Sztuczna inteligencja do rozpoznawania napadów padaczki w zapisach fal mózgowych na modelu zwierzęcym padaczki pourazowej

This study used machine learning to objectively identify seizures in the electroencephalogram of a model of post-traumatic epilepsy based on fluid percussion injury in male rats. We applied transfer learning to a neural-network trained and tested on three potentially distinct electroencephalographic phenotypes: (1) late-onset convulsive seizures associated with rare post-traumatic epilepsy, (2) early-onset convulsive seizures that often occurred after sham or injury treatment (independent of post-traumatic epilepsy), and (3) spike-wave discharges, which occurred in both injured and sham-control rats. The neural network was able to detect seizure events within individual animals and across different cohorts and showed that early and late seizures have similar electroencephalographic phenotypes. Additionally, crossover training and testing on spike-wave discharges from injured and sham-control rats distinguished a convulsive-seizure phenotype from normal spike-wave discharges. Convolutional neural network modeling of the electroencephalogram can identify spectro-temporal phenotypes that reliably distinguish spike-wave discharges from convulsive seizures, indicating that: 1) Spike-wave discharges are normal, not to be falsely classified as non-convulsive epileptic seizures, 2) The automated detection of convulsive seizures over months revealed rare post-traumatic epilepsy with low seizure frequency, 3) Early and late (epileptic) seizures were indistinguishable within and across rats, thus suggesting similar underlying neuronal circuits and ictogenic pathways. This commonality, however, may also obscure important differences between seizure types, 4) Convolutional neural network modeling may facilitate objective comparison of seizures within and between laboratories, supplementing subjective expert visual classification, and 5) The rarity of injury-induced epilepsy argues fluid-percussion injury is poorly suited for effectively testing anti-epileptogenesis therapies. CNN/EEG modeling can identify fundamental spectro-temporal phenotypes that reliably distinguish SWDs from convulsive seizures, indicating that SWDs are normal EEG, not to be falsely classified as non-convulsive epileptic seizures. Accurate, automated detection of convulsive seizures over months revealed PTE was rare after FPI, with low seizure frequency. Early and late (epileptic) seizures were indistinguishable within and across rats, thus suggesting similar underlying mechanisms and neuronal circuits. This commonality, however, may also obscure important differences between seizure types. CNN modeling can facilitate objective comparison of seizures across laboratories, supplementing subjective expert visual classification. The rarity of FPI-induced PTE, and the low seizure frequency in FPI-induced PTE argues that FPI is poorly suited for effectively testing anti-seizure and anti-epileptogenesis therapies.