Data processing

Our analysis focused on signals from the forehead because it is well perfused and has a large, homogeneous and plane area, which helps to avoid ballistocardiographic effects^15,20. The same procedure to define the ROI was applied to iPPG and LSI recordings (Fig. 3). The ROI was kept statically throughout each measurement epoch. A tape on the subject forehead served as base for ROI definition (a). The area within the tape was identified (b) and the smallest circumscribing rectangle (c) was fitted. Through moving inward, the largest inscribing rectangle (d) was defined. This rectangle was reduced by 20 pixels in length and width (e) in order to account for slight body movements that otherwise might cause the tape to enter the static ROI (f) and introduce undesired effects. All ROIs were visually verified.

ROI definition algorithm. See text for details.

For each measurement epoch, we analysed the iPPG video in an interval of 10 s duration starting 8 s after beginning of the respective measurement epoch. Afterwards, we averaged over the ROI to yield a time varying signal. A temporal linear interpolation was applied to construct an equidistant time series at a sampling frequency of 100 Hz.

iPPG was quantified in terms of a perfusion index PI from each interval. PI uses the iPPG signal’s AC and a DC component¹⁸ and was defined by PI=ACiPPGDCiPPG. To calculate the DC component, iPPGt was filtered by a Butterworth low-pass filter of 3rd order and cutoff frequency of 0.4 Hz to yield a DC signal iPPGDCt. To define the AC component, another Butterworth low-pass filter was applied (5th order, cutoff frequency 8 Hz). The AC signal iPPGACt resulted as difference between suchlike filtering and iPPGDCt. Using iPPGACt, single heart beats were detected by the method proposed by Lázaro et al.²¹. Heart cycles were then defined around the time instants of detection using the median interval between all detections. Heart cycles showing a mean correlation lower than 0.3 in a pairwise comparison between all cycles were sorted out. From the remaining cycles, an average cycle was calculated. The average cycle was used to determine the AC amplitude as difference between its maximum and minimum value. The DC component was taken as average value of cbPPGDCt at the time instants of single beats.

Laser speckle processing and reference PPG used the same time intervals as iPPG. The procedure to define the ROI for LSI was identically (but using the video images from the device). As feature, the mean flux value (Flux) was used (mean value over the ROI and time interval). Laser speckle data thus did not require any preprocessing as the feature extraction was readily provided by the device. The reference finger PPG was evaluated in terms of the perfusion index like iPPG and denoted as reference-PI.