The relationship between disease status and acoustic analysis of speech in Multiple Sclerosis

© 2021 Gustavo NoffsMultiple Sclerosis (MS) is a common cause of disability in young adults. Modern MS treatment has markedly improved long-term outcomes for the most common type of MS, i.e., relapsing-remitting MS. However, optimal pharmacological treatment decisions depend on early detection of disease progression. Validated assessment of progression includes disability scales and standard structural magnetic resonance imaging (MRI) of the brain (brain volume changes), but these are not sensitive to the early disability worsening. More dynamic tools are required to increase the responsiveness of MS monitoring. Objective speech analysis offers several potential advantages for disease monitoring, including short assessment time, low patient effort, automation, remote assessment (via smartphone or mobile device), improved reliability via elimination of inter and intra-rater variability, and wide disability-wise suitability as speech continues to be produced even by severely disabled patients. The general aims of this thesis were to (1) review and identify gaps in the literature regarding the validity of speech to monitor MS progression, (2) examine the relationship between acoustic metrics and general neurological status or cerebellar impairment in people with MS (pwMS), and (3) describe differences in health-related acoustic metrics between mobile devices' recordings and professional-grade microphones, both of which are commonly used in research settings. Most previous studies described only differences between pwMS and healthy controls for a narrow set of acoustic speech features, with almost nonexistent information on disability or neuroimaging correlates of speech impairment in MS. In the current study, we observed a moderate strength relationship between speech acoustic features and general disability, cerebellar ataxia, whole brain and cerebellar MRI-measured white matter volume and lesion load. Single-feature metrics independently contributed to an acoustic composite score that differentiated between mild, moderate, and severe levels of disability and correlated with EDSS scores in people with or without overt dysarthria (subclinical speech impairment. A similar ataxia acoustic composite score predicted an abnormal time to complete the 9-hole peg test with 85 percent accuracy, including in a naive dataset. Acoustic models were non-specific, with the same speech features compounding both the general and cerebellar models. Lastly, we observed wide differences in acoustic values for most speech features of interest between mobile devices and a professional recording set. Differences did not affect the relationship between acoustic scores and EDSS, although tested only in a small cohort. The current studies describe various speech acoustic features in their relationship with clinical disability and neuroimaging metrics in the same participants and further explores subpopulations of interest according to disability level and presence of dysarthria. Present results support the validity of speech models to, at least in part, reflect neurological status in MS. Current and previous findings strongly suggest against the interchangeable use of recording methods, particularly between different microphones and different distances between microphone and sound source, but do not preclude the use of mobile devices for health-related speech data acquisition