Hypophonia detection from voice recordings of patients with Parkinson’s disease

Jenko, Jaka (author), Sadikov, Aleksander (mentor), Georgiev, Dejan (comentor), Pesek, Matevž (comentor)

Parkinson's disease affects approximately 10 million people worldwide. With the recent advancements in the development of neuroprotective drugs aimed at slowing or halting the progression of PD, early diagnosis has become increasingly important. These early diagnostic methods are split between invasive and usually costly methods (DatScan, biomarker, and genetic testing) and noninvasive methods (gait analysis, olfactory function, emotion recognition, and voice analysis). Our research focused on detecting PD using voice analysis, which can theoretically be done quickly and remotely, thus being accessible to more people. We have created a new dataset with 24 control subjects and 9 patients with PD. The dataset comprised of demographical data (age, gender, and education), results from psychological tests (GDS-15, MMSE, FAB, and MDS-UDPRS III), and recordings of participants reading short compositions (three neutral compositions and six compositions with emotional content). Using this dataset, we tried to classify the subjects based on their reading accuracy, detected emotions in their speech, and extracted MFCC features. We have shown that using reading accuracy, we were able to correctly classify 6 out of 9 experimental subjects and all of the control subjects. Similarly, using the MFCC features of words containing 2 or more syllables, we could classify all subjects correctly. Unfortunately, we could not draw any conclusions using detected emotions from the subject's speech, as the used emotion detection model predicted most of the recordings to express a neutral emotion while labeling voice recordings from males with deeper voices as sad. This study strengthened the results from the previous research and showed that voice analysis could serve as a viable, cost-effective, and noninvasive method for detecting PD.

• Language: English
• Keywords: Parkinson's disease, dataset, speech impediments, voice recordings, reading accuracy, emotions, MFCC
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2024
• PID: 20.500.12556/RUL-165183
• Publication date in RUL: 26.11.2024

Secondary language

• Language: Slovenian
• Title: Zaznavanje hipofonije iz posnetkov govora bolnikov s Parkinsonovo boleznijo

Abstract

Parkinsonova bolezen prizadene približno 10 milijonov ljudi na svetu. Zaradi napredka pri razvoju nevroprotektivnih zdravil, katerih cilj je upočasniti ali zaustaviti napredovanje PB, postaja zgodnja diagnoza bolezni vse pomembnejša. Metode zgodnje diagnostike se delijo na invazivne in običajno drage metode (DatScan, biomarkersko in genetsko testiranje) in neinvazivne metode (analiza hoje, funkcija voha, prepoznavanje čustev in analiza glasu). Naša raziskava se je osredotočila na odkrivanje PB z analizo glasu, ki jo je teoretično mogoče opraviti hitro in na daljavo ter je tako lahko dostopna večjemu številu ljudi. Ustvarili smo nov nabor podatkov s 24 kontrolnimi osebami in 9 bolniki s PB. Nabor podatkov je obsegal demografske podatke (starost, spol in izobrazbo), rezultate psiholoških testov (GDS-15, MMSE, FAB in MDS-UDPRS III) ter zvočne posnetke udeležencev, ki so brali kratke sestavke (tri nevtralne sestavke in šest sestavkov s čustveno vsebino). Udeležence smo poskušali razvrstiti na podlagi njihove natančnosti branja, zaznanih čustev v njihovem govoru in pridobljenih značilk MFCC. Na podlagi natančnosti branja smo uspeli pravilno razvrstiti 6 od 9 bolnikov s PB in vse kontrolne osebe. Podobno smo z uporabo značilk MFCC besed, ki vsebujejo dva ali več zlogov, lahko pravilno razvrstili vse subjekte. Žal na podlagi zaznanih čustev iz govora subjektov nismo prišli do relevantnih zaključkov, saj je uporabljeni model za zaznavanje čustev napovedal, da večina posnetkov izraža nevtralno čustvo, medtem ko je glasovne posnetke moških z globljimi glasovi označil kot žalostne. Naša študija je okrepila rezultate prejšnjih raziskav in pokazala, da bi analiza glasu lahko služila kot izvedljiva, stroškovno učinkovita in neinvazivna metoda za odkrivanje PB.

• Keywords: Parkinsonova bolezen, nabor podatkov, motnje govora, glasovni posnetki, natančnost branja, čustva, MFCC