Bonn EEG as Graphs: Structural vs. Recurrence Networks for Tiny-Data Seizure Screening

We study whether time-series-to-graph ($\text{TS} \rightarrow \mathrm{G}$) mappings enable accurate, lightweight seizure screening on the tiny, single-channel Bonn EEG corpus. Each segment is converted into either (i) a structural graph (two line-of-sight variants) or (ii) a recurrence network via delay embedding; permutation-invariant graph descriptors (degree/clustering/path statistics, Laplacian spectra, and 3-node motifs) feed simple classifiers (logistic regression, linear SVM, and random forest). On two tasks, T1 (binary: ictal vs. non-ictal) and T2 (3class: healthy/interictal/ictal), we evaluate with repeated, stratified$5 \times 5$-fold cross-validation, reporting Accuracy, Macro-F1, and AUROC (T1), and contextualize results with commonly reported Bonn baselines from prior work. Our compact pipeline attains T1 results of$\text{Acc}=98.7 \pm 0.8 \%$, Macro-$\mathrm{F} 1=98.5 \pm 0.9 \%,\ \text{AUROC} =0.996 \pm 0.003$; and T2 results of$\text{Acc}=95.9 \pm 1.5 \%$, Macro-F1${}={9 5. 5} \pm {1. 7} \%$(best: RN with density$p=0.10$, B+S features, RF). Robustness tests under additive noise (SNR 20 and 10 dB) show a Macro-F1 drop of only 0.7 / 1.9 points; inference remains CPU-friendly at {1.12 ms / segment. These results indicate that structural and recurrence graphs provide strong, reproducible baselines for tiny-data EEG screening without deep neural networks, while noting that Bonn may overestimate separability relative to clinical multichannel settings.