Using i-vectors from voice features to identify major depressive disorder

Area of Science:

• Computational linguistics
• Clinical psychology
• Machine learning

Background:

• Machine learning for major depressive disorder (MDD) diagnosis using acoustic features lacks large-scale validation.
• Existing methods require more robust evidence from clinical trials and diverse samples.

Purpose of the Study:

• To evaluate the effectiveness of the i-vector method for diagnosing major depressive disorder (MDD) in a large sample of women.
• To assess the robustness of the i-vector method across different speech durations.
• To provide an acoustic interpretation of i-vector features in MDD.

Main Methods:

• Collected speech utterances from 785 women with MDD and 1,023 healthy controls.
• Extracted Mel-frequency cepstral coefficient (MFCC) features and MFCC i-vectors.
• Compared binary logistic regression performance between i-vectors and MFCCs, testing robustness with varying utterance durations.

Main Results:

• i-vectors enhanced the area under the curve (AUC) by 7-14% compared to MFCC features.
• Classification accuracy stabilized for utterance durations exceeding 40 seconds.
• i-vectors showed consistent correlations (positive or negative) with MFCC feature variations.

Conclusions:

• i-vectors significantly improve MDD diagnostic accuracy (up to 14% AUC increase) in a large clinical sample.
• The i-vector system demonstrates robustness for utterance durations over 40 seconds.
• This study establishes a foundation for the clinical use of voice analysis in MDD diagnosis.