Automated Audio Analysis of Dysarthria in Huntington’s disease (Audio-HD)
The diagnosis of Huntington’s disease (HD) is based primarily on the motor exam component of the Unified Huntington’s Disease Rating Scale (UHDRSTM). The UHDRSTM grades dysarthria from 0-4 (normal to anarthria), however automated speech feature detection could capture subtle changes that might serve as early disease biomarkers. Using an exploratory approach, our goals were 1) identify features of speech that differentiated HD from controls, 2) use key speech features to train a classifier model and 3) compare these results to clinical assessments.
Audio-HD protocol using Canary Speech has significant potential as an efficient and sensitive tool for identifying speech biomarkers in HD. Further training of the model and longitudinal follow up will be critical for determining the overall utility of this approach. HD average UHDRSTM total motor score was 16.1 (SD=15.72) and CAG repeat length 43.1 (SD=3.62). Welch’s t-test (95% CI) identified >1000 features of speech that differentiated HD from HC. The 20 most significant features were entered into a Random Forest classifier to generate a trained model. Classifier accuracy (HD vs. HC) was 96% (91% sensitivity, 98% specificity) and sensitivity was 96% for detecting dysarthria in HD. Interestingly, the model also correctly labeled 40% of HD cases lacking clinical dysarthria, suggesting it was capturing speech changes not apparent to the examiner.