REPOSITORY > RESULTS

Abstract

Omnipresence of sensor-equipped devices spurred rapid development of e-health and mhealth
applications in the past decades. Despite their wide-spread adoption in the form of
wearables, such devices are ultimately not a universal or ideal solution for regular health
monitoring due to their reliance on battery, requiring skin contact and general obtrusive
nature. Ideally, the need for direct user-device interaction should be completely removed in
the paradigm of ubiquitous and pervasive computing, which can be achieved using contactfree
sensors such as radars and cameras that monitor different parts of electromagnetic
(EM) spectrum. These devices can be used to monitor different physiological parameters
in an unobtrusive manner, making them feasible for subjects who cannot use wearables
(e.g., neonates, burn victims, elderly with dementia).
We initially explored the potential of radio-frequency part of the EM spectrum, measured
by radars, for detection of complex hemodynamic states. These are expressed via
several physiological parameters, including respiration. Radars allow for measurement of
periodic thoracic expansion and contraction even in challenging conditions, such as night
time and occlusion, making them ideal for sleep monitoring. We proposed a novel branched
neural network architecture that can take a different number and type of input signals. We
showed that we can detect five different hemodynamic states available in a public dataset
(including apnea) with up to 0.83 accuracy and F1 score when using only radar signals
as input. These results were only 4-5% behind traditional contact sensors, confirming
feasibility of radar-based physiological monitoring.
In the second part, we investigated the feasibility of using the visible part of the
EM spectrum, specifically the feasibility of a modified consumer RGB camera for multiwavelength
(MW) pulse transit time (PTT) measurement between different skin layers.
Different wavelengths penetrate to different depths and allow for depth-specific photoplethysmogram
(PPG) reconstruction. These can be used for MW PTT computation and
subsequent blood pressure (BP) estimation. We found that algorithmic channel separation
of mentioned PPGs is mandatory due to the imperfect nature of image sensor design, which
causes spectral overlap between PPGs from different depths. We thus developed several
algorithms that allow for data-driven camera-independent channel separation, which in
turn allows for precise measurement of MW PTTs. Finally we confirmed on an in-house
dataset that such MW PTTs are well-correlated with BP and that a personalized regression
model can be trained to predict both systolic and diastolic BP with errors within clinical
standards.
Overall we showed in this dissertation that contact-free sensors leveraging the information
from the EM spectrum are an affordable unobtrusive alternative to wearables, and
can achieve similar performance in monitoring of important physiological parameters and
states. While some limitations and challenges remain, such as difficult uncontrolled conditions
and privacy concerns, there is potential for implementing the proposed methods in a
single device, which could immensely improve the speed, cost and comfort of physiological
monitoring both at home and in hospitals.