Dysarthria detection based on a deep learning model with a clinically-interpretable layer.

Area of Science:

• Speech-language pathology
• Machine learning in healthcare
• Computational linguistics

Background:

• Deep neural networks (DNNs) show promise for classifying dysarthric speech.
• Current DNNs lack clinical interpretability, limiting their practical application in diagnosing speech impairments.
• Interpretable models are crucial for understanding the underlying speech production deficits in dysarthria.

Purpose of the Study:

• To develop a clinically interpretable deep learning model for classifying dysarthric speakers.
• To jointly learn speech classification and interpretable features related to dysarthria subtypes.
• To enable a flexible trade-off between classification accuracy and the interpretability of speech deficit patterns.

Main Methods:

• A deep neural network with a bottleneck layer was designed.
• The model was trained to simultaneously predict a classification label and four clinically relevant features.
• Shapley additive explanation (SHAP) was employed for model interpretability analysis.
• The model was evaluated on two distinct subtypes of dysarthria.

Main Results:

• The proposed method achieved a balance between classification accuracy and the identification of interpretable speech deficit patterns.
• The learned representations were found to be consistent with the known disturbances characterizing the studied dysarthria subtypes.
• The model demonstrated the ability to uncover clinically meaningful patterns in dysarthric speech.

Conclusions:

• This interpretable DNN approach offers enhanced clinical value for dysarthria classification.
• The method provides insights into the specific speech characteristics associated with different dysarthria subtypes.
• Future research can leverage this framework for more precise diagnosis and targeted speech therapy interventions.