A New Regression Model for Depression Severity Prediction Based on Correlation among Audio Features Using a Graph...

Area of Science:

• Computational linguistics
• Psychiatry
• Machine learning

Background:

• Depression severity can be predicted using audio features from patient voices.
• Existing deep learning models often assume audio features are independent, limiting predictive accuracy.
• Correlated audio features in depressed patients' voices offer a basis for improved characterization.

Purpose of the Study:

• To propose a novel deep learning regression model for predicting depression severity.
• To leverage the correlations among audio features, unlike previous independent-feature methods.
• To utilize graph convolutional networks for modeling feature interdependencies.

Main Methods:

• Developed a graph convolutional network (GCN) model.
• Trained the model using graph-structured data representing correlations among audio features.
• Evaluated the model on the DAIC-WOZ dataset for depression severity prediction.

Main Results:

• Achieved a root mean square error (RMSE) of 2.15 and a mean absolute error (MAE) of 1.25.
• Reported a symmetric mean absolute percentage error of 50.96%.
• Demonstrated superior performance in RMSE and MAE compared to state-of-the-art methods.

Conclusions:

• The proposed GCN model effectively utilizes correlated audio features for depression severity prediction.
• The model shows significant improvements over existing methods, indicating its potential.
• This approach offers a promising new tool for aiding in depression diagnosis.