Abstracts

We describe a novel method for detection of seizure onset using a quantum neural network (QNN) that was progressively trained by quantitative features (in both the time and frequency domains) extracted from intracranial EEG (iEEG) signals. QNNs are particularly helpful in the classification of uncertain data without any restricting assumptions. [Purushothaman and Karayiannis, [italic]IEEE Trans Neural Networks[/italic] 1997; 8:679-693]. Raw iEEG data (subdural and depth recordings) from 4 subjects, sampled at 1000 Hz, were analyzed. Bandpass was set from 1.6 to 300 Hz. Seizure onset was marked at the time of earliest change in EEG. The EEG segment, 10 sec before and after seizure onset, was divided into onset epochs (OE), each of 2 sec length, with 50% overlap between consecutive epochs. Background epoch (BE) was defined as the average of 30-sec EEG during the interictal period, preceding the OE by at least 2 min.
The signal was decomposed using a Daub-2 wavelet and scales 2 to 6 (corresponding to frequency bins 250-125, 125-63, 63-32, 32-16, 16-8 and 8-4 Hz respectively) were used for feature analysis. Two parameters were computed for each scale: relative energy (ratio of the energy in a given scale in the OE to the energy of the same scale in the BE); average of the distribution tail (ratio of the average of the top 80th percentile squared coefficient to the average of the squared coefficient for a given scale). In addition to this average peak-to-peak amplitude and curve length (measure relating to the fractal dimension) of the OE are also computed and used as features.
QNN was initially trained using seizure onset patterns from multiple subjects. With each subsequent seizure, the QNN was progressively re-trained by populating the training with 50% of the epochs belonging to the last seizure, 25% belonging to the previous to last seizure and 25% from a bank containing a variety of patterns from multiple subjects. Decision-making was based on the inherent ability of the QNN to identify and quantify uncertainty, represented by its output values between 0 (non-seizure) and 1 (seizure), in conjuction with a spatio-temporal rule. Thus, seizure onset was reported when the activity was observed in at least 2 adjacent channels and the activity was sustained for at least 6 sec (with QNN output value [gt]0.65). If a channel had reported seizure in the past, its membership threshold was reduced to 0.4 and the spatio-temporal rule was applied. The QNN classifier was tested on 8 seizures from 2 subjects (each with 4 seizures). All the seizure onsets (100%) were detected using the above method. There were only a few false detections. Progressively trained QNNs appear to be promising in detection of intracranial seizure onset. Further investigation will focus on detection of a wider variety of seizure onset patterns in multiple subjects and different decision-making strategies.