Abstracts

Rationale: The ability to reliably predict seizures would make preventive intervention possible for patients with intractable epilepsy. A seizure prediction algorithm (SPA) typically estimates dynamical variables computed from EEG, ECoG or depth recordings, and flags an abnormal trend as a sign of impending seizure. Existing SPAs show limited specificity and are prone to false prediction. Our hypothesis is that normal state(s) of vigilance (SOV)—e.g., wakefulness, REM sleep or non-REM sleep—and transitions between these states, play a significant role in determining seizure probability and SPA performance. In order to test this hypothesis, we propose to discriminate SOV in an animal model of epilepsy and correlate SOV with seizure occurrence. Methods: We have implemented a chronic, behaving animal model of limbic epilepsy as a test-bed for feedback seizure control strategies using polarizing low-frequency electric fields. Tetanus toxin implanted in the ventral hippocampus of rats generates spontaneous, intermittent, generalized seizures for several weeks following a latent period of 2-4 days. We have also developed methods for discrimination of instantaneous SOV from EEG and kinematic measurements in non-epileptic controls (J. Neurosci. Meth., 163:373-83, 2007). In the present work we extend this methodology, using the computational framework of the hidden Markov model (HMM), to track SOV in chronic recordings from epileptic animals and to test for characteristic SOV changes in the preseizure period. Results: A simple HMM, constructed using estimates of the EEG theta power fraction (4-8 Hz), and motion derived from a head-mounted MEMS accelerometer in 12 s epochs of data, was able to discriminate between non-REM sleep, REM sleep, quiet wakefulness, and active awake behavior with over 80 % accuracy in epileptic animals in the latent period prior to development of any seizures. The HMM was then used to track SOV in recordings containing spontaneously recurring seizures. Examination of a sample of 42 seizures showed that seizures were more likely to occur during transitions from REM sleep (55 %) than from non-REM sleep (10 %) or awake behavior (35 %). In a five minute period leading up to seizure onset, the probability that the animal was awake increased gradually from about 10 % to 35 %, suggesting that seizures in this model may accompany arousal from sleep. Similar observations are common in clinical monitoring. Comprehensive analysis of recordings from a number of animals is ongoing to investigate the precise nature of preictal SOV transitions in this epilepsy model. Conclusions: Quantitative correlations between SOV and seizures in epilepsy will enable improved state-dependent seizure prediction and control. Our long-term goal is to translate this approach for use in human patients. Computational models of SOV dynamics are expected to impact the understanding and treatment of epilepsy, with potential applications in sleep research and neural prosthetics as well.