Development of a Procedure for Sound Quality Evaluation based on Electroencephalography and Assessment Analysis

One of the big challenges in the XXI century, as an essential part of human brain analysis procedures, is the determination of mathematical models capable to explain and forecast the relationships between human activities and electroencephalography (EEG) signals. EEG signals produce data organized in temporal sequences with a structured behaviour and have been used for different purposes, from seizure detection and epilepsy diagnosis, to automatic detection of abnormal EEG, and recognition of Alzheimer’s disease brain activity, the detection of awareness, or the use of brain–computer interfaces (BCI) [28]. This thesis aims to carry out in parallel two pre-processing methods for the automatic removal of bad channels and portions of the signal with many artifacts, including the data filtering part as well, to finally try to figure out with which of the two methods better results could be expected for the study of brain waves subjected to different types of sound stimuli. The first approach was to use EEGLAB, a toolbox that can be used in MATLAB, while the second approach was to use AUTOREJECT, a library to automatically reject bad trials and repair bad sensors in magneto-electroencephalography (M/EEG) data. Signals from 43 healthy volunteers (22 males and 21 females) were acquired for the experiment using a commercially available wearable device: the Interaxon MUSE headband. The data acquisition process involved utilizing the MUSE application, which was paired with a smartphone via Bluetooth Low Energy (BLE) technology. After applying the 2 approaches and filtering the raw data, the data were processed by calculating the various power spectra densities and then the outliers were removed. After that, an analysis on the normal distribution of the data and a parametric ANOVA statistical test were conducted. The results show that six of the forty-three subjects analysed with EEGLAB were discarded as not having enough data to pursue this type of study. Differently, the AUTHOREJECT approach was able to retain all forty-three subjects. Regarding outliers removal, the two methods behaved almost the same with a 51.35 percent rejection rate. As far as Gaussianity analysis conducted with the Shapiro-Wilk Test is concerned, the two approaches also appear to be robust, with a slight advantage of the EEGLAB method (92.65%) over AUTOREJECT (85.22%). Even in the alpha frequency range this trend seems to be confirmed with EEGLAB (94.70%) surpassing the AUTOREJECT method (90.91%). The results of the ANOVA test show how data processed with the two approach for TP9-TP10 channels have a relatively high F-statistic, standing for a gap between the groups means (4.29  1.17) while a decrease in F-statistic values can be observed in alpha frequency range (1.73  0.47). Consistent results are also confirmed by the  values showing in the whole frequency range, for both EEGLAB and AUTOREJECT, similar behave, with p-values much less than 0.05 (4.9e-3  7.6e-3). An opposite trend as obtained for the F-statistic occurs for the alpha frequency range (0.16  0.08). A better acoustic physiological response was obtained analysing the Relative Alpha Band Power with the AUTOREJECT (F-statistic = 2.95  1.33,  = 0.042  0.056) method and EEGLAB method (F-statistic = 1.62  0.06,  = 0.16  0.02). In conclusion, despite the presence of some limiting factors such as data integrity, data quantity, device choice and others, is possible to say that due to the three analyses conducted in this thesis, the EEGLAB method and the AUTOREJECT method are very similar in terms of outliers removal and normality analysis of the data. While with regard to the ANOVA statistical test, nothing statistically significant could be stated in the characterization between one sound and another. Differently, we noticed how the AUTOREJECT method seems to perform better in the range of relative alpha power than the EEGLAB approach.