Identifying Parkinson's Disease and Atypical Parkinsonism using acoustic representations of core vowels with Machine Learning models

Research overview

The text below is the source article’s original abstract under a verified reuse license. cc by

Abstract Background Differentiating Parkinson's disease (PD) from Atypical Parkinsonian Syndromes (APS) is challenging due to overlapping speech impairments. This study addresses the gap in utilizing speech biomarkers for early differentiation in Mandarin speakers. Objective To identify distinct acoustic phenotypes in PD and APS and develop machine learning models for differential diagnosis and severity monitoring. Methods Speech samples from sustained vowels (nine Mandarin core vowels) and diadochokinesis tasks were collected from 34 PD, 25 APS (13 multiple system atrophy, 12 progressive supranuclear palsy), and 30 healthy controls. Acoustic parameters including Fundamental Frequency, Jitter, Shimmer, Noise-to-Harmonics Ratio, Vowel Space Area, Vowel Articulation Index, and diadochokinesis rate were extracted. Principal component analysis followed by linear discriminant analysis was used for classification. An XGBoost model predicted dysarthria severity based on UPDRS‑III subscores. Results APS patients showed more severe impairments (reduced Fundamental Frequency, elevated Jitter and Shimmer, decreased sound pressure level, increased noise-to-harmonics ratio for /ər/) compared to PD and healthy controls. PD patients primarily showed milder articulatory deficits (reduced Vowel Space Area and Vowel Articulation Index) and slower diadochokinesis rate. Cognitive impairment correlated with worse vocal instability in both groups. The linear discriminant analysis model achieved a mean accuracy of 0.83 ± 0.09 (10-fold cross-validation) and 0.93 on an independent test set. For PD vs. APS differentiation, area under the curve was 0.8389 ± 0.1693. XGBoost predicted dysarthria severity with area under the curve of 0.75 (PD) and 0.85 (APS). Conclusion Mandarin speech biomarkers combined with machine learning enable accurate differentiation of PD, APS, and healthy controls, as a potential non-invasive adjunctive tool for early diagnosis and severity monitoring.

Related topics