Objective: In this study, we aimed to investigate voice alterations in PD patients, using machine-learning techniques to extrapolate potential useful biomarkers that might help clinicians in the diagnosis of PD and disease progression. Introduction: Parkinson’s disease (PD) is a neurodegenerative disorder characterized by loss of dopaminergic neurons in the substantia nigra and accumulation of abnormally folded α-synuclein. The diagnosis is clinical and based on the presence of bradykinesia associated to rigidity and/or rest tremor. Phonation is a complex activity often altered in early stages of PD. Evidence shows that 70-90% of Parkinson’s disease patients present with hypokinetic dysarthria, a variable impairment of vocal emission characterized by hypophonia, mono-loudness, monopitch, harsh voice, and other alterations that encompass phonation, articulation, and prosody. These changes often manifest in the early stages of PD and worsen over the course of the disease. Instrumental studies found them related to asymmetric rigidity of the intrinsic laryngeal muscles and incomplete glottic closure due to vocal folds hypokinesis and bowing. Materials and Methods: We conducted a clinical study at the IRCSS Fondazione Mondino in Pavia, investigating 40 patients affected by PD (mean age: 62.9 10.2) and 28 age-matched healthy subjects (mean age: 59,9 6,9). We clinically evaluated voices using specific subitems of the UPDRS-III and the Voice Handicap Index. Voice samples related to the sustained emission of the vowel /a/ were recorded with a smartphone and underwent machine learning analysis. After having extracted and selected the 20 most performing features, which were mostly related to frequency, Random Forest classifier proved to be the most accurate algorithm in discriminating PD from HC, with a sensitivity of 89,7% and a specificity of 82,8%. Discussion: Results confirmed that several voice parameters are indeed abnormal in Parkinson’s disease. The limitations that we found were mostly related to the low sample size and to the impact of dopaminergic treatment on voice characterization in the various stages of the disease. In the future, the study will be further implemented in order to analyze all the available vocal samples and to correlate them with the clinical scores. Conclusions: In conclusion, machine learning analysis of voice samples allows to effectively identify Parkinson’s disease patients, thus highlighting the potential of voice alterations as new PD biomarkers.
Obiettivo: Questo studio si pone l’obiettivo di investigare le alterazioni vocali nei pazienti affetti da PD, utilizzando tecniche di machine learning per estrapolare potenziali biomarkers che potrebbero aiutare nella diagnosi del PD e nella progressione della malattia. Introduzione: Il morbo di Parkinson (PD) è una malattia neurodegenerativa caratterizzata dalla perdita di neuroni dopaminergici nella sostanza nigra e dall'accumulo di aggregati patologici di α-sinucleina. La diagnosi è clinica e si basa sulla presenza di bradicinesia associata a rigidità e/o tremore a riposo. La produzione della voce è un'attività complessa spesso alterata nelle prime fasi del PD. Evidenze mostrano che il 70-90% dei pazienti affetti da malattia di Parkinson presenta disartria ipocinetica, un'alterazione variabile dell'emissione vocale caratterizzata da ipofonia, monovolume, monotonicità, voce aspra e altre alterazioni che coinvolgono fonazione, articolazione e prosodia. Questi cambiamenti si manifestano spesso nelle prime fasi del PD e peggiorano nel corso della malattia. Studi strumentali li hanno correlati a rigidità asimmetrica dei muscoli intrinseci della laringe e a chiusura incompleta delle pliche vocali a causa dell'ipocinesia e dell'arcuazione delle corde vocali. Materiali e metodi: Abbiamo condotto uno studio clinico presso l'IRCSS Fondazione Mondino di Pavia, coinvolgendo 40 pazienti affetti da PD (età media: 62,9 c 10,2) e 28 soggetti sani appartenenti allo stesso gruppo di età (età media: 59,9 6,9). Abbiamo valutato clinicamente le voci utilizzando specifici subitems dell'UPDRS-III e del Voice Handicap Index. Sono stati registrati campioni vocali relativi all'emissione sostenuta della vocale /a/ tramite uno smartphone e sono stati sottoposti ad analisi di machine learning. Dopo aver estratto e selezionato le 20 caratteristiche più performanti, principalmente legate alla frequenza, il classificatore Random Forest si è dimostrato l'algoritmo più accurato nel discriminare il PD dagli HC, con una sensibilità dell'89,7% e una specificità dell'82,8%. Discussione: I risultati hanno confermato che diversi parametri vocali sono effettivamente anomali nella malattia di Parkinson. I limiti che abbiamo riscontrato sono stati principalmente legati alla ridotta dimensione del campione e all'impatto del trattamento dopaminergico sulla caratterizzazione vocale nelle diverse fasi della malattia. In futuro, lo studio verrà ulteriormente implementato al fine di analizzare tutti i campioni vocali disponibili e correlarli ai punteggi clinici. Conclusioni: In conclusione, l'analisi tramite machine learning dei campioni vocali consente di identificare in modo efficace i pazienti affetti da malattia di Parkinson, evidenziando quindi il potenziale delle alterazioni vocali come nuovi biomarkers della malattia di Parkinson.
Analysis of voice in Parkinson's disease as a potential biomarker: a Machine Learning study
BARILLI, ANNA
2022/2023
Abstract
Objective: In this study, we aimed to investigate voice alterations in PD patients, using machine-learning techniques to extrapolate potential useful biomarkers that might help clinicians in the diagnosis of PD and disease progression. Introduction: Parkinson’s disease (PD) is a neurodegenerative disorder characterized by loss of dopaminergic neurons in the substantia nigra and accumulation of abnormally folded α-synuclein. The diagnosis is clinical and based on the presence of bradykinesia associated to rigidity and/or rest tremor. Phonation is a complex activity often altered in early stages of PD. Evidence shows that 70-90% of Parkinson’s disease patients present with hypokinetic dysarthria, a variable impairment of vocal emission characterized by hypophonia, mono-loudness, monopitch, harsh voice, and other alterations that encompass phonation, articulation, and prosody. These changes often manifest in the early stages of PD and worsen over the course of the disease. Instrumental studies found them related to asymmetric rigidity of the intrinsic laryngeal muscles and incomplete glottic closure due to vocal folds hypokinesis and bowing. Materials and Methods: We conducted a clinical study at the IRCSS Fondazione Mondino in Pavia, investigating 40 patients affected by PD (mean age: 62.9 10.2) and 28 age-matched healthy subjects (mean age: 59,9 6,9). We clinically evaluated voices using specific subitems of the UPDRS-III and the Voice Handicap Index. Voice samples related to the sustained emission of the vowel /a/ were recorded with a smartphone and underwent machine learning analysis. After having extracted and selected the 20 most performing features, which were mostly related to frequency, Random Forest classifier proved to be the most accurate algorithm in discriminating PD from HC, with a sensitivity of 89,7% and a specificity of 82,8%. Discussion: Results confirmed that several voice parameters are indeed abnormal in Parkinson’s disease. The limitations that we found were mostly related to the low sample size and to the impact of dopaminergic treatment on voice characterization in the various stages of the disease. In the future, the study will be further implemented in order to analyze all the available vocal samples and to correlate them with the clinical scores. Conclusions: In conclusion, machine learning analysis of voice samples allows to effectively identify Parkinson’s disease patients, thus highlighting the potential of voice alterations as new PD biomarkers.È consentito all'utente scaricare e condividere i documenti disponibili a testo pieno in UNITESI UNIPV nel rispetto della licenza Creative Commons del tipo CC BY NC ND.
Per maggiori informazioni e per verifiche sull'eventuale disponibilità del file scrivere a: unitesi@unipv.it.
https://hdl.handle.net/20.500.14239/16220