Deconstructing demographic bias in speech-based machine learning models for digital health.

Area of Science:

• Digital Health
• Machine Learning
• Computational Linguistics

Background:

• Machine learning (ML) shows promise for digital healthcare but faces criticism for perpetuating demographic disparities.
• Speech-based ML algorithms are increasingly used for behavioral and mental health outcomes.

Purpose of the Study:

• To explore gender and race bias in speech-based ML algorithms for detecting mental health conditions.
• To investigate sources of bias in training data and ML decisions.
• To evaluate methods for bias reduction in ML models.

Main Methods:

• Examined acoustic features and labels for demographic bias.
• Investigated bias reduction techniques using less demographic-informative features.
• Employed adversarial methods to transform feature spaces, diminishing demographic information while preserving mental health state information.

Main Results:

• Found statistically significant differences in acoustic features and labels across gender and race groups.
• Observed differential ML performance across demographic groups, with bias partially preserved in decisions.
• Results were mixed regarding model accuracy for sensitive groups in anxiety and depression detection.

Conclusions:

• Highlights the necessity for careful ML model design in digital healthcare.
• Emphasizes the importance of maintaining data integrity and ensuring equitable performance across diverse populations.
• Underscores the need for bias mitigation strategies in speech-based ML for mental health.