Detecting Depression Using an Ensemble Logistic Regression Model Based on Multiple Speech Features

Area of Science:

• Psychiatry
• Computational Linguistics
• Speech Science

Background:

• Early depression intervention is crucial for reducing disease burden.
• Current diagnostic methods for depression have limitations.
• Automatic speech analysis offers a potential avenue for objective depression assessment.

Purpose of the Study:

• To investigate the effectiveness of speech features in classifying depression.
• To develop and evaluate an ensemble model for detecting depression from speech.
• To assess the performance of the proposed model in a native Chinese population.

Main Methods:

• A sample of 170 native Chinese subjects (85 healthy, 85 depressed) was analyzed.
• Prosodic, spectral, and glottal speech features were extracted.
• An ensemble logistic regression model for detecting depression (ELRDD) was proposed and tested, using logistic regression as the base classifier.

Main Results:

• The ELRDD model outperformed other compared classifiers in depression recognition.
• A derived technique, ELRDD-E, achieved high accuracy: 75.00% for females and 81.82% for males.
• ELRDD-E demonstrated favorable sensitivity/specificity ratios: 79.25%/70.59% for females and 78.13%/85.29% for males.

Conclusions:

• Automatic speech analysis using ensemble logistic regression is a viable method for depression detection.
• The ELRDD-E technique shows promise for accurate and sensitive identification of depression.
• Speech-based depression classification can aid in early intervention and reduce the impact of the disease.