Hardware efficient seizure prediction algorithm

Epilepsy affects 2.5 million people in the USA, 15% of which cannot be treated with traditional methods. Effective treatments require reliable prediction of seizures to increase their effectiveness and quality-of-life. Phase synchronization phenomenon of two distant neuron populations for a short period of time just prior to a seizure episode is utilized for such prediction. This paper presents a hardware efficient prediction algorithm using phase-difference (PD) method instead of the commonly used phase-locking statistics (PLS). The dataset has been collected from publicly available CHB-MIT Scalp EEG Database that consists of scalp EEG recordings from 10 pediatric subjects with intractable seizures. The seizure channel is selected based on the maximum value of the standard deviation during seizure, while the reference channel has the minimum value of the standard deviation. Data from these two channels is conditioned with a band-pass (f(lc) = 10Hz, f(hc) = 12.5Hz) 6th order Chebyshev Type II filter or a FIR filter. The analytical signals are derived using Hilbert Transform to allow phase extraction. PLS and PD are calculated from the mean of the phase-differences using an overlapping sliding-window technique. PD method demonstrates the same characteristics as PLS, while achieving 2.35 times faster computation rate in MATLAB than PLS. With 51 seizure episodes, prediction latency was between 51 seconds to 188 minutes with sensitivity of 88.2%. PD yields to lower hardware requirement and reduces computational complexity.