Automatic speech analysis combined with machine learning reliably predicts the motor state in people with Parkinson’s disease

Abstract

It is still under debate whether levodopa treatment improves speech functions in Parkinson’s disease (PD). Therefore, speech functions of people with PD were compared in medication-OFF condition (withdrawal of PD medication for at least 12 h) and medication-ON condition (after receiving 200 mg of soluble levodopa). A total of 78 participants, including 51 males and 27 females, performed predefined standard speech tasks. Acoustic speech features were automatically extracted with the algorithm given by the Dysarthria Analyzer. Results suggest that acute levodopa intake improves phonatory-respiratory speech functions and speech planning abilities, while the articulatory system remains unaffected. Furthermore, the study provided preliminary evidence that speech function is able to predict the medication status in individuals with PD as the constructed speech-based biomarker score did not only correlate with established measures of (speech) motor impairment but could also differentiate between the medication OFF and ON status. A post-hoc machine learning model yielded similar results.