AM), or the inter-beat-intervals (FM). 3) Extracting a respiratory signal from the time series of features. 4) Estimating RR from the respiratory signal using either a time-or frequency-domain technique. 5) (optionally) Combining RR estimates derived from different respiratory signals. There are challenges to estimating RR from wearable PPG signals, such as identifying the most appropriate respiratory signal(s) for use with the target population [14], only estimating RR when the respiratory modulations in the PPG are sufficiently strong [97], and avoiding erroneous detection of other low frequency phenomena rather than breathing [98]. Nonetheless, some wearables do now provide RR (see the Biostrap Evo). It could prove to be a helpful parameter for the detection and management of COVID-19 [95], [99].

Blood pressure (BP)
PPG-based wearables would be greatly enhanced if they could assess BP, since it is used for a wide range of clinical purposes. It may be feasible to assess BP from the PPG because it influences two aspects of the pulse wave: the shape of the pulse wave and the speed of its propagation from the heart to the periphery. A plethora of approaches to assess BP from the PPG have been proposed, mostly exploiting one or both of these phenomena. Firstly, BP can be assessed from the shape of the pulse wave and its derivatives [100] (see [101], [102] for lists of PPG features). Secondly, BP can be assessed from the pulse arrival time measured between a marker of cardiac contraction (such as the QRS complex of the ECG) and the subsequent PPG pulse arrival [11], [103]. Thirdly, BP can be assessed from the pulse transit time between PPG pulse waves measured closer to, and further away from, the heart [104]. Machine learning has been widely used to develop techniques to estimate BP from PPGbased measurement [105]. It is challenging to assess BP from the PPG because: changes in BP may have only a relatively small effect on the PPG; the nature of the effect of BP on the PPG is partially subject-specific; and, other cardiovascular properties (such as arterial stiffness) can cause similar changes to the PPG. Consequently, PPG-based devices may require calibration with reference BP measurements in order to monitor BP (see the Samsung Galaxy Watch3). In addition, manufacturers may decide to only assess changes in BP (or the direction of BP changes), rather than absolute values [106].

Sleep assessment
Wearables are extensively used for sleep assessment, based primarily on monitoring activity levels (actigraphy) [107]. PPG-based wearables which also include an accelerometer provide opportunity to enhance sleep assessment through the inclusion of HR and PRV analysis. Indeed, consumer PPG-based wearables have been found to identify periods of wake and sleep with a performance similar to, or better than, research-grade actigraphy devices [107], [108]. Error rates vary considerably between PPG-based wearables [109]. Further research is required to determine whether sleep stages can be accurately classified using PPG-based wearables [110].

Energy expenditure (EE)
Energy expenditure (EE) is now measured by some PPG-based wearables. EE primarily consists of three components [111], [112]: basal (i.e. baseline) EE, the thermic effect of feeding, and energy expended from physical activity. The most important component to monitor is that relating to physical activity, since it accounts for the most variable portion. On the other hand, basal EE can be predicted from age, sex and height [112], and the thermic effect of feeding is relatively small. Energy expended from physical activity can be estimated from HR, which indicates the intensity of exercise [113], [114]. This is based on the assumption of a linear relationship between HR and oxygen consumption, which can be considered consistent for an individual performing submaximal activities [115]. However, the true relationship varies between individuals and between exercise types. Furthermore, care must be taken not to include periods of rest and light activity in calculations of energy expended from physical activity, since the relationship does not hold in these cases. The accuracy of EE estimates has been questioned [116], [117], and none of the wearables included in a recent systematic review were found to accurately estimate EE [78].

Maximal oxygen consumption
The maximum rate of oxygen consumption during exercise (denotedVO 2max ) is an important marker of cardiorespiratory fitness [118]. Broadly, two approaches are taken to estimateVO 2max . The first consists of estimatinġ VO 2max from an individual's resting HR and maximum HR (the maximum possible HR for an individual, which can be estimated as a function of gender and age [119]). The second approach is based on the relationships between oxygen consumption and running speed, and between running speed and HR [120]. HR and speed measurements are obtained during exercise, and used to derive a relationship between HR and running speed for the individual. This relationship is then extrapolated to calculate a maximal running speed corresponding to the maximal predicted HR [120]. Finally,VO 2max is estimated from the maximal running speed [120]. Devices such as Fitbits estimatė VO 2max from HR, speed (obtained via GPS), resting HR, and demographics [118]. Studies of the accuracy oḟ VO 2max estimates provided by PPG-based wearables have reported mixed results [116], [118], [121], [122]. In the future it will be important to assess whether this approach has utility in subjects who do not or cannot run [118].

Pulse rate variability (PRV)
PRV is the variability in inter-beat-intervals assessed from the PPG. PRV has several potential applications, including assessing mental stress levels, identifying sleep stages, and assessing cardiovascular health [123]. PRV is highly related to HR variability (HRV), the variability in inter-beat-intervals assessed from the ECG (the gold standard). However, PRV and HRV are not equivalent because they are caused by different physiological mechanisms [124]. A range of summary statistics can be used to quantify PRV and HRV from inter-beat-intervals [125]. The level of agreement between the PRV and HRV statistics is dependent on several technical factors. Firstly, PPG sampling frequency affects PRV measures [49], [126]. Secondly, interpolation of the PPG to higher sampling frequencies can improve agreement [126]. Thirdly, PRV has been found to agree more closely with HRV when calculated from the timings of particular fiducial points on the PPG (namely the middle-amplitude point, the apex point of the first derivative, and the tangent intersection point) [127]. In addition, agreement between PRV and HRV is dependent on several physiological factors [123]. Agreement tends to be higher when measurements are performed at rest than during exercise [124], [128]. In the future it will be important to determine the minimum duration of PPG signal required to accurately assess PRV so that it can be assessed during periods of low activity during daily living [129].

Arterial stiffness
Some PPG-based devices now offer an assessment of arterial stiffness (see Biostrap Evo's Arterial Elasticity and Peripheral Elasticity metrics). Such measurements are based on analysis of the shape of the PPG pulse wave. Several pulse wave features have been identified as potentially indicative of arterial stiffness [102], [130], based on similar physiological mechanisms to those exploited when assessing BP from a single pulse wave. Gold standard measurements of arterial stiffness are associated with cardiovascular events and all-cause mortality [131], as are features of the PPG pulse wave [132]. However, it is not yet clear how accurately arterial properties can be assessed from wearable PPG signals, and whether such measurements are indicative of cardiovascular risk.

COMMERCIALLY AVAILABLE DEVICES
This Section provides an overview of the wide range of commercially available PPG-based wearables. Many wearables are used alongside smartphone apps, which enable data storage, visualisation and analysis. Readers interested in apps for PPG-based wearables are referred to [133].
The consumer market for PPG-based wearables has exploded in recent years. Smartwatches, wristbands, smart rings, and ear buds are all widely available. Although not all consumer wearables use photoplethysmography, those which monitor HR invariably use photoplethysmography to do so, such as those shown in Fig. 7. The functionality of selected PPG-based wearables is summarised in Table I. This table provides an overview of the field, demonstrating the range of form factors taken by wearables, and parameters estimated by them. The list is by no means comprehensive: devices are shown from selected wearable manufacturers, and only one device is listed per manufacturer. Furthermore, the functionality of devices can be enhanced with software updates, or through use with additional apps. The interested reader is referred to [4], [134] for details of additional consumer devices. Devices which are not available to consumers are beyond the scope of this review (see for instance the CE-marked SOMNOtouch NIBP device [135] and the FDA-approved Biobeat watch [136]).

Form factors
PPG-based wearables take a variety of form factors and can be attached to different parts of the body. As shown in Table I, the vast majority of PPG-based wearables are worn on the wrist, either as a watch (with a larger screen), a band (with a smaller screen), or a strap (without a screen). Other form factors have emerged more recently, such as finger rings, ear buds and ear wrap arounds, ankle socks, and arm sleeves. In addition, the Polar OH1 sensor is a disk-shaped sensor which provides the user with flexibility to mount it anywhere on the skin, and comes with an armband and swim goggle attachment. Potential users have expressed a preference for wrist-worn devices over other form factors [138]. Nonetheless, there are advantages to other measurement sites (see Section II-A), and use cases in which other form factors are advantageous. For instance: a sock may be more suitable for monitoring babies (see Owlet Smart Sock); ear buds may provide more reliable fall detection than devices attached at peripheral sites such as the wrist (see Bodytrak); and, an arm sleeve can incorporate an ECG electrode on the chest (see AIO Sleeve), which is not possible with many other form factors. Additional form factors have been used in devices sold to industry rather than consumers directly: the Equivital eq02+ LifeMonitor takes the form of a chest belt with shoulder straps, and the Spire Health Tag attaches to clothing. Other form factors have been investigated in research too, such as incorporating PPG sensors into arm bands and glasses [22], [31]. Whilst the wrist form factor dominates the market at the moment, the finger ring may become increasingly popular as it shares many of the same characteristics. Other form factors are likely to be more suitable for specific use cases.

Functionality
Wearables estimate several physiological parameters from the PPG signal. The PPG is widely used for HR monitoring, as shown by HR being provided for all of the devices in Table I. Some devices also capture the ECG either on demand or continuously, from which HR can be obtained. Since PPG sensors can provide frequent or continuous monitoring without the need for additional electrodes, it is likely that photoplethysmography will remain the dominant method for HR monitoring in consumer wearables. The remaining PPG-derived parameters are less widely calculated, and the technology for estimating these parameters is still in the development phase in many cases. Some devices use the PPG to identify irregular heart beats which may indicate arrhythmias. This approach has been shown to have utility in the Apple Heart Study [57], and is likely to become more widely used as shown by the Huawei Heart Study [139], FitBit Heart Study [140], and work on detecting atrial fibrillation by other manufacturers [141]. SpO 2 has recently become more widely measured, and is now offered by several wrist-worn devices. SpO 2 measurement has been facilitated by devices incorporating LEDs of multiple wavelengths (such as the Apple Watch Series 6 using red, green and infrared light). RR and BP are more difficult to estimate reliably from PPG signals, and are not yet widely offered. Approaches to overcome the challenges of monitoring these parameters include: monitoring RR at night when there is likely to be less motion artifact (see the Withings Smartscan's Respiratory Scan function) [8], [142]; and calibrating BP estimates with a traditional cuff measurement (see the Samsung Health Monitor App). In the future PPG-based wearables might routinely measure many or all of the mentioned parameters, in a similar way to the current widespread measurements of HR.
Wearables also estimate some parameters through analysis of both the PPG and other signals. Sleep can be identified and assessed using both activity levels from accelerometry signals, and HRs from PPG signals. Indeed, since most wearables incorporate an accelerometer, most of those listed in Table I assess sleep. Similarly, most of the wearables assessed calories burnt, which can be estimated from activity levels with (in some cases) the additional use of HRs.VO 2max can be estimated from HRs acquired during both rest and exercise [118], and several of the listed wearables do estimateVO 2max .
Wearables also provide further parameters derived from other signals. The recent addition of an ECG sensor to some wrist-worn devices provides opportunity to complement PPG-based assessments with ECG signals which could potentially be used for diagnostics. For instance, on detecting an irregular heart rhythm from a PPG signal, the user could be prompted to take a short ECG recording which could then be clinically reviewed. The use of accelerometry to identify exercise, and barometers to identify stair climbing [143], provides complementary information which can be used in conjunction with PPG-derived HRs to assess fitness. Finally, several wearables measure temperature which, if closely associated with body temperature, may be a helpful marker of illness [144]. When measuring temperature, the ear bud form factor may be advantageous as temperature measured in the ear may agree more closely with core temperature than measurements at other measurement sites.

Marketing models
Several different marketing models have been used to commercialise wearables. In most cases the consumer pays for the device upfront, and any necessary apps are provided by the manufacturer without charge (for instance, Withings' HealthMate app). An alternative approach is to use subscription charges, and provide the wearable to subscribers (see WHOOP membership). In addition, a hybrid approach can be used in which additional data analytics and visualisation functionality is enabled with a subscription (see Biostrap's Sleep Lab).

Batteries
The way in which the battery is incorporated into the device affects its usage. Most PPG-based wearables contain built-in batteries and so must be removed from the body for charging, resulting in periods of downtime. For instance, the Apple Watch is charged using a magnetic charging cable, the Oura Ring is charged using a dock, and the Fitbit Charge is charged using a cable with clip attachment. In contrast, the Equivital eq02+ LifeMonitor and WHOOP Strap have interchangeable battery packs, virtually eliminating downtime due to battery constraints.

APPLICATIONS
A wide range of applications have been proposed for PPG-based wearables. These range from clinical applications to fitness and lifestyle monitoring. The reader is referred to [7] for a discussion of several clinical applications. A few additional applications are now discussed.

Menstrual cycle monitoring
The ability to monitor the menstrual cycle and predict ovulation is valuable for family planning. The data acquired by wearables could be used for this purpose, particularly when measured whilst asleep. Temperature measured at the finger using an Oura Ring has been found to be useful for menstrual cycle monitoring [145], and temperature measured in the ear canal has been found to be useful for predicting ovulation [146]. In addition, PPG-derived parameters including HR, PRV, RR and skin perfusion have all been found to be useful for predicting the fertility window [147], [148]. Potentially, such parameters could be used in combination to track the menstrual cycle.

Identifying orthostatic hypotension
Orthostatic hypotension is a prolonged drop in BP upon standing of at least 20 mmHg systolic BP, or 10 mmHg diastolic BP [149]. It is caused by a dysfunction of the autonomic nervous system, which would normally compensate for the reduction in BP when standing up due to gravity causing blood to pool in the lower body. In older adults, orthostatic hypotension is associated with an increased risk of falls [150], which in turn result in increased morbidity, mortality and healthcare costs [151]. Methods to identify orthostatic hypotension in daily life could potentially help prevent falls. It has been proposed that by estimating BP from the PPG, one could identify drops in BP indicative of orthostatic hypotension [152]. Accelerometry and gyroscope measurements from wearables could also be used to identify when a user stands following a prolonged period of lying down or sitting, which could be combined with PPG-based detection of drops in BP to identify possible orthostatic hypotension for further assessment.

Seizure detection in epilepsy
Epilepsy is one of the most common neurological disorders, affecting almost 1% of the population worldwide [153]. Patients are often required to keep a seizure diary outside the hospital in order to follow-up the disease and evaluate treatment. Reliable seizure detection might help optimise antiepileptic treatment, which could in turn reduce the risk of sudden unexpected death in epilepsy. Manual seizure diaries are unfortunately not highly reliable, thus there is a potential role for automated, preferably wearable, seizure detection devices. Some types of epilepsy affect the autonomic nervous system (e.g. temporal lobe epilepsy), and in turn the cardiovascular system. Temporal lobe seizures are often accompanied with a strong increase in HR, providing opportunity to detect this type of epilepsy through PPG-based HR monitoring [154], [155]. Another study has investigated cardiorespiratory effects of epilepsy by monitoring SpO 2 and found that 63%-73% of generalized convulsions and 20%-28% of focal seizures can be detected by using SpO 2 thresholds of 80%-86% [156]. Relatively little research has been conducted to date on using PPG-based wearables to detect seizures, although a planned multicenter study holds promise for providing new insights into the utility of PPG-based wearables for this application [157].

Anaesthesia and pain monitoring
Monitoring of painful stimulation is routinely used to assess the adequacy of analgesic medication for pain control during anaesthesia [158]. Pain causes sympathetic responses in the autonomic nervous system, and these are associated with morphological changes in the PPG [159]. Consequently, PPG-based devices may have utility for monitoring analgesia. The PPG-derived Surgical Plethysmography Index (SPI), calculated from the pulse wave amplitude and duration, has been proposed as an approach to assess analgesic state [160]. The SPI was integrated into the Aisys® Carestation device (GE Healthcare, Finland). The SPI does not appear to be valid in children [161], which may be due to both differences in blood vessel distensibility and baseline HRs in children versus adults. The autonomic nervous system state (ANSS) metric was introduced in [162], and is calculated from the same features of the PPG pulse wave as for the SPI. Additional approaches have been proposed in [163], which assessed the performance of a plethora of PPG pulse wave indices to assess postoperative pain, and [164], which used deep learning to classify pain level. Further details of methods for analgesia monitoring are provided in [158], including methods based on HRV and wavelet cardiorespiratory coherence [165], which could be suitable for implementation in PPG-based devices.

Chronic kidney disease monitoring
Chronic kidney disease (CKD) has a high global prevalence of 11 to 13% [166], [167], and advanced CKD may require frequent haemodialysis to replace lost kidney function. Consequently, interventions to delay the progression of CKD and improve outcomes are valuable. Cardiovascular diseases (CVDs) such as coronary artery disease (CAD), congestive heart failure, arrhythmias, and sudden cardiac death are the main causes of morbidity and mortality in patients with CKD [168]. These CVD pathologies impact haemodynamics and so could potentially be detected and monitored by PPG-based devices. In [169], the authors demonstrated the potential of dual wavelength (infrared and green) reflectance photoplethysmography to identify CAD in CKD patients. Differences were observed in the upslopes and downslopes of PPG pulse waves between CKD patients with and without CAD. In the future, such features could be adjusted to account for changes which occur with age. Specifically, in haemodialysis patients, an arteriovenous (AV) fistula in the arm significantly increases the steepness in blood volume change during systole compared with the arm without AV fistula. PPG-based monitoring may also be useful for detecting arrhythmias such as bradycardia and tachycardia, which are common in haemodialysis patients [170]. PPG-based monitoring might also be useful for preventing intradialytic hypotension [171] and for hypotension management [172], [173]. A recent review [174] identified only a few studies investigating the use of PPG-based monitoring in haemodialysis patients; however, they concluded that wearable health devices will enter clinical practice in the near future, including for haemodialysis patients.

Biometric authentication
There is growing interest in new approaches to user authentication. PPG-based wearables provide opportunity to authenticate users based on analysis of their PPG signal, since PPG signals differ from one user to the next. Broadly, two approaches have been taken to identify an individual from their PPG signal: classification algorithms based on either pulse wave features [175]- [177] or deep learning frameworks [178]- [180]. The PPG can be transformed into the angle domain before feature extraction to reduce intra-subject variability [181], [182]. Time-frequency analysis has also been used to extract features without the need to identify individual pulse waves [183]. The main challenge is to reliably classify users in the presence of intra-individual variability. Consequently, it is helpful to train a system using data from one recording session, and assess its performance on data from the same individuals in a separate recording session [180]. Performance can also be assessed during situations which affect pulse wave morphology and variability, such as changes in emotions [181] and activities of daily living [176], [182]. PPG-based authentication could be used to provide access to confidential and sensitive information over the internet [180], or could act as a pre-cursor to existing authentication systems such as fingerprint scanning [179].

Health insurance
Some insurance providers are now offering rewards or discounts to policy holders who use a PPG-based wearable to track their exercise [184], [185]. In doing so, insurance providers can promote healthier lifestyles [186], personalise insurance premiums according to individual risk, and create personalised products based on users' data [187]. Indeed, many of the largest insurance companies in the USA offer self-tracking health and life insurance schemes [188]. However, concerns have been raised that the use of wearables in health insurance is not in keeping with the principles of solidarity, fairness and equality on which collective insurance schemes are based [188].

CONCLUSION AND FUTURE WORK
The growing use of PPG-based wearables provides opportunity to monitor health and fitness in daily life. The potential utility of wearables is directly linked to their design, with several form factors used in commercial devices. Several physiological parameters can be estimated from the PPG signal, which if incorporated in wearable systems, could provide a wealth of information on the cardiovascular, respiratory and autonomic nervous systems. Additional parameters and features have been incorporated into commercially available wearables over time, expanding their functionality and potential utility. Wearables may now, or in the future, have utility in a wide range of applications in clinical, health and fitness settings. However, much further work is required to realise the full potential of PPG-based wearables.

Areas for future work
Key areas for future work concerning PPG-based wearables include: 1) Identifying use cases for wearables in clinical practice, and assessing their cost-effectiveness for these use cases [189]. The detection of possible atrial fibrillation using PPG-based wearables is a promising initial use case, since the potential utility of devices for this purpose has been demonstrated [57], and there is an established treatment strategy (anticoagulation to reduce risk of stroke). 2) Assessing whether the use of wearables for behaviour change, such as prompting increased physical activity, is associated with improved health and wellbeing outcomes. 3) Harnessing the wealth of data provided by wearables for large-scale research [190], as achieved in the Apple Heart Study [57], Huawei Heart Study [139], and studies using data from Fitbit devices [191], [192]. 4) Thorough evaluation of the limitations of wearables to ensure they are not used outside of their capabilities [193]. 5) Ensuring that any clinical benefits of wearables are made accessible to all, noting the vast geographical discrepancies in the current use of wearables [1].