A Network Disruption Theory of Epilepsy Supported by Artificial Intelligence Tools and Electrophysiological Fingerprints
W skrócie
[Preprint - wstępne wyniki] Badacze wykorzystali sztuczną inteligencję do analizy aktywności mózgu u 87 pacjentów z padaczką i odkryli, że ataki padaczkowe nie rozprzestrzeniają się przez odrębny chorobiczy obwód mózgowy, jak dotychczas sądzano, lecz zaburzają normalnie funkcjonujące sieci mózgu. Wyniki sugerują, że ognisko ataku stanowi punkt dysfunkcji w normalnej sieci, a nie część osobnego systemu chorobowego. Te ustalenia mogą zmienić podejście do leczenia chirurgicznego padaczki - zamiast usuwania zmian, należałoby przywracać prawidłową funkcję sieci mózgowych.
Oryginalny abstract (angielski)
Abstract Background : The dominant model of epilepsy at the systems level - the tripartite epileptic network, comprising seizure onset zone (SOZ), propagation zone, and uninvolved cortex - has guided clinical practice for decades. This model rests on a largely untested assumption: that epileptic seizures propagate within a disease-specific pathological circuit that is distinct from the normal functional architecture of the brain. We asked a more fundamental question: is this assumption correct, or do seizures instead disrupt and propagate within normal functional brain networks, with the SOZ representing nothing more than a focal point of pathological vulnerability embedded in an otherwise intact network? Methods : We trained a 6-layer Transformer encoder via masked autoencoding and subject-contrastive learning on 2,656 hours of continuous intracranial EEG (iEEG) from 18 patients (SWEC-ETHZ long-term dataset) - entirely without human-defined labels - and applied it in zero-shot fashion to two additional independent datasets (HUP dataset: 54 patients; SWEC-ETHZ short-term: 15 patients), yielding 87 patients and 537 seizures in total. We characterised the temporal dynamics, spatial organisation, and functional network affiliation of seizure electrophysiological fingerprints using the Schaefer 400-region/Yeo-7 network parcellation. Results : Four convergent findings emerged, all inconsistent with the disease-specific circuit model. First, seizures are abrupt phase transitions in AI feature space, with seizure termination producing a significantly larger feature-space displacement than initiation across all three datasets (SWEC long-term: offset jump 0.541 vs. onset jump 0.304, p=0.0003; HUP: 1.423 vs. 0.982, p Conclusions : These results directly challenge the epileptic network model. We propose the Network Disruption Theory of Epilepsy: epileptic seizures do not travel through a disease-specific circuit but instead originate at a focal disruption point (the SOZ) within a normal functional network and propagate along its pre-existing connections. This reconceptualisation shifts surgical focus from "resect the lesion" to "restore network integrity".