Quote
Content
Photoplethysmography imaging (PPGi) allows for estimating the cardiac activity with a camera due to subtle skin color variations with a frequency associated to blood volume pulsations. To enhance signal-to-noise ratio, it is common practice to average all pixel values in a region-of-interest (ROI) over time, resulting in a single waveform. As recently demonstrated using RGB cameras, considering, e.g., two pixel subsets within the ROI reveals mutually greatly dissimilar waveforms attributed to skin inhomogeneity. In this work, we utilized hyperspectral imaging (HSI) with 16 bands for increased wavelength resolution and principal component analysis (PCA) for a more rigorous analysis of waveform distribution.A camera (VRmagic D3, CMOS CMV2000, [470,620nm] wavelength range, 25 Hz) was usedto record the forehead of three volunteers (male, [27,30y]) from a distance of 10cm for 20s duration. A ring of ultraviolet and white LEDs (Falcon FLDR-i100B-UV24-W) was mounted around the lens providing frontal illumination. For each video (ROI: 60x250pixels≈0.96x4cm) and camera band, we performed PCA by regarding pixel indices as variables and pixel waveforms as observations. In the space spanned by the first two principal component scores, we identified three disjoint classes, each containing 5% of all pixels. We computed an averaged waveform for each class and a reference using all pixels. Our results showed a cardiac component in the reference waveform of HSI bands covering [530,580nm]. Additionally, one class exhibited a waveform similar to the reference (mean absolute percentage error: 9.35±2.25% (mean±std)), despite containing only 5% of all pixels. On the other hand, the other two classes contained mutually greatly dissimilar waveforms with inverted sign and did not show a distinct cardiac component. The spatial distribution of pixels within classes did not reveal any larger clusters of connected pixels.In conclusion,