Improved Pulse Pressure Estimation Based on Imaging Photoplethysmographic Signals

Matthieu Scherpf1, Hagen Malberg1, Martin Schmidt2
1TU Dresden, Institute of Biomedical Engineering, Dresden, Germany, 2TU Dresden


Objective: Imaging photoplethysmography (iPPG) enables the extraction of physiological signals from standard RGB video recordings. For the assessment of the human health condition, pulse pressure is of utmost importance and is usually determined from conventional blood pressure signals. Since blood pressure signals can so far not be measured contactless, we investigated the relationship to the pulse strength (iPPG signal amplitude difference). By comparing two different approaches for skin segmentation and iPPG signal extraction, we considered different methods in signal processing. Methodology: We analyzed 70 RGB videos from patients recovering from cardiac surgery to take patient specific pulse pressure variations into account. Each video recording has a duration of 30 minutes and was synchronously recorded with continuous blood pressure. We split each signal into non-overlapping 10 second segments resulting in 180 segments per recording. The processing pipeline started with the extraction of framewise skin segmentation using Deeplab and Levelset. Next, the optimal color channel combination (O3C) and green (G) color channel was computed for iPPG signal extraction. For the extraction of pulse pressure and pulse strength, the segment-wise median of amplitude differences was calculated. Disturbed segments have been excluded by signal-to-noise ratio (SNR). Finally, we calculated the Pearson correlation (r) between pulse pressure and pulse strength for each segment and compared the different signal processing methods. Results: Deeplab (r(O3C): 0.60, r(G): 0.48) achieved higher correlations between pulse strength and pulse pressure compared to Levelset (r(O3C): 0.45, r(G): 0.38). Our results also demonstrate O3C’s superiority regarding iPPG signal extraction and pulse pressure estimation with an average increase of 0.10 in r. Conclusion: Our results demonstrate the suitability of O3C and Deeplab regarding the iPPG signal extraction in the context of contactless pulse pressure estimation. Further work will integrate pulse shape analysis to investigate the estimation of absolute pulse pressure.