Development of an Interpretable Deep Learning-Based Segmentation Algorithm for Automated Assessment of Oral...

We aimed to develop a universal, fully automated segmentation algorithm that allows robust analysis of oral diadochokinesis across various neurological diseases, dysarthria types, and dysarthria severities.

Recordings of sequential motion rates were collected from 231 subjects, including 80 healthy controls and 151 patients with neurological diseases such as amyotrophic lateral sclerosis, essential tremor, Huntington's disease, multiple sclerosis, multiple system atrophy, Parkinson's disease, progressive supranuclear palsy, and cerebellar ataxia. A robust automatic segmentation algorithm utilizing convolutional neural networks and rule-based postprocessing was developed and evaluated across disease type, dysarthria type, and dysarthria severity. The performance of the developed artificial intelligence-based algorithm was compared with a traditional signal processing-based segmentation approach.

Our deep learning-based algorithm was able to correctly identify the position of individual syllables with a very high F1 score of 99.1%, compared to a signal processing-based approach with an F1 score of 97.7%. Using a 10-ms tolerance window, the deep learning-based algorithm achieved an average accuracy of 92.0% for the temporal detection of individual phoneme positions. Performance was strongly influenced by dysarthria severity, with accuracy reaching 94.7% in mild, 91.0% in moderate, and 83.1% in severe dysarthria. Disease and dysarthria type did not appear to have a substantial effect on algorithm performance.

Our proposed deep learning-based algorithm provides reliable segmentation of syllable and individual phoneme positions during oral diadochokinesis across various disease types, dysarthria types, and dysarthria severities. The deep learning-based segmentation approaches have the potential to outperform the traditional signal processing methods for assessing oral diadochokinesis.