Feasibility of Remote Blood Pressure Estimation via Narrow-band Multi-wavelength Pulse Transit Time

2.1 MW Monitoring of Physiological Signals

A review study on MW wearable PPG was conducted recently by Ray et al. [23] in which they highlight the importance of a variety of factors for MW skin monitoring, such as melanin (skin tone) and skin temperature. They show an extensive overview of wavelengths and hardware used in these studies, showing that in the majority of cases the blue, green, and infrared (IR) part of the light spectrum is used, as these wavelengths have different penetration depths in accordance with relevant vascular content at different depths of skin. Their study was broad in terms of monitored vitals, however, for BP estimation they highlight the work of Liu et al., who proposed single-site BP estimation using a custom MW contact sensing probe [19]. They showed a decent correlation between SBP and PTT using blue and red PPG trace. They later expanded their research [18] showing that MW PTT is correlated with SVR – a metric correlated to (traditional) PTT – which can in turn be used for continuous BP estimation. In a study using 20 subjects they achieved low errors of around 3 mmHg for SBP using a four-wavelength (470, 570, 590, 940 nm) sequential illumination sensor, which is not practical in a remote setting (fast-flickering lights). These errors were achieved on per-subject basis using a formula connecting PTT and BP that requires calibration values.

An early contact-free proposal of MW imaging for HR and SpO2 monitoring was given by Wieringa et al. [30]. They built a custom recording setup with a CMOS camera and light sources emitting wavelengths of 660, 810, and 940 nm. They acquired respiratory, HR and SpO2 information from PPGs at multiple wavelengths, which correlated well with ground truth. Another study in remote MW PPG analysis was reported by Moço et al. [20] in which they highlight skin inhomogeneity as an important factor in phase differences of PPGs from different wavelengths. They suggest that the observed temporal delays reflect morphological heterogeneity of the PPG waveform due to tissue-specific factors to a larger extent than actual propagation-related properties, which can be a major challenge for PTT applications. They initially investigated only the spectrum of visible light, focusing on green and red. However, in later work they highlight NIR wavelength as being a good choice for SpO2 measurement, as it reaches the deeper pulsating arterioles instead of only the capillary bed. They express concerns about the latter exhibiting pulsatility as a consequence of elastic deformations and subsequent capillary density changes in the monitored region, rather than only blood volume changes, which they support by showing that skin compression can importantly influence the PPG morphology (e.g., when using a fingertip contact sensor). They also offer detailed models for skin layers microvasculature [21]. Their work offers important fundamentals in terms of understanding and modeling the non-trivial blood perfusion in the complex skin microvasculature.

2.2 Remote BP Estimation

There is a myriad of recent work dealing with BP estimation from rPPG. Some of it relies heavily on morphological properties of the waveform or several waveforms obtained from smaller sub-regions of the chosen measurement site, such as the forehead. Iuchi et al. [14] recently proposed one such method in which they used RGB facial videos that were converted to what they call spatial pulse wave signal. This is in essence a sequence of frames divided into small sub-regions where each sub-region shows the intensity of hemoglobin in the skin in the form of a heatmap. This sequence of frames is then fed into a branched 2D CNN to estimate SBP and DBP with mean absolute errors (MAE) as low as 6.7 mmHg for SBP and 5.4 mmHg for DBP. Importantly they also showed good correlation coefficients between prediction and ground truth – up to 0.81 for SBP and 0.84 for DBP – indicating that BP variations (even though quite limited in the dataset) were successfully modeled. In their experiment they also used personalization, as the data was split into training and testing using the 80-20 split, where data of same individual is present in both sets. Some other recent approaches are complete black-box end-to-end neural network implementations, taking a sequence of images as input and predicting BP as output, without manual data processing intermediate steps. Botina et al. [5] proposed a truly end-to-end 3D CNN architecture capable of real-time rPPG reconstructions and subsequent physiological parameter estimation, in which they reduce the dimensionality of the input frames to achieve up to 88% increase in speed. They still maintained comparable accuracy in terms of correlation coefficients to other state-of-the-art architectures. Despite good results, the challenges mentioned previously remain – personalization and calibration requirements [24], good skin exposure with enough skin pixels, lighting conditions, and so on.

2.3 Synthesis and Findings

Looking at related work as well as review articles of contact-free physiological parameter estimation, we came to the following findings:

Our proposed solution aims at answering these by using a contact-free sensor in the form of a modified triple-band-pass high-frequency camera and a channel separation method preceding conventional data processing. This type of sensor modification allows us to leverage information from deeper skin layer via NIR part of the spectrum. In contrast to existing single-site approaches, which use contact PPG sensors, our proposed setup and method do not require skin contact, thus not compressing nor distorting the tissue and the waveform, meaning the precise location of systolic peaks is completely preserved. This is important for computation of short PTTs between skin layers. Usage of narrow triple-band-pass filter in combination with algorithmic channel separation resolves the inter-channel influence and overlap, allowing for robust PTT observations. Additionally, contact PPG sensors are always dependant on either battery life (consumer wearables) or cables (most often clinical devices), making them more obtrusive and not suitable in all circumstances. In different contexts, different recording setups are preferred, as each has its pros and cons. While a traditional contact approach circumvents many technical challenges such as light source and image sensor specifics, the proposed contact-free approach would be more desired in cases where wearables are not feasible due to some subject specifics such as skin conditions, allergies, preterm infants, and so on. Personalization or person-specific calibration remains a fundamental challenge, which we acknowledge and work around by using a small amount of each subjects data in the training of regression models. We describe the proposed methods in detail, followed by our experimental results, in Sections 4–6.

3.1 Hemodynamics and Skin Physiology

The underlying idea behind PTT is to measure the time it takes for a specific heart pulse wave to traverse a certain distance. This idea was historically most commonly implemented using pulse arrival time (PAT), where an electrocardiogram (ECG) and a distal PPG (usually at the fingertip) are synchronized and then the R peak in the QRS complex of the ECG and the corresponding systolic peak in the PPG are used as the reference points between which the PAT is measured [8]. When the temporal delay is measured using two PPG signals, it is referred to as PTT [22]. This delay is shorter if vascular stiffness and BP are higher, and longer when vascular stiffness and BP are lower.

While the relationship between PTT and BP seems simple and intuitive, the hemodynamics of humans are neither trivial nor general. There are two crucial variables determining systemic BP, the cardiac output (CO, the amount of blood ejected by the heart into aorta in some time, typically one minute) and the systemic vascular resistance (SVR, the fluidic resistance of blood vessels opposed to the flow of blood), as given in 1 [13]. (1) \(\begin{equation} BP = CO \cdot SVR, \end{equation}\) To introduce PTT into the dependancy via SVR, we rely on the Moens–Korteweg Equation (2): (2) \(\begin{equation} PWV = \frac{L}{PTT} = \sqrt {\frac{E_{inc} \cdot h}{2r\rho }}, \end{equation}\) where PWV is pulse wave velocity, which is the velocity at which a pulse wave propagates a certain distance L, \(E_{inc}\) is the incremental elastic modulus of the vessel wall, h is the vascular wall thickness, r is the vessel radius, and \(\rho\) is the blood density, assuming the arterial wall is isotropic, meaning uniform in all directions. SVR is related to the elastic modulus on the right-hand side of the equation [13]. PWV is complex and varies drastically as the pulse traverses the cardiovascular system due to different types of vessels in the body. Additionally, especially in thinner vessels, wave reflection makes the PWV difficult to estimate. Literature reports different values spanning 2-3 orders of magnitude. Precise measurements in capillaries, which are the dominant vessels between different skin layers of the palm, show expected PWV in the range of 6.4–17.6 mm/s [4]. Assuming average palm skin thickness of 1 mm, we can expect PTTs of around 156–57 ms. Importantly, this depends on specific elastic properties of the arterial/capillary wall and can vary between subjects.

Modern PTT approaches simplify the traditional setup by commonly replacing the ECG with a PPG or a camera sensor, as these sensors are cheaper, simpler to use, less obtrusive and more accessible [15]. Despite this, the multi-site nature of measurement remains, and comes with its set of challenges, such as the noise on each sensor and precise synchronization required. MW approach circumvents these by using a single sensor at a single measuring site. Instead of measuring the delay of the pulse reference points between two different locations, such an approach is instead based on the idea that different wavelengths of light penetrate to different skin depths, reflecting blood perfusion as it arrives from deeper towards the upper skin layers [18, 19]. Related work established that a direct light source with wavelength 550 nm (green) can reach depths between 0.1 and 1 mm (mostly dermis layer), while wavelength 850 nm (infrared) penetrates beyond 2 mm deep (hypodermis layer), depending on skin tissue specifics [7]. The obtained waveforms from different skin depths are expected to be slightly delayed due to blood traversal. The fundamental difference between approaches is shown in Figure 1.

Fig. 1.

View Figure

Figure 1 also shows a layered structure of the skin microvasculature consisting of three layers: the topmost epidermis, which is mostly bloodless and consists of dead or dehydrated cells; the middle dermis containing capillary loops and small arterioles; followed by the deepest hypodermis containing subcutaneous fat and larger vessels. We are mainly interested in the dermis and hypodermis, as they contain a majority of the vasculature that exhibits blood flow.

3.2 Camera Physics

Obtaining different wavelength information is an inherent property of RGB cameras, as each color pixel is sensitive to specific wavelengths. This is vital for obtaining delayed PPG waveforms from different skin depths. However, the image sensor of the camera is not perfect but is instead governed by the quantum efficiency of each wavelength pixels, which shows the relative amount of specific wavelength photons successfully registered by the image sensor pixels and translated to a digital signal. An example is shown in Figure 2, which belongs to our specific image sensor, but also serves well for illustration purposes. The relative intensity of sensor response (photons registered and translated to digital signal) on the y axis is in the range [0, 1], where 0 means no photons are registered and 1 means all of them are. The x axis simply shows different wavelengths of light in the visible and near infrared (NIR) spectrum.

Fig. 2.

View Figure

Each of the colored lines in Figure 2 represents the relative response of the corresponding specific color pixels of the image sensor (red, green, and blue) at different wavelengths. We can discern two important facts and corresponding challenges:

These challenges are fundamental to out-of-the-box cameras and vast majority of image sensors, and must therefore be understood and considered when dealing with MW approaches. We propose the following mathematical description of each color channel obtained from our camera equipped with a triple-bandpass filter, producing the quantum efficiency signature shown in Figure 2: (3) \(\begin{equation} \begin{aligned}R &= r \\ G &= g + x \cdot b + r \\ B &= b + y \cdot g + r \end{aligned}, \end{equation}\) where R, \(G,\) and B are the red, green and blue values produced by the camera, while r, g and b are the corresponding channel-separated red, green and blue values. The parameters x and y are describing the ratio of the corresponding color present in mixture that must be determined.

To elaborate further, let us for example consider the G component from Figure 2, looking only at the areas that the triple-bandpass filter passes through (under the black curve) – it comprises of the actual green response g in the band around 550 nm, but we can also observe non-trivial response of the green line in the blue band around 450 nm, as well as in the NIR band around 850 nm. The latter is (almost) equal to the \(R = r\) response, while the former can be expressed as some fraction of the (channel-seperated independent) blue response in the form of \(x \cdot b\). Some simplifications are made here, specifically, we truncated the relative response under 0.1 to 0 (since the influence is very small) and we equalized all three responses in the NIR band, since the sensor responses are almost the same.

4.1 Camera

As discussed in Section 3, skin layers are thin at 1–2 mm and blood traversal velocity in capillaries expected around 12 mm/s, so the PTT measured between different skin layers is expected to be short at around 100 ms. Knowing this, we required a camera capable of high-frequency recording. Additionally looking at related work dealing with contact MW PTT measurement [18, 19], we discerned that our camera should be capable of recording at around 250 fps. This allows measuring PTT with precision of up to 4 ms, which is more than an order of magnitude shorter than expected PTT, so it should be able to capture PTT variations related to BP. Additionally, we wanted no on-board image processing since any changes to the pixel values might influence and distort the obtained rPPG waveforms. Following these requirements we opted to use the iDS 3040SE-Q RGB camera with the Sony IMX273 1/3” CMOS image sensor and iDS-5M23-C1618 16 mm lens for our experiments. This camera allows for variable frame rates of up to hundreds of fps while also offering programmatic access to the raw images as registered by the sensor. For our experiments we chose 250 fps due to aforementioned reasons.

As seen in Figure 1 and shown by Moço et al. [21], capillary loops dominate the papillary dermis layers reachable by shorther wavelengths, while smaller arterioles and arteries are found in the deeper layers of the skin, so we wanted to use light that also reaches those depths. NIR in the wavelengths around 850 nm seems suitable for penetrating to deeper dermis layers, however, it is not registered by the image sensor of a traditional RGB camera out of the box, since they come equipped with a default IR filter with a cutoff at 650 nm. We thus modified our camera by first removing the on-board IR filter and then replacing it with a triple band-pass MidOpt TB475/550/850 filter, which allows only light in narrow bands of 475 \(\pm\) 10 nm (blue), 550 \(\pm\) 10 nm (green) and 850 \(\pm\) 22 nm (NIR) to pass. The band is wider in the NIR due to lower sensitivity of the image sensor in the NIR range, as seen in 2. Furthermore, human skin often exhibits sweating, especially when exposed to heat or after physical activity, which can result in specular reflections on the surface. These might distort our signal reconstructions, so we additionally used the MidOpt PR1000 VIS/SWIR Wire Grid Linear Polarizer which is effective in the range of 400–2000 nm.

The combination of these modifications gives us access to NIR information from deeper arterioles in the reticular dermis while also partially addressing the first challenge mentioned in Section 3 – channel separation. As the selected filter has narrow bands, it isolates the response of the pixels producing a “cleaner” image in terms of specific wavelength fidelity, albeit this does not solve the problem entirely, as explained in Section 3.

4.2 Light Source

To further address the second challenge listed in Section 3 – poor quantum efficiency in NIR part of the spectrum – we had to use a light source that emits the full spectrum that we are interested in, as well as (ideally) increases in energy towards the NIR. Daylight has a suitable spectrum, however, we wanted good control and consistency in our light source, so we used an artificial source instead. The spectrum requirement meant that conventional LED sources are not suitable, since they have distinct peaks in the blue and green part of the spectrum while emitting virtually no energy in the NIR. An obvious initial candidate was a traditional incandescent filament bulb, which has a suitable spectrum for our use case. Upon making initial recordings with a standard 230V 50W bulb and our high-frequency camera, we observed flickering in the recordings, which is a consequence of the 50 Hz alternating current. To circumvent this we instead used a direct current power source – a 12V 650W power supply unit – alongside Osram Decostar 51ALU filament bulbs. These bulbs are coated with thin aluminum coating and unlike their dichroic counterparts, the heat is fully emitted in front, meaning we get a good amount of (N)IR emission, compensating for the lower quantum efficiency in this part of the spectrum. We decided to use two 50W bulbs to ensure enough light to get reasonable visibility, as the exposure time is very short at such high fps. The light was directed towards the region of interest (ROI) of the skin and the bulbs were placed perpendicular to one another in order to negate any shadows created by the uneven surface of the recording ROI.

4.3 Blood Pressure Monitor

Ground-truth BP measurements were obtained with a clinical-grade Omron M10-IT cuff-based digital BP monitoring device in order to obtain precise and trustworthy measurements. The cuff was placed on the upper arm in accordance with the official guidelines. The whole setup with an anonymous example subject is shown in Figure 3.

Fig. 3.

View Figure

4.4 Data Collection Protocol

The described recording setup was used to record the central part of the palm skin of test subjects, as seen in the example frame shown in Figure 3. The palm was chosen for measurement since it is one of the places where skin exhibits the most pulsatility (similar to fingertip and forehead, which are the most commonly used sites) and is relatively comfortable to record. The spectral range between blue (450 nm) and NIR (850 nm) was selected as such light penetrates between 0.5 and 4 mm deep in human skin tissue, which is suitable accounting for variations in skin thickness due to calluses or other anomalies [3].

We obtained palm recordings from 9 subjects alongside their ground-truth BP measurements. Written consent for recording and data processing was obtained from all subjects. Seven were male and two were female, the total mean age was 32 \(\pm\) 3.4 years. Each subject was asked to report any potential existing health pre-conditions and 7 of them reported no existing pathologies or diseases, while also having normal body mass index. There were two outliers in terms of health status, since they were years-long cigarette smokers, while also being above the average age.

Since we were interested in very precise waveforms, and some related work shows potential influence of cuff compression on the waveform itself via blood vessel compression [20, 21], we recorded videos of the left palm while placing the cuff on the right upper arm to minimize the influence. Each recording lasted for 30 s, during which a single ground-truth measurement was taken, since the measurement itself takes more than 20 s. Thus a single SBP, DBP, and HR ground truth is associated with the whole 30-sec recording, which is feasible since BP does not change a lot in 30 s outside of extreme circumstances (e.g., arterial bleeding or strong medication). Each subject was recorded several times on different days, following 2 scenarios:

As mentioned, we repeated this recording sequence on different days in order to avoid some daily specifics of individuals (e.g., had a stressful meeting coming up), which might not reflect their typical hemodynamic state. In total, we obtained 99 recordings from 9 subjects and first checked the SBP and DBP distributions using a Shapiro–Wilk test, which returned \(p \gt 0.05\), for both cases, confirming it to be a normal distribution. SBP and DBP were on average increased by 22.3 and 11.0 mmHg respectively between rest and activity. Since we also have dependencies between measurements (a sort of before/after scenario), as we assume that subjects with higher BP during rest will also have higher BP during activity, we used the paired T-test to check for statistically significant differences between resting and post-activity scenarios. We used the standard significance level of 0.05, and obtained p-values of 0.01 and 0.001 for DBP and SBP respectively, successfully rejecting the null hypothesis that our rest and activity BP samples come from normal distributions with the same mean, thus confirming that the two scenarios are different in a statistically significant manner.

5.1 Frame Acquisition Software

Once our recording setup was finalized at the physical level, we had to ensure that we removed any on-board default image processing. We did this by developing a piece of recording software based on the official iDS SDK together with their Peak library. In short, we first set the fps to 250, which means that we had to make a tradeoff by setting the exposure time to only 4 ms. We then turned all the white-balance corrections off and set the color gains to the default value of 1.0. At this point, we obtained raw 3D matrices corresponding to each frame, which we then manually Debayered as given in Equation (4): (4) \(\begin{equation} R, G, B\ =\ {\left\lbrace \begin{array}{ll}R\\ round(\frac{G_1}{2} + \frac{G_2}{2})\\ B \end{array}\right.}\!\!, \end{equation}\) where R are the (nearinfra)red pixels, \(G_1\) and \(G_2\) are the green pixels and B are the blue pixels in a standard half green, one quarter red and one quarter blue Bayer mosaic pattern. Doing this lowered the resolution of the final frames (due to halving of the green pixels) to 540 \(\times\) 720 pixels.

5.2 Signal Processing

The sequence of frames was first globally spatially averaged (taking all pixels, since only skin is present in each frame) in order to obtain three temporal RGB traces. These traces were then zero-phase filtered with a 2nd order Buttherworth band-pass filter with cutoff frequencies of [0.5, 6.0] Hz. This removed low-frequency baseline wandering and eliminated high-frequency noise. Finally, we normalized our signals amplitude to a constant range of [\(-\)1, 1]. At this point we obtained relatively clean traces with obvious cardiac pulsatility, as seen in Figure 3. Methods used in this processing pipeline were universally applied to all signals, as they were carefully designed to not influence the waveforms in any way that would distort the relationship under investigation. More specifically, both the amplitude normalization and the zero-phase filtering were chosen specifically to be generally applicable and to not influence the temporal position of the systolic peak or the steepest systolic rise point, regardless of the subject.

In order to measure PTT, we opted to use a reference-point method, meaning we used a single robustly detectable stable point per cardiac cycle, between which we measure PTT in different waveforms. Specifically, we chose the steepest point on the systolic slope, as this is often more robust and stable compared to more commonly used systolic peaks. The systolic rise is more apparent even in noisy signals, while systolic peaks are sometimes difficult to precisely determine even for state-of-the-art peak detection algorithms [11]. Stability is important in this case because even a slight temporal missdetection of a few samples can represent a large part of the very short PTT. We detected these reference points by using the derivative method, taking the more clearly-expressed peaks in the first derivative. As precision and stability of these reference points is vital, we cross-checked our detections with state-of-the-art systolic rise/peak detection algorithms proposed by Elgendi et al. and Han et al. [6, 9], which were specifically developed to tackle challenging conditions and cardiac anomalies respectfully. Since we ensured our signals were as clean as possible, by both stabilizing the measurement site, as well as using aforementioned algorithmic cleaning, all algorithms detected the same locations of reference points. Detected example reference points can again be seen in Figure 3, represented by the star markers.

At this point of the pipeline, we can make an initial observation of the PTTs by looking at the steepest points locations in each channel, and then computing the difference between these. To do this, we first took the green channel and its corresponding steepest point detections as the reference, because green has been both historically and empirically (from our observations) identified – and also widely accepted in the field – as the trace with the most expressed and stable pulsatility. We then checked a very narrow area before and after these reference points by defining a short threshold of 40 samples and searching for steepest point detections in the other two traces (red and blue). If such points were found, we computed the temporal difference (PTT), otherwise we ignored the cycle.

Looking at early results, we had two initial important observations. First, the waveforms were relatively clean and stable in NIR and green, while the blue often expressed very shallow and noisy pulsatility. Second, the PTTs were extremely short, lasting just a sample or two. Given the former, while also knowing that blue wavelengths have the shallowest penetration mostly reaching the upper less blood-perfused skin layers, and concerns expressed in related work about lower wavelengths [21], we decided to simply remove the blue channel from further analysis, since it was too unstable. Regarding the unobservable PTTs, this was not entirely unexpected, since we already highlighted the problematic inter-channel influence in previous sections, which was only partially resolved using the narrow-band triple band-pass filter.

5.3 Channel Separation and PTT Analysis

To further reduce the inter-channel influence, we first had to determine x and y from (3), which we computed from our image sensor response as given in Equation (5): (5) \(\begin{equation} \begin{aligned}x &= \frac{S_{g_1}}{S_{b_1}} \hspace{10.0pt} in \hspace{10.0pt} [450, 500] \hspace{1.99997pt}~nm \\ y &= \frac{S_{b_2}}{S_{g_2}} \hspace{10.0pt} in \hspace{10.0pt} [520, 570] \hspace{1.99997pt}~nm, \end{aligned} \end{equation}\) where \(S_{g_1}\) and \(S_{b_1}\) are the areas under the curve in the blue band of the filter, while \(S_{b_2}\) and \(S_{g_2}\) are the areas under the curve in the green band of the filter. This essentially tells us the amount of one color present in the other one, where the unwanted “incorrect” color is the numerator and the dominant “true” color is the denominator as shown in Figure 2. For our specific camera and image sensor we obtained \(x = 0.61\) and \(y = 0.13\). Once x and y are obtained via definite integration of the curves in Figure 2 bound by the filter bands, we can follow (3) to express channel-separated r, g and b traces as given in Equation (6): (6) \(\begin{equation} \begin{aligned}r &= R \\ g &= G - R - x \cdot b \\ b &= B - R - y \cdot g. \end{aligned} \end{equation}\)

With some expression manipulation we can express the channel-separated traces as follows. First we express b by inserting g: (7) \(\begin{equation} \begin{aligned}b &= B - R - y \cdot (G - R - x \cdot b) \end{aligned} . \end{equation}\)

Inserting this back into g gives us (8) \(\begin{equation} \begin{aligned}g &= G - R - x \cdot b \\ \end{aligned} , \end{equation}\) where all the variables are known and can be obtained from the system. The factor \(\frac{1}{1 - x \cdot y}\) is a camera-specific constant that only influences the amplitude, which was normalized, so we can omit it, giving us the final forms as (9) \(\begin{equation} \begin{aligned}r &= R \\ g &= G - R - x \cdot (B - R) \\ b &= B - R - y \cdot (G - R). \end{aligned} \end{equation}\)

The effects of this channel separation method are summarized in Table 1. We can observe that they are positive and align with our expectations, since the PTTs become more obvious and mostly appear in the expected order [10].

5.4 BP Correlation Analysis

We analyzed the correlations between the measured ground-truth SBP and DBP values, and PTTs in both recording scenarios to investigate the feasibility of BP estimation. We used absolute Pearson’s correlation coefficient as the correlation metric. As there is a general consensus in the community that BP is subject-specific, our analysis was first done on a per-subject basis, where we were interested in both the actual correlation and the detection of the two subjects who reported overall higher BP and for whom we could reasonably assume to have stiffer vessels due to a history of smoking and higher age. However, since we only have a handful of samples for each subject, we cannot make statistically significant conclusions. Subsequently, we also checked the correlation between SBP/DBP and PTT in general, looking at the whole population at once. Our aim was at investigating whether we find consistent correlation patterns on both general and subject level, separating the different elicited hemodynamics and corresponding BPs. Looking at general correlations on the full population, we can also make conclusions about statistical significance of our observations. We used the MATLAB internal computation of p-values in the corr() function, which checks the probability of getting a correlation as large as observed by random chance, when the true correlation is zero.

5.5 BP Prediction Analysis

As the ultimate goal (that is most valuable to medical professionals) is to numerically estimate BP, we also trained a Random Forest regression model to predict SBP and DBP. We used each of the recordings as one instance with the corresponding ground-truth SBP and DBP label. We then used the average PTT in the recordings as our first feature, and then extended the feature vector with additional features describing the morphological properties of the waveform in that recording. In accordance with related work [17] we decided to use the average cycle width (\(T_c\)), time from start of cycle to systolic peak (\(T_s\)), time from systolic peak to cycle end (\(T_d\)) and areas under the curve for each of these sections (\(AUC_c\), \(AUC_s,\) and \(AUC_d\)). We did not choose any features relating to the diastolic notch, as it is difficult to precisely detect even in a traditional contact setup, while in our case the waveform is additionally distorted due to the remote nature of the measurement. Finally, we also added age and sex as demographic features describing each individual.

We implemented a leave-one-subject-out (LOSO) evaluation framework with personalization, where in each iteration we use the data of \(n-1\) subjects for training and the left-out subject for testing. A small part of data belonging to the test subject (in our case one instance of rest and one instance of activity scenario) is added to the training data in each iteration to personalize the model. While we are aware that this is slight overfitting to the test subject, we decided to still use personalization, as consensus in the field is that a general BP estimation model usually works much worse compared to a personalized model [24], even when the data is obtained in a highly controlled setting with contact sensors. We trained a Random Forest regression model due to relatively low time complexity and good historical performance, and evaluated the performance for SBP and DBP individually using the Mean Absolute Error (MAE) metric.

7.1 Contributions

The main contributions are as follows:

We can also conclude, in agreement with related work [21], that using the NIR band is preferred in MW PTT monitoring compared to traditional visible spectrum, since NIR light reaches deeper skin layers with arterioles, exhibiting better pulsatility compared to shallow layers reachable by blue wavelengths, while also increasing the length of the blood traversal path and delay between the waveforms. Feasibility of the proposed method was confirmed in controlled laboratory conditions, showing that usage of NIR and green wavelengths allows for acquisition of stable rPPG waveforms suitable for careful signal processing [18, 21] and subsequent PTT calculation. Despite this, some challenges remain before practical application, which are listed and discussed in detail in the following section.

7.2 Limitations

We acknowledge that a non-clinical experiment with 9 subjects cannot be used to make statistically significant general conclusions. A large-scale clinical validation would be required instead, but this requires access to more (varied) subjects and prolonged observations with continuous ground-truth BP. The former is always preferred to a cuff, as it offers much richer BP information, but is measured using an intravenous catheter, which can only be feasibly used with trained medical personnel. Such a large-scale clinical validation (in a hospital) is exceptionally challenging to organize and makes more sense for subsequent steps beyond a feasibility study. Ideally, such an extensive evaluation would include hypotensive and hypertensive patients, both with mild (no medication) and more severe (using medication) cases.

We also made some simplifications in our work when doing channel separation – we approximated some signal contributions which might be refined further. Additionally, channel separation might be inaccurate when the light source spectrum does not cover the relevant spectral bands we are interested in. In future work, we will consider strictly data-driven algorithms for channel separation that are independent of the quantum efficiency and light spectrum information, for instance using a genetic algorithm where the channels are initially a random linear combination of all three colors and the coefficients corresponding to each color are iteratively derived by using the error of a BP regressor as the fitness function. Another option would be to use a projection method, generating a densely populated space of candidate coefficients and try to find the best performing ones.

Furthermore, we were quite strict with removal of noisy data, since our interest was in fundamental feasibility rather than practical application. Signal noise (e.g., movement) still remains a challenge when considering wide-spread use. During our lab experiments, we circumvented this by ensuring a stable sitting position and fixed arm position immediately after the exercise finished. This combined with the processing pipeline allowed for very robust detection of reference points, however, the method used to detect those should always be verified for performance and stability when dealing with noisier waveforms, as it is vital for PTT computation. In practical BP screening applications it could be partially resolved with a similar setup using a foam padding on which a subject places their hand. The setup would also consist of an enclosure ensuring no interference of ambient lighting while providing a light source with the desired spectrum. Generally, sunlight has a very broad uniform spectrum containing all the bands we are interested in, so it could be used for our proposed method, however, using common LED bulbs is not suitable due to their poor spectrum lacking NIR and IR part. Still, an LED array containing diodes that also emit longer-wavelength light is feasible and relatively inexpensive to construct [12]. Such a setup could be used in a doctor’s office, especially when screening many subjects (e.g., groups of school children). Even more feasible (due to required personalization) is home use for telemedicine (e.g., elderly people), as self-measurement of BP with a cuff-based device (most common in general population) is very impractical and cumbersome, potentially causing incorrect measurements or even refusal to do regular measurement.

Finally, as we decided to use the palm as the measurement site due to tissue properties, our work is palm-specific to a degree, although we expect these findings to hold for other well-perfused skin locations. This remains to be validated, as variations in both skin tissue (e.g., melanin content, thickness) and measurement location should be investigated. However, as our purpose was to validate the early feasibility of remote MW PTT measurement and subsequent BP estimation, we limited our prototype setup and scenarios to be robust and reproducible at the cost of variations in measurement sites.

We used a relatively simple regression model to check the feasibility of explicit BP prediction and get a baseline result to compare against related work. The evaluation framework was designed to use personalized calibration, meaning some small part of data of each subject was used in the training in order to personalize the regression model. While this is not ideal, the consensus in the community is that building a robust generalized model is not feasible due to subject-specific relationship between PPG features and BP [24], meaning that personalization or calibration is required (or at least desired) in all models for BP estimation, including those based on multi-site PTT [14]. More specifically, there can be subject specifics in many biological properties, such as blood thickness, arterial wall thickness, elasticity, and other parameters related to PPG and BP. While early results seem promising, this evaluation is still in early stages and will be extended with more sophisticated models, additional features and other robust evaluation schemes.

Despite these limitations and further research opportunities, this work showed the importance of a fundamental understanding of both human physiology and camera physics for contact-free single-site MW measurement, while highlighting and addressing important challenges using novel methods. We showed that it is feasible to estimate PTT in such a way and confirmed that it highly correlates with BP, offering potential for remote MW BP estimation, which could be a useful part of a potential telemedicine system, as one of the parts in the growing framework of AIoT.