Objective: To detect differences of speech in patients with PD due to a variant motor state, i.e. ON and OFF.

Background: Speech is a highly relevant “motor behavior“ that is poorly represented in clinical decision making. Understanding speech by advanced methods of data analytics could be instrumental for the evaluation of patients with PD (PwP), e.g. for detection of therapy response, or guiding personalized evaluation.

Methods: We assessed PwP at the Schön Klinik München in their practical ON and OFF conditions by standardized digital speech recordings with a Rode NT1-A microphone and a Steinberg UR22 MK2. The speech recording protocol consisted of isolated and connected speech. For speech analysis, we used the phone-attribute probabilistic features related to phonation and articulation [1, 2]. The quantized feature representations were used for building ON and OFF codebooks using the data of the full cohort as training data (Fig 1). Given the codebooks, the condition of a new (test) patient can be detected by comparing the binary codes. We conducted statistical analyses of phone-attributes that identify which attributes are most discriminative among ON and OFF. We investigated the correlation among different protocols to detect the one most suitable for personalized assessment.

Results: In total, we included 16 subjects (10 female, 6 male) with a mean age of 66 yrs (SD ±7) and a mean disease duration of 11y (SD±5). Average length of each recording session was about 4+4 minutes. The UPDRS motor score during ON was 21±11, during OFF 32±1. Mean UPDRS-Item III.1 score in ON was 1,27±1,10, in OFF 1,93±1,28. The classifier detected the motor state with 100% sensitivity and 100% specificity.

Conclusions: Phone-attribute probabilistic features obtained from deep neural networks are proven to be effective for analysis of impaired speech in PwP. The posture of phonation and articulation can be characterized in terms of binary phone attribute codes; unique codes are observed in ON and OFF motor states that enable unambiguous classification. Among the comprehensive set of linguistic phone attributes, a small number were highly varying in ON and OFF, whereas others demonstrated a steady behavior not affected by medication. Contrary to published opinion, our data argues convincingly for an important contribution of dopaminergic medication on speech [3]. PwP need to be evaluated individually; our analysis guides to a personalized clinical assessment procedure.

References: [1] Asaei, Afsaneh, Milos Cernak, Hervé Bourlard, and Dhananjay Ram. “Sparse Pronunciation Codes for Perceptual Phonetic Information Assessment.” In Proceeding of Workshop on Signal Processing with Adaptive Sparse Structured Representations (SPARS), 2017 Online: https://pdfs.semanticscholar.org/660e/df2270165ec00f9cde0e842202994a46d3e4.pdf. [2] Cernak, Milos, Afsaneh Asaei, and Hervé Bourlard. “On structured sparsity of phonological posteriors for linguistic parsing.” Speech Communication 84 (2016): 36-45. Online: https://arxiv.org/pdf/1601.05647.pdf. [3] Brabenec L, Mekyska J, Galaz Z, Rektorova I. Speech disorders in Parkinson’s disease: early diagnostics and effects of medication and brain stimulation. J Neural Transm. 2017;124(3):303–34.

« Back to 2018 International Congress

MDS Abstracts - https://www.mdsabstracts.org/abstract/distinguishing-on-from-off-motor-state-by-phone-attribute-codes-in-patients-with-parkinsons-disease/