Abstract Objective
Automated EEG-based epilepsy diagnosis has reported near-perfect accuracies for almost two decades on a benchmark dataset, yet virtually no system is used in routine care. We critically re-examined this translation gap by reproducing five widely cited AI models spanning statistical feature extraction, classical machine-learning and deep-learning paradigms, and assessed their ability to generalise from the benchmark dataset to a newly curated, clinically verified scalp-EEG cohort. 

Approach
All models were implemented as originally described and trained with 10-fold cross-validation on the benchmark dataset. External validation was performed on our independently curated dataset made publicly available, comprising 30 subjects (15 epilepsy, 15 healthy) recorded with 19-channel scalp EEG under standard clinical protocols. We further examined the influence of subject-level data leakage by contrasting performance when training/testing samples overlapped with those when completely independent patient partitions were enforced. Accuracy, sensitivity, specificity, and AUC of ROC (with 95% Confidence Intervals) were the primary metrics. 

Main results

When transferred unchanged to the external cohort, overall accuracy fell from >= 94% on the benchmark dataset to 42-53%, with sensitivities as low as 0.97% for the deep CNN and specificities dropping to 4% for time-frequency methods. Permitting subject overlap artificially elevated accuracy to 59-96%, whereas strict patient separation reduced it to 41-53% in 95% Confidence Interval. Deep-learning models exhibited the steepest decline, confirming over-fitting to subject-specific artefacts. Statistical feature-based approaches, though less affected, still under-performed clinically acceptable thresholds.

Significance
Our results expose key translational barriers in AI for EEG-based epilepsy diagnosis--data leakage, acquisition bias, and overfitting to patient idiosyncrasies--leading to severe performance erosion on clinical data. Rigorous patient-independent validation, transparent reporting (aligned with CARE principles), and well-curated multi-channel scalp-EEG datasets are essential to ensure clinically dependable AI tools for epilepsy diagnosis.