Comparative evaluation of tabular-to-image deep learning pipelines for EEG-based epileptic seizure recognition: A...

Epilepsy affects roughly 50 million people worldwide and is diagnosed primarily through electroencephalography (EEG), yet the manual review on which this diagnosis depends remains slow, subjective, and inconsistent across readers. A rapidly expanding literature reframes EEG signals as two-dimensional images, allowing ImageNet-pretrained vision backbones to be repurposed for seizure recognition via transfer learning, but no prior work has disentangled the competing tabular-to-image encodings from the choice of architectural family, nor characterized their joint behavior within a single leakage-controlled protocol. The study closes this gap by factorially screening seven encodings (DeepInsight, IGTD, REFINED, Gramian Angular Field (GAF), Markov Transition Field, Tab2Img, and an HSV channel scheme) against ten backbones spanning the convolutional, pure-transformer, and hybrid families across binary, three-class, and five-class tasks from the Bonn EEG corpus. Evaluation employed GroupKFold partitioning, dual seeds, Optuna hyperparameter optimization, paired non-parametric inference with Holm-Bonferroni correction, Cohen's d effect sizes, and Grad-CAM/Attention-Rollout explanations. Across more than 600 final runs, DeiT-Small with IGTD attained a macro-F1 of 0.9868 ± 0.0070 on the saturated binary task, whereas the hybrid CoAtNet-0 with GAF dominated the harder three- and five-class settings (0.9538 and 0.7799). Two reciprocal within-superset confusions jointly accounted for 67.5% of the residual five-class error, while hybrids achieved Cohen's d advantages of +1.2 to +2.0 and occupied the accuracy-efficiency Pareto frontier. Architecture and encoding interact non-trivially with task granularity, and the persistent five-class gap is best closed by enriching the input representation rather than enlarging the backbone.