Fig. 3: Construction of the PPG (a) and TOI (b) signals from a facial video.

To obtain the PPG signal from the regions of interest, the frames are first processed to reduce noise and amplify the relevant information. Then, three signals are created per region of interest (ROI) by spatially averaging the pixel brightness of red, green, and blue over the entire ROI in each frame. In a last step, independent component analysis (ICA) is applied and the signal is band-pass filtered and smoothed to obtain the PPG signal. In the case of transdermal optical imaging (TOI), the information in each frame is divided into three color channels that are each comprised of an 8-bit color stack. Each bit of the stack constitutes a bitplane. Subsequently, bitplanes of the stack with characteristic fluctuations in one's are isolated using a machine learning model. With this data, a spatiotemporal map of hemoglobin concentration is created, regions of interest are identified and the hemoglobin concentration is spatially averaged in each frame to obtain a hemoglobin signal for each region of interest.