Revealing physiological changes through colour

Thesis

This dissertation extends remote vital sign monitoring by examining whether it is possible to track oxygen saturation with visible-light video cameras by filming an exposed skin region (the face). Two further questions are also investigated as part of the presentation of this thesis: the in uence of skin melanin on the ability of video cameras to track oxygen saturation, and the ability of video cameras to assess the absolute value of oxygen saturation at any given point in time.

A study involving human volunteers was designed to answer these research questions in a controlled environment, and 46 subjects were recruited. The subjects were divided into six different skin phenotype groups (Fitzpatrick Scale I{VI). For a period of up to one hour, the volunteers' oxygen saturation was varied between 80% and 100% and videos of their faces were recorded in that time. The volunteers' reference oxygen saturation was recorded using two pulse oximeters placed on the index fingers of each hand. Another study was carried out in a neonatal high-dependency unit to test the algorithms for remote vital sign monitoring in a challenging clinical setting.

A theoretical basis for the derivation of oxygen saturation monitoring algorithms is developed throughout the dissertation, and two methods (baseline and RoR) were built upon this. The baseline method has parallels with oximetry and is based on tracking overall changes in the colour of the skin while minimising the effect of lighting, while the RoR method mimics pulse oximetry in form and assesses the relative change in amplitude as measured by each colour channel while oxygen saturation changes. Both methods are found to work well for subjects of Fitzpatrick phenotypes I{IV (average coeficient of determination r² of 0.79 and 0.71 for baseline and RoR respectively across all Type I{IV subjects), variably for Type V subjects (average r² values of 0.64 and 0.48), and not work for Type VI subjects (average r² values of 0.19 and 0.01). Although the RoR method yields noisier results than the baseline method, it was found that the timecourse of the latter varied with respect to the oxygen saturation.

A simple linear model used to derive oxygen saturation directly from the video camera data was thus devised using the RoR method only, and applied to Type I{V subjects. The subjects were divided into randomly generated but equally split training sets and test sets and the model parameters learned from the training set were evaluated in the test set. This was repeated over 10 iterations and the overall mean absolute error between the estimated oxygen saturation using the video camera data and the reference oxygen saturation using the pulse oximeters was, for Type I{V subjects respectively: 3.0%, 2.3%, 2.5%, 2.5%, and 4.6%.

Authors: Guazzi, A

• DOI: 10.5287/ora-mvam4ny4w
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford