COunterfactual Reasoning for Temporal EXplanations: Plausible and Robust Explanations for EEG-Based Seizure Detection

TMLR Paper6783 Authors

Abstract: Identifying the drivers of change in time-sensitive domains like healthcare is critical for reliable decision-making, yet explanations must account for both temporal dynamics and structural complexity. While counterfactual explanations are well-studied for static data, existing methods often fail in dynamic, spatio-temporal settings, producing implausible or temporally inconsistent explanations. To address this, we introduce COunterfactual Reasoning for Temporal EXplanations (CORTEX), a search-based explainer for multivariate time series modeled as spatio-temporal graphs, tailored to seizure detection from EEG recordings. CORTEX generates temporally robust and plausible counterfactuals by retrieving relevant past instances and sieving them via structural dissimilarity, temporal distance, and instability. Evaluated on clinical seizure detection data, CORTEX outperforms state-of-the-art methods with a $2.73\times$ improvement in validity and $5.32\times$ in fidelity, and achieves zero implausibility, demonstrating consistency and practical relevance. By shifting the focus from mere validity to plausible and time-consistent explanations, CORTEX enables more reliable and controllable counterfactual explanations.

Submission Type: Regular submission (no more than 12 pages of main content)

Assigned Action Editor: ~Fabio_Stella1

Submission Number: 6783