Related Experiment Video
Updated: May 12, 2026
05:56
Objectification of Tongue Diagnosis in Traditional Medicine, Data Analysis, and Study Application
Published on: April 14, 2023
Development of an Interpretable Deep Learning-Based Segmentation Algorithm for Automated Assessment of Oral
Jan Melechovsky1, Michal Novotny2, Tereza Tykalova2
1Information Systems Technology and Design, Singapore University of Technology and Design, Singapore.
Summary
A new deep learning algorithm accurately segments speech in patients with neurological disorders. This automated tool shows promise for analyzing oral diadochokinesis across various conditions and severities.
Area of Science:
- Speech-language pathology
- Computational neuroscience
- Artificial intelligence in medicine
Background:
- Oral diadochokinesis (ODK) analysis is crucial for diagnosing and monitoring neurological diseases affecting speech.
- Current segmentation methods for ODK can be labor-intensive and lack universal applicability across diverse patient populations.
- Automated segmentation algorithms are needed to improve the efficiency and robustness of ODK assessment.
Purpose of the Study:
- To develop a universal, fully automated segmentation algorithm for robust ODK analysis.
- To evaluate the algorithm's performance across various neurological diseases, dysarthria types, and severities.
- To compare the deep learning algorithm's performance against traditional signal processing methods.
Main Methods:
- Collected sequential motion rate recordings from 231 subjects (80 healthy, 151 with neurological diseases).
- Developed a convolutional neural network (CNN) based algorithm with rule-based postprocessing for automatic segmentation.
- Evaluated algorithm performance using F1 scores and temporal accuracy, comparing it to a signal processing-based approach.
Main Results:
- The deep learning algorithm achieved a 99.1% F1 score for syllable segmentation, outperforming the 97.7% F1 score of the signal processing method.
- Average accuracy for temporal phoneme detection was 92.0% within a 10-ms window.
- Performance was most affected by dysarthria severity (94.7% mild, 91.0% moderate, 83.1% severe), with less impact from disease or dysarthria type.
Conclusions:
- The proposed deep learning algorithm reliably segments syllables and phonemes in ODK tasks across diverse neurological conditions.
- Deep learning-based segmentation offers superior performance compared to traditional signal processing methods for ODK assessment.
- This automated approach has the potential to significantly advance the analysis of speech motor control in neurological populations.
