Fig. 6: Illustration of the physiological processing system for EEG epileptic seizure detection.
a Flow chart of the VO2 memristor-based physiological signal processing system for epileptic seizure detection. Each EEG clip is encoded by the VO2 encoder and then classified by the LSNN. During training only, the cross-entropy loss and the spike regularization term are calculated and propagated backward to update the weights. A final post-processing step consisting of moving average and thresholding improves the performance of the system. b Spike raster of the encoded epileptic EEG clip in (a) with a CUE signal prompting a valid output period. c Spike raster of the LIF neurons. d Spike raster of the ALIF neurons. e The Vg evolution of five ALIF neurons. f The evolution of output probabilities showing that the system correctly classified the epileptic EEG clip. g Visualization of the contiguous target labels. h Visualization of the contiguous classification results of the LSNN. i The results obtained after applying moving average with a window of 9, showing reduced values for regions with false positives, while maintaining high values for regions with true positives. j The final classification results after thresholding by a value of 0.8. k A zoomed-in plot of the area marked by the right red rectangle in (h), showing the contiguous nature of true positives. l A zoomed-in plot of the area marked by the left red rectangle in (h), showing the distributed nature of false positives. m A zoomed-in plot of the area marked by the red rectangle in (j), showing the correctly predicted epileptic seizure clips. n Confusion matrix of the raw LSNN classification results before post-processing. o Confusion matrix of the post-processed classification results, illustrating the benefits of the post-processing step in increasing specificity and accuracy.
