Investigating multi-wavelength photoplethysmography for hemodynamic monitoring

Abstract

Photoplethysmography (PPG) is the predominant technique used by modern wearable devices to measure cardiovascular parameters, from simple heart rate and its variability to respiratory rate, and cardiovascular age. In clinical practice, PPG is the technique behind the pulse oximeter that provides vital information about peripheral oxygen saturation – a parameter that modern wearables also measure. The confguration of the sensor in the common refectance mode is relatively straightforward with a light source illuminating the skin and an adjacent photodetector detecting the backscattered light. The detected intensity is modulated by the light-absorbing cutaneous blood volume which has a periodic component that varies with the beating of the heart. An important factor in obtaining the signal is the wavelength of the light source. Due to the optical properties of the skin chromophores, the penetration depth of light varies so that in general in the visible light and in the very near-infrared (NIR) part of the spectrum, longer wavelengths penetrate deeper than shorter wavelengths. Multi-wavelength PPG (MWPPG) utilizes this by using several wavelengths of light simultaneously with the aim of providing information from the large arteries buried deep in the tissue to the more superfcial arterioles and capillaries.

This thesis contributes to MWPPG research by developing a fngertip-based method that enables the recording of MWPPG signals under controlled and measurable sensor contact pressure. The method is initially demonstrated with a device consisting of fve light-emitting diodes (LEDs) with central wavelengths ranging from 465 to 880 nm. The method is then developed further by inverting the typical MWPPG sensor setup of (relatively) narrow-band LEDs and a wideband photodiode to a confguration consisting of a wideband light source and a spectrometer capable of extracting 99 different channels with a range of approximately 434–731 nm. Using the oscillometric blood pressure (BP) measurement technique, the method is demonstrated to extract BP simultaneously from large arteries using wavelengths over 630 nm and from more superfcial vessels using wavelengths under 590 nm. The part in the middle, 590–630 nm, forms a transition band where a jump from probing superfcial, low-pressure, blood vessels to deeper, high-pressure, vessels occurs. The experimental results are corroborated with a Monte Carlo (MC) photon propagation model.

MWPPG signals are then further studied under different levels of sensor contact pressure. Static pressure held at relatively high pressure levels is shown to cause different responses between the wavelength channels, and, based on the MC results, is hypothesized to reveal information about the blood vessels primarily responsible for vasodilation in response to pressure stimulation. Additionally, a separate study is performed to study the effects of sensor contact pressure on measured peripheral oxygen saturation (SpO2), pulse arrival time (PAT) and pulse waveform features. The results show that especially sensitive are the parameters that are based on amplitudes or DC level.

Finally, in the age of rapidly advancing artifcial intelligence, a mathematical model is developed to generate synthetic PPG signals. The model can help tackle problems related to, for example, dataset sizes, bias, and data privacy. The synthetic model is demonstrated with a wearable device that uses convolutional neural networks (CNNs) trained with only synthetic data to estimate heart rate (HR) from measured signals. The performance of the CNNs often exceeds the more traditional methods used to calculate HR, providing a path to apply the synthetic model to more complex problems. This is briefy demonstrated by combining the synthetic model with another mathematical physiological model to generate synthetic oscillograms similar to those obtained with the presented MWPPG instruments.