Withdrawn paper proposed interpretable AI model for speech-based depression screening

A machine learning approach to automated depression detection has aimed to bridge the gap between technical accuracy and clinical usability by making AI models more transparent.

The challenge#

Depression is one of the most prevalent mental health conditions, and early detection significantly improves treatment outcomes. Automated screening tools powered by machine learning offer promise for identifying at-risk individuals. However, two major obstacles have prevented widespread adoption in clinical settings: unreliable predictions that arise from analyzing speech in small chunks, and the inability of these systems to explain why they make their decisions—a critical requirement for healthcare professionals who must understand and trust the systems they use.

The proposed solution#

The team developed a bi-modal speech-level transformer, a deep learning model that simultaneously processes both audio and text from patient speech samples. Rather than breaking recordings into fragments and analyzing each piece independently—an approach that introduces errors—the model examines entire speech samples at once. This preserves context and reduces noise.

Crucially, the researchers added a hierarchical interpretation layer that traces the model's reasoning through multiple levels. By analyzing attention patterns across all network layers and using gradient-weighted attention maps, the system produces explanations at three scales: which sentences in a speech matter most for depression detection, and within those sentences, which specific words and audio frequencies are most influential.

Performance gains#

Testing demonstrated substantial improvements over segment-based approaches:

• Precision: 0.854 vs. 0.732
• Recall: 0.947 vs. 0.808
• F1 score: 0.897 vs. 0.768

Clinical impact#

The interpretability layer allows clinicians to verify predictions by seeing which parts of a patient's speech contributed to a depression diagnosis. This transparency is essential for moving automated mental health screening from research settings into actual clinical practice, where trust and explainability are non-negotiable.

Note: This paper was withdrawn by the authors and is no longer available in full on arXiv.