Comparing Post-Hoc Explainable AI Methods for Interpreting Black-Box EEG Models in Depression Detection

Researchers compared five post-hoc explainability methods for interpreting deep learning models trained to detect Major Depressive Disorder from EEG data. While different attribution approaches showed partially overlapping patterns emphasizing frontal and temporal brain regions, the study reveals methodological assumptions significantly influence interpretability results, cautioning against treating findings as definitive clinical biomarkers.

This research addresses a critical challenge in medical AI: making black-box deep learning models interpretable for clinical applications. The study evaluates how different explainability frameworks—Shapley-based, gradient-based, and perturbation-based methods—interpret an InceptionTime architecture trained for depression detection from EEG signals. The investigators applied rigorous methodology using subject-level stratified cross-validation to ensure robust findings.

The convergence of some attribution patterns across methods suggests genuine signal in identifying relevant brain regions, particularly in frontal, temporal, and posterior areas associated with depression in neuroimaging literature. However, the substantial divergence between DeepSHAP and other approaches highlights a fundamental problem: explainability methods may reveal more about their own mathematical assumptions than ground truth about model decision-making. This distinction matters considerably for psychiatric applications where clinical stakes are high.

For AI developers and healthcare institutions deploying EEG-based diagnostic tools, this research demonstrates that no single explainability method provides complete transparency. Organizations cannot rely on one interpretation technique to validate model decisions for patient care. The findings suggest practitioners should employ multiple complementary methods and treat results as exploratory rather than confirmatory evidence.

Looking forward, the psychiatric AI field needs standardized protocols for explainability validation in clinical contexts. Future work should integrate explainability analysis with prospective clinical validation and neurophysiological grounding. This research underscores the ongoing maturation required before deep learning models achieve trustworthy deployment in mental health diagnostics, where interpretability remains non-negotiable for regulatory approval and clinical adoption.

• →Multiple post-hoc explainability methods show partially overlapping but methodologically-dependent attribution patterns for EEG-based depression detection models.
• →Gradient-based and perturbation-based approaches demonstrate substantial agreement, while DeepSHAP produces distinctly different interpretations highlighting algorithmic assumptions.
• →Brain regions identified (frontal, temporal, posterior) align with prior depression neuroimaging studies, but findings remain exploratory rather than definitive biomarkers.
• →No single explainability method provides complete transparency, requiring practitioners to employ multiple complementary approaches for robust interpretation.
• →Clinical deployment of AI-driven psychiatric diagnostics requires standardized explainability validation protocols beyond current post-hoc analysis techniques.