Development of problem-oriented strategies for biomedical and raman spectroscopic signal analysis

• Main Author: Bai, Yanru
• Other Authors: Liu Quan
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2020
• Subjects: Engineering::Bioengineering
• Online Access: https://hdl.handle.net/10356/137051
• Institution: Nanyang Technological University

The dissertation is concerned with developing techniques for the analysis of biomedical signals and Raman spectroscopic data. Three studies are covered in this dissertation, involving two non-invasive and powerful analytical tools, i.e. Raman spectroscopy and electroencephalography (EEG), particularly pain-related EEG. Raman Spectroscopy can detect the biochemical ‘fingerprint’ of a material and has been widely accepted as a mature analytical tool for the characterization and identification of molecules, cells and tissues in diverse biomedical applications. Despite its several advantages, such as being rapid and low-cost, non-invasive, non-destructive, and highly specific, Raman spectroscopy suffers from low signal intensity especially for biological samples. Thus, an effective method that can recover decent Raman spectra from Raman measurements with low signal-to-noise ratio (SNR) is highly demanded. On the other hand, the target species of interest is usually mixed with several other species in many circumstances. Quantitative analysis of target species is seriously influenced by the Raman contributions of non-target species in a complex mixture. Multivariate analysis has been extensively used as powerful assistance in the processing of Raman spectroscopic data due to its capability of coping with multidimensional datasets and exploring the complete spectral information simultaneously. There is an increasing demand for problem-oriented multivariate analysis methods to address challenges specific to a study. This dissertation is focused on developing chemometric methods with regards to these two aspects as described next. In the first study of this dissertation, a method based on Wiener estimation using a numerical calibration dataset is proposed to recover Raman spectra with low SNR. The model has been evaluated in both Raman phantoms and biological samples. The unique advantage of this method is that the calibration dataset is created numerically and thus can be easily extended for analyzing Raman spectroscopic data from diverse categories of samples. More importantly, the method demands fewer experiences from users to process spectra with unknown noise levels when compared to the two commonly used denoising methods, Moving-Average (MA) filtering and Savitzky-Golay (SG) filtering. The second study is to address the challenge of quantifying the target species in a mixture with multiple species. A strategy for target quantification based on orthogonal projection and basic component analysis is proposed, which can be potentially used in optical computing for fast data analysis. The effectiveness of the method has been demonstrated on both Raman phantoms and spectra synthesized from the experimental measurements of doxorubicin (DOX) and leukemia cells. Furthermore, our method has been extended to a preliminary study of the potential biomedical application in non-invasive diagnosis of cornea infection. Pain is an unpleasant experience related to substantive or potential tissue damage. An effective physiological pain assessment method is highly desired in pain clinical research and practice. EEG-based pain assessment has attracted a growing interest and demonstrated to be a useful alternative to self-report. The third study is related to EEG-based pain assessment at cross-subject level. We comprehensively investigate the relationship between pain-evoked EEG (pEEG) responses and spontaneous EEG (sEEG) activities as well as the relationship between subjective pain ratings and pEEG responses. Results reveal a significant correlation between one’s pain-evoked EEG responses and his/her spontaneous EEG in terms of magnitude, and a nonlinear relationship between the level of pain perception and pain-evoked EEG responses. Based on the observations, a normalization strategy is proposed to reduce the inter-individual variability in pain-evoked EEG and a new two-stage prediction strategy is developed. Results demonstrated significant improvement in the accuracy of cross-subject pain prediction.