Assessment of the vocal status of multiple sclerosis patients - comparison with healthy subjects and evaluation of vocal rehabilitation

This study was carried out in collaboration with the team of speech therapists at the Don Gnocchi Hospital in Milan, analysing the speech performance of 70 patients that suffer from MS: one group was treated with standard therapy, a second group with innovative LSVT-LOUD therapy and a third group was not treated. The recordings provided include 3 repetitions of the vowel /a/ and a free speech (about 1 min), which were provided for each patient before and after therapy. After excluding not valid recordings (saturated or too noisy), two subsets were created for the analysis of the /a/ and the free speech. Using scripts developed in the Matlab R2020a environment, 9 descriptive statistics of Harmonic-to-Noise Ratio (HNR), intensity, Cepstral Peak Prominence Smoothed (CPPS) and fundamental frequency (fo) were extracted; in the case of the /a/ other 9 amplitude and period stability parameters were obtained. The script first performs a pre-processing that selects non-silent harmonic signal frames with frequency jumps between adjacent frames not greater than half octave. The parameters of the recordings of 60 healthy subjects, which are available at the Turin Polytechnic, were extracted with the same script. By comparing the vocal parameters of SM patients before therapy with those of healthy subjects, the most distinguishable values have been identified. In addition, comparing the values of the patients pre and post therapy, the parameters mainly affected by the therapy have been identified. Parameters identified for the vowel /a/: stability parameters of amplitude and period, 5° percentile (prc) and standard deviation (std) of CPPS, 95° prc, range and std of fo, mean, median and 5° prc of HNR. Parameters identified for the free speech: 95° prc and std of CPPS, mode of fo, std of HNR. Then, SM patients were analysed observing the difference between the parameters extracted at T1 (post-therapy) and T0 (pre-therapy). The patients were divided into three classes according to the therapy (LSVT-LOUD, ACTIVE, no therapy) and the most representative features were sought for distinguishing the therapies. For this aim, a combinatorial algorithm based on the logistic regression model was used. Taking two classes at a time, the model was tested with single features and with all possible combinations of 2, 3 and 4 features and selecting those that exhibited the best classification performance. The same operation was performed by weighting the features with the reciprocal and the complement to one of the std of the three repeated vowels, but no relevant improvements were observed. Subsequently, the best feature combinations were validated (5-fold cross-validation) through the Matlab APP Classification Learner. The best performances were obtained for the /a/ using the stability parameters vfo, apq and ppq and the statistics 95° prc, 5° prc, range, mode of HNR and std, 95° prc of fo; for free speech using the statistics mean, median, std, 5° prc of CPPS, range, mode, kurtosis of HNR and mean, range, skewness, 95° prc of fo. The best validated accuracy was of about 75% for free speech. The consistency between the perceptual evaluation of the experts, using the G value of the GIRBAS scale, and the obtained outcomes was assessed. The differences in G values between T1 and T0 were compared to the feature differences, reporting the results as a confusion matrix and taking the experts' assessment as a reference. In this last analysis, many errors were related to the poor resolution of the GIRBAS scale (0 to 4).