Ashkan Vaziri1, Adonay Nunes1, Meghan Pawlik2, Ram Kinker Mishra1, E. Ray Dorsey3, Jamie L. Adams3
1BioSensics LLC, 2Center for Health + Technology, University of Rochester Medical Center, 3Department of Neurology, University of Rochester
Objective:
To explore the potential of speech data as a digital measure in assessing the progression of Huntington disease (HD) and its clinical implications.
Background:
Speech changes are an early symptom of Huntington disease (HD) and may occur prior to other motor and cognitive symptoms. Traditional HD assessments utilize clinician-rated outcome measures. These can be limited by observer variability and episodic administration. Digital assessment of speech can provide a quantitative evaluation apt for frequent and remote administration.
Design/Methods:
Recordings of 33 participants (15 HD, 7 prodromal HD, and 11 control) were collected while reading a passage speech task. Biodigit Speech was used to analyze the recording and extract speech features such as pausing, pitch, and intelligibility. The features were compared between HD, prodromal HD and controls, and also correlated with the Unified Huntington Disease Rating Scale (UHDRS) subscales. Machine learning models based on random forest algorithms were employed to predict clinical status and scores from speech features.
Results:
Observable differences in pausing, intelligibility, and accuracy for passage reading among the HD, prodromal HD, and control groups were identified. The random forest classifier correctly predicted the three clinical status from speech tasks with an accuracy rate of 73%. Random forest regression predicted UHDRS clinical scores from speech features with explained variance between 14% and 54%.
Conclusions:
Speech data presents a great potential as a key digital measure in monitoring neurodegenerative progression. It can facilitate remote, regular assessments of the disease in prodromal HD and HD stages. Through the extracted speech features, random forest machine learning models adeptly predicted clinical status and disease severity. This suggests that speech measurements might be a sensitive marker for determining clinical onset and progression in upcoming clinical trials.