Assessment of Vocal Fatigue of Multiple Sclerosis Patients. Validation of a Contact Microphone-based Device for Long-Term Monitoring

Multiple Sclerosis (MS) is a neuro-degenerative disease of the central nervous system, involving motor symptoms such as spasticity, weakness, language disorders and dysphonia; also, fatigue is often considered one of the most debilitating symptoms of MS. An acoustic analysis of MS is performed with vocal material supplied by the speech department of Don Gnocchi hospital (Milan). The data-set includes voice recordings of two subgroups of identical dimension (i.e., sixteen subjects) correspondent to healthy subjects (HS) and MS. Speech material consists in the vocalization of the sustained vowel /a/, the reading of a phonetically balanced speech (Notturno) and one minute of free speech for each subject. Additionally, long-term recordings are carried out with VH device only, covering a maximum period of four hours. About signals acquired with in-air microphone, pre-processing and harmonic frame selection phases are executed in Matlab (R2022b) environment and then, the extraction of parameters is operated. Using a Logistic Regression (LR) model, data are classified dividing them into two classes (HS and MS). The LR model is trained using a single and a combination of 2, 3, 4 features; the combinations exhibiting the best performance in terms of accuracy and Area Under The Curve (AUC) are selected; then, for these, 5-fold cross-validation is implemented. Best performances are obtained for the reading task (accuracy equal to 92.3%) by selecting 3 features, which are gender, 5° percentile of Cepstral Peak Prominence Smoothed, and harmonic frames ratio V/uV. The expanded uncertainty U(p) is evaluated, providing a confidence interval; when the confidence interval includes the discrimination probability, the classification of the subject is considered "non-classifiable" and new classification metrics are defined, such as Realistic Accuracy and Fraction of Classified (FoC). The implementation of this procedure to combination showing best performances (gender, CPPS (5,prc) and V/uV), results in FoC of 92.3% and higher accuracy. To validate VH device, parameters from the microphone in air are compared to ones from VH by calculating differences Δ between these measures. Considering sustained vowel /a/ task, the analysis is performed on parameters local jitter (%), local shimmer (%), CPPS (median) (dB) and CPPS (std) (dB); for balanced and free speech task differences are carried out for fundamental frequency f0 (Hz) and CPPS (dB). For local jitter, CPPS (median) and CPPS (std) the validation can be considered passed, while for others, such as local shimmer, significant differences are noted. Since the two devices have different characteristics they receive as input different signals, the use of VH device requires the definition of specific cut-off values for the extracted parameters. A proposal to assess fatigue is conducted using differences δ between parameters extracted from long-term and short-term monitoring; this comparison is performed considering the parameters fundamental frequency f0, CPPS (dB), Background Noise Level (90° percentile) in dBA and Sound Pressure Level (dB). No significant difference with regard to the fatigue is found; however, limitations can be overcome through both an increase in the data-set and in the time interval of the records, being acquisitions too short to demonstrate fatigue. Eventually, an evaluation of five vocal dose measures as indicators of long-term vocal folds tissue exposure to vibration is provided.