Deep learning models for detection and foreseeing of epileptic seizures

Epilepsy is one of the most widespread globally diseases and nowadays several steps to its classification and treatment have been made. Starting from the electroencephalogram ( EEG ) of epileptic patients, the first step analysis consists of choosing those TRC files with an effectively recording channels into a range included between 24 and 35: the number of channels changes a lot from patient to patient, so it is threshold to collect as much as possible quite homogeneous data. Converting each selected file into a dataframe, the analysis goes on with splitting them into no-ictal and ictal phases: using Brainstorm is actually possible to extract the start and the end events related to epileptic seizure ( ES ). After checking that every ES is truely indipendent from the others ( this happens after 4h from the end of the previous one ), the time data order for each no-ictal part is reversed and every 5 seconds the signal is first downsampled to 256 Hz and then log-spectrogram is computed for each channel. Finally a filter system composed by 1 Hz high pass, 125 Hz low pass and notch filter to 50 and 100 Hz is applied to clean the spectrogram from noise frequencies. If some dataframe has less then 35 channels an empty spectograms with null signal are added in order to have 35 images for each 5 seconds time period. As the aim of the work is to predict the ES in increasingly detailed time brackets, each spectrogram-packet ( composed of 35 images ) is first labeled and then the new spectrogram dataset is balanced between time categories. Next, a split between the training and test set is necessary in order to train a CNN model in supervised learning. The test set is the 10 % of the entire dataset and it’s composed by chosen random item from the various categories: that why it’s called random test. To better tuning the model hyperparameters a 3-fold cross validation is performed and to boost up the accuracy both training and random test are normalized with mean and std computed on training part. Despite on random test an 88 % of accuracy is achieved, on the last two patients benchtest the performance falls down near to 28 %. So, using a dataset with unknown information about seizure type and inhomogeneous recording channels ( both for number and type ), it’s not possible, for now, to build a model able to predict the arrival of a generic epileptic attack with reasonable temporal precision.