Interpretable Transformer-Based Depression Detection in Low-Resource Malay Speech Using SHAP-Based Feature Analysis

Abstract

Speech-based depression detection has emerged as a scalable and non-invasive approach for mental health assessment; however, the clinical relevance of self-supervised speech representations remains insufficiently understood, particularly in low-resource languages. This study investigates the interpretability of feature representations for depression detection in low-resource Malay clinical speech using a Transformer-based classification framework. Three feature configurations are systematically evaluated: i) 27-dimensional handcrafted acoustic biomarkers, ii) 768-dimensional frozen Wav2Vec2 embeddings, and iii) a 795-dimensional hybrid fusion of handcrafted and self-supervised representations. Experiments are conducted on a Malay Clinical Speech dataset and the DAIC-WOZ benchmark to assess performance across two independent clinical speech datasets under participant-independent five-fold cross-validation. Experimental results demonstrate that handcrafted acoustic representations consistently outperform both Wav2Vec2 and hybrid representations, achieving F1-scores of $0.748~\pm ~0.067$ on the Malay dataset and $0.599~\pm ~0.090$ on DAIC-WOZ. In contrast, hybrid representations exhibit reduced precision and increased performance variability, indicating limited robustness in low-resource settings. To interpret the observed behavior, SHAP-based analysis using GradientExplainer is employed to quantify feature importance, attribution stability, and cross-fold consistency. The analysis reveals that low-dimensional spectral acoustic features account for the majority of model attribution (52.9%) with high cross-fold consistency, whereas Wav2Vec2 embeddings exhibit more dispersed and less stable attribution patterns despite their higher representational capacity. These findings indicate that high-dimensional self-supervised speech embeddings do not necessarily improve interpretability or predictive performance under low-resource clinical speech conditions. Overall, this study provides a comparative representation-level evaluation and a stability-aware SHAP analysis that offer practical insights into selecting clinically meaningful speech representations for interpretable depression detection in low-resource settings.