Document resource
Introduction Approximately 90% of PD patients experience hypokinetic dysarthria, which worsens as the disease progresses making it challenging to understanding the patients‘ speech. The application of artificial intelligence to the classification of on-off speech patterns in Parkinson’s disease offers a transformative approach to managing motor fluctuations.By leveraging advanced machine learning algorithms, we developed models that analyze subtle variations in speech characteristics to distinguish between ‘on’ periods-when medication improves motor control-and ‘off’ periods, marked by diminished control and speech impairment. Features such as rhythm, prosody, articulation clarity, and voice intensity effectively capture nuanced transitions between on and off states.Methods A CreateML VGG framework was trained on audio datasets of 200 Parkinson’s patients, split into training, validation, and testing subsets in an 80:10:10 ratio. Specifically, 160 patients (80%) were allocated to the training set, 20 patients (10%) to the validation set, and 20 patients (10%) to the testing set.Results The model achieved a training accuracy of 93%, validation accuracy of 62%, and testing accuracy of 75%. For Class 1 (On), the model demonstrated a precision of 78%, a recall of 70%, and an F1 score of 0.74, while for Class 2 (Off), the precision was 73%, the recall was 80%, and the F1 score was 0.76.Discussion Despite promising results, descrepancies in AI performance highlight opportunties for optimisation, offering valuable insights for the clinicians to improve PD management.jdavids@ic.ac.uk