Vasomotor assessment by camera-based photoplethysmography

2.1 Vasomotor stimuli

Different interventions are known to induce vasoconstriction. The cold face test (CFT) is a non-invasive method that provokes a reaction in the autonomic nervous system by applying a low temperature stimulus on facial regions. Beside bradycardia, CFT causes peripheral vasoconstriction [5]. An inverse effect is simply achievable by an abrupt increase of the ambient temperature which leads to an acclimatization of the body and causes vasodilation [6]. Within this study, dilation is generated by a halogen lamp (heat radiation) whereas constriction is induced by CFTs.

2.2 Data and technical setting

Experiments were conducted indoors with 10 healthy subjects (9 male, 1 female) between 21 and 33 years (24.9 ± 3.35 years). All participants were informed about experiment’s procedure and gave a written agreement. During cbPPG measurements, they lied relaxed on a tilt table and were asked to remain calm. We recorded the inner arm (shoulder to wrist) using an industrial camera (IDS UI-3370CP-C-HQ) at a frame rate of 300 fps, a resolution of 1024 × 200 pixels and a color depth of 12 Bit. Figure 2 shows the recording area for an example. Driven by a constant current, a three spot halogen lamp (3 × 75 Watts electrical power) was exclusively adopted as illumination. The lamp was aligned directly towards the arm and therefore also had a heating effect. In addition to the video recordings, we synchronously captured subjects’ electrocardiogram (ECG).

2.3 Experimental protocol

Figure 1 depicts the experimental protocol. Each experiment started with a resting phase. In order to treat this phase equally regarding the following CFTs, it was separated in three tests (W₁, W₂, W₃), each of them holding three subwindows of 30 s (W_{i ∈ {1, 2, 3}, j ∈ {1, 2, 3}}). After the resting phase three subsequent CFTs were carried out for every participant (W₄, W₅, W₆). Each CFT started with a 30 s baseline (W_{i ∈ {4, 5, 6}, 1}). After the baseline, an ice pack was applied to the forehead and remained there for further 30 s (W_{i ∈ {4, 5, 6}, 2}). The ice pack reached laterally to the temples. Eyes were not covered in order to avoid the ocular reflex. After removal of the ice pack, the subjects were recorded for further 30 s (W_{i ∈ {4, 5, 6}, 3}). The time between all tests was approximately 120 s. For each test, one video (90 s) was recorded separately leaving a total of 10 × 6 videos.

Figure 1 also depicts the assumptions for the peripheral vasomotor activities. The vasodilation effect, induced by warming, should manifest in a trend over the whole experiment. The vasoconstriction effect, caused by the CFT, should occur in each window W₄, W₅ and W₆. Note that such assumptions are qualitative, and interindividual differences must be expected (e.g. on the onset of vasoconstriction during CFT).

Figure 1:

Experimental protocol. In gray, the assumption for the peripheral vasomotor activity is shown.

2.4 Image and signal processing

For further processing, we only used videos’ green channel since it was shown to provide the strongest plethysmographic signal [3]. In each video, the arm was manually selected as static region of interest (ROI). The wrist and skin edges were excluded because they often yield cbPPG signals that are in counter phase to wanted signals [7]. Blocks of 32 × 32 pixels were fitted in the ROI and truncated when they exceeded the border (see Figure 2). Over all videos, between 70 and 100 blocks were built (82.2 ± 7.9). As the signal quality varies considerably between blocks, only a subset of blocks should be used. Therefore, from each block, a cbPPG signal was extracted by averaging its containing pixels for every frame. All signals were detrended and bandpass-filtered by a Butterworth filter (order 3) with cutoff frequencies of 30 and 200 bpm. The minima of signals’ beats were detected based on R-peak’s positions that were annotated in the reference ECG. Time delays (TDs) were then calculated as temporal differences between these peaks and the found minima (see Figure 2, bottom). The reason behind this procedure is that in high quality signals, minima will correspond to the start of systoles. In that case, variations in TDs just result from variations in the pulse transit time. Such variations are small compared to variations which appear in low quality signals. For poor signals, minima are not necessarily related to the beginning of systoles but occur randomly with respect to the R-peak resulting in a high standard deviation of TDs.

For further analysis, the signals from 30 blocks holding the lowest standard deviation of TDs were averaged (unfiltered) to build video’s optimal cbPPG signal s_opt. The raw signal s_opt was detrended and bandpass-filtered, and this time, minima as well as maxima were detected using the reference ECG. We determined beats’ height (distance between minimum and maximum) and built a pulse signal h_opt. Every h_opt was divided in consecutive 30 s windows and the median over each window was calculated resulting in 6 × 3 pulse values h_{cam(i, j)} per subject.

Figure 2:

Recording area with set ROI (black) and ROI (green) to extract optimal cbPPG signal (bottom). The plot shows the raw (gray) and the filtered (black) signal with detected extrema (red), determined by using ECG’s R-peaks (green).

2.5 Evaluation and statistics

The trend effect (vasodilation) was analyzed using the first window of each test W_{1, 1} − W_{6, 1}. We assumed the pulse values continuously to increase over the six windows. To measure this dependence, we first calculated Spearman’s roh between h_{cam(i, 1)} and W_i,1 over all subjects (10 values per window). Second, Page’s test was applied to validate the significance of the trend. The test provides a test statistic L which was transformed into a common p-value.

For the three CFTs (vasoconstriction), we assumed the pulse values to decrease from subwindow 1 to 2 and increase again from subwindow 2 to 3. Due to high intraindividual variations, values h_{cam(4−6, j)} were normalized on the median of h_opt from the respective video (4–6): h^_{cam(4−6, j)}. First, a Friedman test was used to determine if significant differences among the groups h^_{cam(4−6, 1)}, h^_{cam(4−6, 2)} and h^_{cam(4−6, 3)} are existent. Second, a post-hoc analysis according to Schaich and Hamerle was applied to explicitly test for significant differences between subwindows 1 and 2, 1 and 3, and 2 and 3.