Classification of multiple sclerosis patients through vocal features - Performance evaluation by using different vocal indexes and software applications

Multiple sclerosis (MS) is a chronic disease of the central nervous system that affects the brain and spinal cord. MS is an autoimmune disease, meaning that the body's immune system mistakenly attacks its own healthy tissue, which in the case of MS is myelin, a substance that lines nerve fibres and helps transmit nerve impulses efficiently. MS involves various body activities of patients, including language. This study was conducted in collaboration with Don Gnocchi Foundation in Milan. Vocal recordings of 16 subjects with MS and 16 subjects (HS) without MS were acquired, which include 3 repetitions of the vowel /a/, free-speech of about 1 min, and the reading of a phonetically balanced text. For each subject, an air microphone (MI) and a contact microphone (VH) were used to simultaneously acquire the vocal signals. Only for the MI, the available traces were manually analysed to exclude invalid recordings (saturated or too noisy) and select the parts of interest. Then, Matlab scripts were specifically developed to subdivide the vocal signal in frames and extract the parameters Harmonic to Noise Ratio (HNR), Cepstral Peak Prominence Smoothed (CPPS), fundamental frequency (fo), and signal intensity (RMS). Each parameter is represented by means of 9 descriptive statistics (mean, median, mode, standard deviation, range, 5 percentile, 95 percentile, skewness, kurtosis). For the vowel /a/, other 9 stability parameters of amplitude (shimmer) and period (jitter) were extracted. The purpose of this thesis is to identify the parameters that better distinguish MS vs HS classes and two indexes that group some of the extracted parameters have been investigated: the well-known Acoustic Voice Quality Index (AVQI) and the Warning Score (WS), which is a new index proposed to assess the vocal health status of subjects. AVQI depends on the parameters jitter, shimmer, CPPS, HNR, Spectral Slope and Tilt extracted by a concatenation of 3 seconds of sustained vowel /a/ and 3 seconds of reading. These parameters were extracted through 3 applications (Matlab, Praat and VOXplot) and they show significant differences: in particular, Matlab CPPS mean values differ of about 3 dB compared to Praat. In terms of AVQI, only parameters extracted with Praat and VOXplot were used: subjects were classified using the Logistic Regression (LR) model, by comparing their accuracy (Acc=70.8% with VOXplot and Acc= 61.5% with Praat) and area under curve (AUC=0.63 with VOXplot and AUC=0.54 with Praat). The index WS depends on the parameters local jitter, local shimmer, mean and standard deviation of CPPS extracted using Matlab scripts from the vowel /a/ by both MI and VH. Subject were also classified according to the index WS by the LR (Acc=41.7%, AUC=0.36). Classification results between HS and MS are not outstanding, thus highlighting that even if HS do not have MS, they could exhibit dysphonic behaviour. For this reason, from a data set of 58 True Healthy Subjects (THS), 12 subjects were extracted in order to have a balanced data set with 12 MS and the classification was repeated using the same LR model, obtaining significantly better results: Acc=100% (AVQI evaluated by VOXplot), Acc=92.3% (AVQI by Praat), Acc=87.5% (WS by Matlab). The LR classification using both AVQI and WS does not provide improvement (Acc=91.7%). Furthermore, a comparison between perceptual assessment (G and A indexes of the GIRBAS scale) and input parameters of the WS index showed negligible correlation both for MS and HS.