Mitigating Gender Bias in Depression Detection via Counterfactual Inference

TL;DR

The paper tackles gender bias in audio-based depression detection caused by imbalanced data. It introduces a counterfactual debiasing framework grounded in causal inference to isolate and remove the direct effect of gender on predictions, leveraging two neural components for gender and acoustic fusion and using counterfactual inference at test time. Empirical results on DAIC-WOZ with two acoustic backbones show reduced bias (lower EA and DI) and improved overall performance, including substantial gains for male depression detection. This approach offers a principled, interpretable method to disentangle gender from true pathological vocal cues, with implications for fairer automated mental health assessment. The authors also outline future multimodal extensions to further mitigate bias across data modalities.

Abstract

Audio-based depression detection models have demonstrated promising performance but often suffer from gender bias due to imbalanced training data. Epidemiological statistics show a higher prevalence of depression in females, leading models to learn spurious correlations between gender and depression. Consequently, models tend to over-diagnose female patients while underperforming on male patients, raising significant fairness concerns. To address this, we propose a novel Counterfactual Debiasing Framework grounded in causal inference. We construct a causal graph to model the decision-making process and identify gender bias as the direct causal effect of gender on the prediction. During inference, we employ counterfactual inference to estimate and subtract this direct effect, ensuring the model relies primarily on authentic acoustic pathological features. Extensive experiments on the DAIC-WOZ dataset using two advanced acoustic backbones demonstrate that our framework not only significantly reduces gender bias but also improves overall detection performance compared to existing debiasing strategies.

Figures (5)

• Figure 1: (a) Example of causal graph. (b) Example of factual scenario. (c) Example of counterfactual scenario. White nodes are at the state under $X = x$, while gray nodes are at the state under $X = \bar{x}$.
• Figure 2: (a) Csusal graph for depression detection. (b) Comparison between factual and counterfactual scenarios in depression detection. White nodes are at the state under $G = g$ and $C = c$, while gray nodes are at the state under $G = \bar{g}$ and $C = \bar{c}$.
• Figure 3: Architecture of Proposed Counterfactual Debiasing Framework. The framework consists of a training stage that jointly optimizes a gender-only model $\mathcal{M}_G$ and a fusion model $\mathcal{M}_F$, and a testing stage where the bias effect is subtracted from the total effect via a counterfactual path.
• Figure 4: Architecture of STA-based backbone model. STA network has two branches, including a 2D CNN for spatial features and an LSTM for temporal dynamics. Their outputs fuse via Attention Pooling to highlight key speech frames. EEP uses eigendecomposition to capture temporal correlations and identify principal components for weighted sequence aggregation.
• Figure 5: Architecture of the NetVLAD-based backbone model. NetVLAD module assigns local acoustic features to learnable cluster centers (VLAD Core) and sums weighted residuals to produce fixed-dimensional vectors. GRU network uses internal gating mechanism to selectively update its hidden state at each step, with the final state summarizing the sequence.