Calibrated Drift Monitoring for EEG Representations

Research question

Differences between EEG recordings include variation already present among acceptable sessions. This project asks how a monitoring rule can use that reference variation to identify atypical recordings in the representations of a frozen EEG model.

Current approach

The framework, called WARD-MMD, combines kernel-based discrepancy scores with reference-session calibration. Calibration and monitoring use matched window lengths, and the analysis examines session-to-session variation, decision uncertainty, and sensitivity to different drift directions.

Current evidence

A reference implementation and simulation studies are complete. The validation targets have mixed outcomes across conditions. Some drift directions remain poorly detected, and calibration depends on the available reference sessions and model assumptions. The present evidence is computational and simulated; real EEG results have not been obtained.

The intended output is evidence about whether a recording is atypical relative to acceptable reference sessions. Its relationship to classification performance requires a separate empirical evaluation.