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

No single explanation method tells the whole story: The study found that different explainability techniques (DeepSHAP, Integrated Gradients, GradCAM, and others) produced "partially overlapping" rather than identical results, meaning each method captures different aspects of how the AI model makes decisions. This suggests that relying on just one explanation method could give an incomplete or misleading picture of why an AI system diagnosed depression.

Methodological choices significantly shape AI explanations: The research demonstrated that the technical assumptions built into different explanation methods substantially influence their outputs, with DeepSHAP producing "comparatively distinct attribution distributions" compared to other approaches. This is important for responsible AI because it means the explanations we get depend partly on which tool we choose, not just on what the model actually learned.

Brain regions show consistent patterns across methods: Despite methodological differences, the explainability analysis consistently highlighted frontal, temporal, and posterior EEG regions—particularly in the right hemisphere—as important for depression detection. This convergence across multiple methods provides some reassurance that these brain regions genuinely matter to the model's decisions.

Transparency in mental health AI requires multiple verification approaches: For high-stakes applications like depression detection, this research suggests that clinicians and developers should use multiple explanation methods together rather than trusting a single approach, creating a more robust understanding of how the AI system reaches its conclusions.