Large-scale Clustering of People Diagnosed with Parkinson's Disease using Acoustic Analysis of Sustained Vowels: Findings in the Parkinson's Voice Initiative Study

The heterogeneity of symptoms in Parkinson's Disease (PD) has motivated investigating PD subtypes using cluster analysis techniques. Previous studies investigating PD clustering have typically focused on symptoms assessed using standardized clinical evaluations and patient reported outcome measures. Here, we explore PD subtype delineation using speech signals. We used data from the recently concluded Parkinson's Voice Initiative (PVI) study where sustained vowels were solicited and collected under non-controlled acoustic conditions. We acoustically characterized 2097 sustained vowel /a/ recordings from 1138 PD participants using 307 dysphonia measures which had previously been successfully used in applications including differentiating healthy controls from PD participants, and matching speech dysphonia to the standard PD clinical metric quantifying symptom severity. We applied unsupervised feature selection to obtain a concise subset of the originally computed dysphonia measures and explored hierarchical clustering combined with 2D-data projections using t-distributed stochastic neighbor embedding to facilitate visual exploration of PD subgroups. We computed four main clusters which provide tentative insights into different dominating speech-associated pathologies. Collectively, these findings provide new insights into the nature of PD towards exploring speech-PD data-driven subtyping.