Researchers propose transparent speech-based depression detection
New framework achieves 82% accuracy while ensuring fairness and interpretability...
A team of researchers led by Mariel Estevez and colleagues has developed a novel framework for speech-based depression detection that prioritizes transparency and fairness alongside performance. Published on arXiv (arXiv:2606.31730), this work addresses a critical gap in clinical AI adoption where opaque models and potential biases have limited real-world deployment.
The framework leverages explainable AI techniques (LIME and SHAP) to select interpretable acoustic features like MFCCs and eGeMAPS, combined with low-complexity machine learning models (Random Forest, SVM, and MLP). Through rigorous statistical validation and demographic fairness analyses, the team achieved state-of-the-art 82% test accuracy on the extended DAIC-WOZ dataset while mitigating spurious correlations that could lead to biased predictions. The methodology is designed to be adaptable to other binary classification tasks beyond depression detection.
- Achieved 82% accuracy on DAIC-WOZ dataset using explainable features (MFCCs, eGeMAPS)
- Combined XAI methods (LIME, SHAP) with low-complexity models (RF, SVM, MLP) for transparency
- Addressed demographic bias through fairness analyses and statistical validation
Why It Matters
Could enable ethical, clinically viable AI mental health tools with verifiable fairness guarantees.