Leakage-Controlled Recording-Level Evaluation of Machine Learning, Deep Learning, and Convolutional-Kernel Models for EEG-Based Epileptic Seizure Detection

This is the authors' abstract. We don't add key points for this paper.

Electroencephalogram (EEG)-based seizure detection is a challenging time-series classification problem because EEG signals are nonlinear, non-stationary, and temporally heterogeneous. This study presents a unified comparison of classical machine-learning, neural and deep-learning, convolutional-kernel time-series, and probability-level ensemble models using the Epileptic Seizure Recognition benchmark dataset. The 11,500 EEG segments were reconstructed into 500 recording groups, and recording-grouped nested cross-validation was applied to prevent segments from the same recording from crossing training and test partitions. Depending on the modelling paradigm, segments were represented either by 67 handcrafted time-, frequency-, and nonlinear-domain features or by ordered 178-sample EEG sequences. Recording-level average precision was used as the primary evaluation metric, supported by ROC-AUC, threshold-dependent measures, Brier score, bootstrap confidence intervals, ablation analysis, explainability, and computational-efficiency benchmarking. RBF-SVM achieved the highest average precision of 0.997985 and ROC-AUC of 0.999475. MultiRocket produced the lowest family-representative Brier score of 0.006148, while MLP achieved the lowest standalone inference latency after feature extraction. The selected meta-ensemble produced strong threshold-dependent performance but did not improve overall precision–recall ranking. Paired bootstrap analysis found no statistically significant differences among the family representatives after multiple-comparison correction. These findings demonstrate that handcrafted-feature and time-series models can both achieve strong recording-level performance when evaluated under a consistent, leakage-controlled protocol. Because verified patient identifiers were unavailable, the results should not be interpreted as evidence of patient-independent clinical generalization.