Abstract
Cross-session EEG mental workload classifiers degrade severely when applied to new recording sessions from the same individual. A key but overlooked cause is the model selection criterion: standard within-session validation rewards checkpoints that exploit session-specific noise, systematically selecting against cross-session generalisation. This article describes Target-Session Early Stopping (TSES), a method that replaces the within-session validation set used for early stopping with a small held-out set of 50 labelled epochs from the target session. TSES requires no gradient updates on target-session data, no architectural changes, and no additional hyperparameter tuning. It improves binary cross-session accuracy on all 14 directional transfer pairs tested across three EEG subsets spanning two drift regimes. Combined with a two-sample Kolmogorov-Smirnov drift screen on the same 50 epochs, the complete pre-deployment protocol requires approximately four minutes of dedicated target-session recording. • TSES improved binary cross-session accuracy on all 14 directional transfer pairs tested across three EEG subsets spanning both high-drift (100% feature shift) and low-drift (52% feature shift) conditions. • Target-session early stopping is more data-efficient than calibration fine-tuning, which requires >100 labelled target epochs before showing any benefit under high drift. • Combined with a Kolmogorov-Smirnov drift screen on the same 50 epochs, the complete pre-deployment protocol requires approximately four minutes of dedicated target-session recording.