Real-time seizure prediction from local field potentials using an adaptive Wiener algorithm

Approximately 30% of individuals with epilepsy have refractory seizures that cannot be controlled by current pharmacological treatment measures. For such patients, responsive neurostimulation prior to a seizure may lead to greater efficacy when compared with current treatments. In this paper, we present a real-time adaptive Wiener prediction algorithm implemented on a digital signal processor to be used with local field potential (LFP) recordings. The hardware implementation of the algorithm enables it to be a miniaturized portable system that could be used in a hand-held device. The adaptive nature of the algorithm allows the seizure data to be compared with baseline data occurring in the recent past rather than a preset value. This enhances the sensitivity of the algorithm by accounting for the time-varying dynamics of baseline, inter-ictal and ictal activity. The Wiener algorithm was compared to two statistical-based naïve prediction algorithms. ROC curves, area over ROC curves, predictive power, and time under false positives are computed to characterize the algorithm. Testing of the algorithm via offline Matlab analysis on kainate-treated rats results in prediction of seizures about 27 s before clinical onset, with 94% sensitivity and a false positive rate of 0.009 min(-1). When implemented on a real-time TI C6713 signal processor, the algorithm predicts seizures about 6.7s before their clinical onset, with 92% sensitivity and a false positive rate of 0.08 min(-1). These results compare favorably with those obtained in similar studies in terms of sensitivity and false positive rate.