Establishing best practices in photoplethysmography signal acquisition and processing

Abstract Photoplethysmography is now widely utilised by clinical devices such as pulse oximeters, and wearable devices such as smartwatches. It holds great promise for health monitoring in daily life. This editorial considers whether it would be possible and beneficial to establish best practices for photoplethysmography signal acquisition and processing. It reports progress made towards this, balanced with the challenges of working with a diverse range of photoplethysmography device designs and intended applications, each of which could benefit from different approaches to signal acquisition and processing. It concludes that there are several potential benefits to establishing best practices. However, it is not yet clear whether it is possible to establish best practices which hold across the range of photoplethysmography device designs and applications.

Potential areas in which best practices could be established include factors relating to device design (hardware and software) and measurement protocols (recording setting and duration). These are summarised in figure 1, and now described.
Several factors in the hardware design influence the PPG signal (Charlton and Marozas, Lemay et al), and are therefore potential areas in which best practices could be established. Firstly, the wavelength of emitted light determines the depth of light penetration, and consequently the level of the vasculature contributing to the PPG signal (Liu et al 2019), which influences signal quality (Fallow et al 2013). Current best practice is to use longer wavelengths (e.g. infrared) for transmission photoplethysmography as these penetrate deeper (Anderson and Parrish 1981), and shorter wavelengths (e.g. green) for reflectance photoplethysmography as these produce higher signal quality for heart rate measurement (Matsumura et al 2020). However, this practice may need to be revisited as the use of green light has been found to result in less accurate heart rate monitoring in subjects with darker skin tones (Fine et al). Secondly, in reflectance photoplethysmography the signal quality is influenced by the geometry of the light emitter, light detector, and sensor casing. Current best practice is to design the surrounding casing to eliminate ambient light as far as possible (Abay and Kyriacou). In the future this may be extended to using geometries in which the LED surrounds the photo detector, as these have been found to give higher signal quality (Khan et al 2019). Thirdly, the contact pressure applied by the device to the skin impacts the shape of the PPG pulse wave (Chandrasekhar et al 2020) and consequently its second derivative (Grabovskis et al 2013). Best practice in the area of contact pressure has not yet been established: higher pressures may reduce probe-tissue movement artifact, and have been found to increase the accuracy of PPG-based heart rate monitoring (Scardulla et al 2020). However, it is not clear whether such pressures would be suitable for longterm monitoring. Ideally, the contact pressure should remain constant when analysing pulse wave shape, such as when tracking changes nocturnal changes in blood pressure (Radha et al 2019). Fourthly, the body site chosen for PPG measurement influences pulse wave shape (Hartmann et al 2019), and the utility of the acquired signal (Charlton and Marozas). Best practice has not yet been established in this area: in clinical devices the finger is often used (Alty et al 2007), whereas in consumer devices the wrist is often used due to user preference (Prinable et al 2017). In summary, the challenge of establishing best practices is not trivial, as several factors can influence the PPG signal, and it is likely that different device configurations would be best suited to different applications.
The software used in PPG devices influences the PPG signal and the parameters derived from it, and therefore presents potential areas in which to establishing best practices. Firstly, there is a compromise between increasing the sampling frequency to capture details of the shape of PPG pulse waves, and reducing it to reduce power consumption (Lee et al 2018). Best practices differ between applications, with minimum acceptable sampling frequencies of 10, 16, and 25 Hz reported for heart rate, respiratory rate, and pulse rate variability measurements respectively (Wolling and Van Laerhoven, Charlton et al 2017, Choi and Shin 2017). Secondly, different approaches can be used to remove motion artifact, ranging from eliminating periods of motion (Guo et al 2021), to denoising the PPG (Zhang et al 2015), to cancelling motion artifact using a reference accelerometer or gyroscope signal (Marozas and Charlton 2021). Here, best practices also differ between applications: in hospital monitoring it has been proposed that periods of motion should be eliminated from analyses (Orphanidou et al 2015), whereas in exercise monitoring the alternative approaches of denoising the PPG or cancelling motion artifact are used (Zhang et al 2015). Whilst it may be challenging to develop a universal strategy to PPG signal quality assessment, recent work has demonstrated that a single approach can perform well across different heart rhythms and different PPG devices (Mohagheghian et al). Thirdly, the analog and digital filtering used to pre-process signals influences both the amplitudes and timings of PPG pulse wave features (Liang et al 2018, Liu et al 2021. For instance, an optimal low-pass filter cut-off of 6 Hz has been proposed to preserve the higher harmonic components of the PPG, and minimise variability in indices calculated from its second derivative (Pilt et al 2013). Fourthly, the choice of signal processing algorithm used to estimate a physiological parameter from the signal can greatly influence the accuracy and precision of the parameter (Charlton et al 2016). Best practices for deriving pulse wave features from finger PPG signals have been proposed (Elgendi 2014. However, best practices have not yet been established for signals acquired at the wrist, which differ from finger signals (Rajala et al 2018). Similarly, it could be beneficial to optimise neural network architectures for PPG analyses, building on existing architectures (Li et al 2021). Further work is also required to identify the best pulse wave features for different tasks from amongst the wide range of features proposed in the literature (Charlton et al 2018, Lin et al 2020. For instance, recent studies have investigated the best features for blood pressure estimation (Xing et al 2021) and pulse rate variability analysis (Peralta et al 2019). The best algorithm design may also depend on a subject's characteristics, as shown by recent proposals of different blood pressure estimation algorithms for subjects of different ages (Xing et al 2020) and subjects of different blood pressure categories (Khalid et al 2020). In summary, it may be difficult to establish best practices for the software used in PPG devices, as the best approach may vary according to the sensor configuration, application, and subjects being monitored.
A further area in which best practices could be established is the protocols used to obtain PPG measurements, where best practices could be used to obtain repeatable and reproducible measurements. Measurement protocols can be tightly controlled in clinical settings, where consideration can be given to room temperature, subject position, and the duration of rest prior to measurement (Allen and Hedley 2019). However, protocols cannot be so tightly controlled when obtaining measurements from consumer devices in daily life. Nevertheless, measurements can be obtained in a repeatable manner during periods of rest, such as resting and night-time resting heart rates (Mishra et al 2020, Radin et al 2020). Future work may consider the required recording durations and acceptable levels of signal quality to estimate different physiological parameters from the PPG (Huthart et al 2020). Whilst it is possible to obtain some parameters during exercise (e.g. heart rate) (Zhang et al 2015), it may only be possible to obtain other parameters accurately whilst at rest (e.g. those derived from the second derivative of the PPG, such as the aging index) (Takazawa et al 1998).
It is clear that there are several potential areas in which best practices could be established for the acquisition and processing of PPG signals. However, it is not yet clear whether it would be possible and beneficial to establish best practices. On the one hand: it may not be possible to establish best practices as they may vary greatly between device designs and applications; it may not be possible to use them widely if they are patented; and, they may not be beneficial if they don't substantially improve device performance. On the other hand, establishing best practices could: reduce the time taken to design and manufacture devices; ensure PPG-based measurements are as accurate and reproducible as possible; and, help advance the field as researchers and developers could build on existing best practices when making novel developments.
Several advances could aid research into determining whether it would be possible and beneficial to establish best practices for PPG signal acquisition and processing. Firstly, wearable devices which provide the raw PPG signal are invaluable for such research, as demonstrated through the use of the Empatica E4 wristband in many research studies (McCarthy et al). Whilst several research devices can provide the raw PPG signal , large-scale studies could be conducted more easily in daily life if consumer devices were similarly able to provide raw PPG signals. Secondly, freely available datasets allow researchers to benchmark their own PPG signal processing algorithms against others on a common dataset. Several such datasets are available , including: the WeSAD and PPG-DaLiA datasets, acquired using an Empatica E4 device in healthy subjects (Schmidt et al, Reiss et al 2019); and the VitalDB and the MIMIC Waveform databases, acquired from critically-ill patients (Johnson et al 2016, Lee and Jung 2018). However, there are limitations to current datasets: they are often collected from either healthy volunteers or a particular patient population, rather than a broad cross-section of society; they often contain PPG signals acquired by only one device, rather than signals acquired using different hardware configurations; and they are often recorded in either laboratory or clinical settings, but few are recorded in daily life. Thirdly, there is a need for widely accepted validation protocols with which to assess the performance of PPG-based devices. Such protocols already exist for devices measuring blood pressure and heart rate (Stergiou et al 2018, Mühlen et al 2021). However, different standards may be required for different applications, such as varying the accuracy and data availability thresholds according to the intended use case and measurement scenario (Consumer Technology Association 2018, Mukkamala et al 2021).
To conclude, there are several potential benefits to establishing best practices for acquiring and processing PPG signals. However, it is not yet clear whether it is possible to establish best practices which hold across the range of PPG device designs and applications. Therefore, much further work is required to investigate whether it would be possible and beneficial to establish best practices, and to understand how they may differ between device designs and intended applications.

Acknowledgments
This paper is based upon work from COST ACTION CA18216 'Network for Research in Vascular Ageing', supported by COST (European Cooperation in Science and Technology): www.cost.eu . The work was supported by British Heart Foundation (BHF) Grant FS/20/20/34626, and by the Estonian Ministry of Education and Research under personal post-doctoral research funding PUTJD815. Certain images in this publication have been obtained by the authors from the Wikimedia website, where they were made available under a Creative Commons licence. Please see individual figure captions in this publication for details. To the extent that the law allows, IOP Publishing and Institute of Physics and Engineering in Medicine disclaim any liability that any person may suffer as a result of accessing, using or forwarding the image(s). Any reuse rights should be checked and permission should be sought if necessary from Wikimedia and/or the copyright owner (as appropriate) before using or forwarding the image(s).