Gender Bias in Depression Detection Using Audio Features

TL;DR

This paper investigates gender bias in depression detection using audio features, revealing dataset biases that skew performance metrics and proposing mitigation strategies based on fair machine learning principles.

Contribution

It identifies gender bias in the DAIC-WOZ dataset and demonstrates how to reduce bias effects using data re-distribution and raw audio features.

Findings

Gender bias inflates performance metrics.

Bias mitigation improves fairness in detection.

Raw audio features help reduce bias effects.

Abstract

Depression is a large-scale mental health problem and a challenging area for machine learning researchers in detection of depression. Datasets such as Distress Analysis Interview Corpus - Wizard of Oz (DAIC-WOZ) have been created to aid research in this area. However, on top of the challenges inherent in accurately detecting depression, biases in datasets may result in skewed classification performance. In this paper we examine gender bias in the DAIC-WOZ dataset. We show that gender biases in DAIC-WOZ can lead to an overreporting of performance. By different concepts from Fair Machine Learning, such as data re-distribution, and using raw audio features, we can mitigate against the harmful effects of bias.