Multiangle Correlation Feature Extraction and Disease Prediction Model Construction for Patients With Post-Stroke...

Area of Science:

• Neurology
• Speech-Language Pathology
• Artificial Intelligence

Background:

• Clinical diagnosis of motor dysarthria in stroke patients is subjective and often overlooks psychological factors.
• Emotional and psychological states significantly impact dysarthria progression and intelligibility.
• Existing diagnostic methods lack comprehensive analysis of multimodal emotional and pathological expressions.

Purpose of the Study:

• To investigate the correlation between emotional expression, psychological state, facial expressions, and motor dysarthria severity.
• To develop an objective dysarthria prediction model incorporating multimodal data.
• To establish a novel Chinese multimodal emotional pathology expression database (THE-POSSD).

Main Methods:

• Collected acoustic, glottal, and facial data from stroke patients and healthy controls under emotional stimuli.
• Utilized grey correlation theory and a grey relational analysis-deep belief network (GRA-DBN) for model construction.
• Optimized the GRA-DBN model using Principal Component Analysis and Variance Inflation Factor.

Main Results:

• Identified 154 significant correlation features between emotional/psychological data and dysarthria severity.
• Achieved high prediction accuracy with an adjusted R² of 0.85 for the GRA-DBN model.
• The optimized model demonstrated a superior adjusted R² of 0.92, highlighting improved predictive power.

Conclusions:

• The developed model offers a more objective and precise method for diagnosing motor dysarthria in post-stroke patients.
• Incorporating multimodal emotional and psychological features enhances the accuracy of dysarthria prediction.
• This research provides a valuable framework and dataset for understanding the interplay between mental state and speech disorders.