Detection of seizure foci by recurrent neural networks

Localization of epileptic foci is an important step in planning surgical treatment of medically intractable epilepsy. The solution to this problem depends on the detection of the earliest time of seizure onset in multiple channels of EEG data. We have been investigating a seizure-detection system based on recurrent neural networks to detect the onset of seizure activity in multi-channel subdural EEG (SEEG) data. The recurrent neural network used in this investigation is a fully-connected, three-layer, recurrent network, in which the output is given by: x(t+1)=f([Σ(P1(t)*W12)+x(t)*W3]), where f(·) is the activation function; P1 is the output of the input unit; W12 is the weight of the connection between the input and the hidden units; W3 is the weight value of the recurrent connection. Using an adaptive backpropagation gradient-descent algorithm, the network is trained on sets of pre-seizure data (target value of zero) and seizure data (target value of unity), specified over discrete time-intervals, from 76-electrode SEEG records from a patient with intractable seizures. The network´s performance is measured by a linear-regression metric (R value) obtained from a fit between the target value and network output for each electrode and time-interval in the record. A metric with a value approaching unity in the pre-seizure component of the record suggests the onset of seizure activity. Ordering these metrics with respect to subdural electrode positions may provide useful information about the epileptogenic foci