Automatická analýza poruch řeči u neurodegenerativních onemocnění

Abstract

Biomarkery získané automatickou analýzou hlasu se těší rostoucímu zájmu logopedů i neurologů v souvislosti s možností rozšířit dosud značně limitovaná neinvazivní měření motorických poruch řeči způsobených neurodegenerativními onemocněními. Akustické řečové příznaky mohou být v klinické praxi vskutku neocenitelné, avšak pouze tehdy, jsou-li podloženy vysvětlitelnými hypotézami a popsány z hlediska dopadu onemocnění, pohlavní dvojtvárnosti a vlivu stárnutí. Spletitost interpretace těchto faktorů tvoří hlavní překážku bránící využití hlasových analýz v klinické praxi, která navzdory značnému rozvoji tohoto oboru nebyla dosud překonána. Tato práce zavádí metodologii pro získání srozumitelných akustických příznaků pomocí číslicového zpracování signálů a strojového učení a modelování pohlavní dvojtvárnosti a vlivu stárnutí; vyšetřuje velkou databázi pacientů s neurodegenerativními onemocněními a diskutuje použitelnost metody na základě experimentálního odzkoušení metody v klinické praxi.

Automated vocal biomarkers are becoming increasingly desired by speech pathologists and neurologists in order to extend current noninvasive measures of speech motor abnormalities associated with neurodegeneration. Clinical information concerning acoustical features and patterns can be invaluable only if the measures are based on interpretable hypotheses and described with regard to the impact of the disease, sexual dimorphism, and any age dependency. The complexity of interpretation is the main barrier between engineering applications and clinical practice. Despite huge developments in the field, no applicable methodology for complex acoustic analysis have been proposed yet. This thesis aims to design and define the automated acoustic analysis that could provide profound insight into speech disorders caused by neurodegeneration.The database used in this research is comprised of 42 subjects with idiopathic rapid eye movement sleep behavior disorder; 32 subjects with early, untreated Parkinson’s disease; 26 subjects with treated Parkinson’s disease; 22 subjects with multiple system atrophy; 15 subjects with progressive supranuclear palsy; 18 subjects with untreated Huntington’s disease; 13 subjects with treated Huntington’s disease; 17 subjects with cerebellar ataxia; 101 subjects with multiple sclerosis; and 284 subjects with no history of neurological or communication disorders (HC). Each speaker performed the sustained vowels /A/ and /I/, took a rhythm test, read a passage, performed a monologue, and completed a diadochokinetic task. Acoustic signals were recorded using a standardized procedure. Signals were processed by fully automated methods. Normative data were estimated by selecting an HC subgroup to match any speaker in terms of age and sex. All measured values were normalized by corresponding normative data and expressed in terms of probabilities and z-scores. A novel approach for supervised learning based on the weighted fusion of z-scores (SWFS) was employed for recognition of certain tendencies of disordered speech. Finally, the methodology was implemented in a software application and tested extensively in a clinical setting by an experienced speech-language pathologist for more than one year. Based on a thorough evaluation, the proposed processing methods represent the most precise technology for the extraction of given acoustic features available up to the date of this thesis. The majority of speech features showed abnormalities in at least one disease group compared to the HC. Individual speech features did not exhibit specificity to disease. Nevertheless, clear tendencies with discriminative qualities were observed in combined features. The SWFS showed the ability to decompose any speech pattern and quantify its severity in terms of abnormalities, whereas the recognition accuracy was comparable with conventional classifiers. The clinician rated the methodology as practicable, clinically relevant, interpretable, and of benefit. Two case studies are presented to demonstrate the capacity of the proposed methodology.This thesis introduces a methodology for the extraction of highly interpretable speech features using a new approach in digital signal processing, machine learning, and the modeling of sexual dimorphism and age dependency; investigates a large database of patients affected by neurodegeneration; and discusses clinical applicability based on the successful experimental use of the implementation in a clinical setting. The methodology was designed to meet the demands of clinical practice with a hope that the presented results will lead, inspire, and bolster the future development of automated methods for the assessment of speech disorders.