Abstracts

Automated sleep-wake state discrimination is required for real-time seizure prediction, detection, and control. We present a novel method for state discrimination in chronically implanted rats by introducing kinematic data utilized with EEG. The methodology uses automated analysis of head acceleration, along with epidural and hippocampal depth EEG using integrated power of low ([lt] 10Hz) and high ([gt] 10Hz) frequencies. Procedures were carried out under GMU ACUC approval. Male Sprague-Dawley rats (300g) were anesthetized and implanted with electrodes into both hippocampi, and with 5 epidural screws, with differential EEG recorded. Head acceleration and tilt were measured with a double axis DC-accelerometer, and behavior documented with a low-light camera. Starting 1 week post-operatively, recordings were made for 24h, and stored in digital format for later processing. 1s epochs were chosen for data reduction. Multivariate linear discriminant analysis was used for classification of groups: stationary-not-moving (G1), stationary-moving (G2), exploratory (G3). State identification over time was independently obtained from video inspection (Fig 1 top). Discrimination variables included mean and variance for each accelerometer axis (front to back, FB; side to side, SS), and EEG power quantified in low (L) and high (H) frequencies. Combining groups 2 and 3, linear discrimination demonstrated that the 2 states were significantly different (p[lt]0.001 histogram Fig 2A), although the overlap and error rate was considerable. The scatter plot of the canonical discriminant functions for 3 states is shown in Fig 2B, demonstrating significantly reduced overlap and error rates.[figure1][figure2] We show that the addition of accelerometer data and multivariate discrimination analysis can improve automated sleep-wake state classification. This methodology will be applied to real-time state discrimination in epileptic rats submitted to adaptive electric field stimulation, allowing the refinement of algorithms to detect, predict, and control seizures. (Supported by NIH grants R01EB001507, K02MH01493, and R01MH50006.)