Comparative analysis of neural network-based EEG classifiers for detecting the effects of anticonvulsants

A comparison of various machine learning algorithms was carried out, including a multilayer perceptron in combination with principal component analysis and linear discriminant analysis, as well as deep learning convolutional neural networks combined with wavelet analysis for the classification of electrocorticographic (ECoG) data to identify the effects of experimental anticonvulsants. The study was performed on 36 non-linear male rats weighing 250–290 g. For electrocorticogram (ECoG) recording, cortical electrodes were implanted in the animals, and after substance administration, signals were recorded (30 minutes) with a sampling frequency of 500 Hz and electrode impedance 5 kOhm. The rats were divided into three groups: group 1 (n = 8) was given intraperitoneal saline solution 1 ml/kg; group 2 (n = 8) was injected with carbamazepine 25 mg/kg; group 3 (n = 6) was injected with compound RU-1205 (20 mg/kg). The obtained signals were segmented into 5-second epochs. For each channel, 15 features were extracted and processed by principal component analysis and linear discriminant analysis. Alternatively, the Morlet wavelet transform was used to obtain spectrograms. Classification was carried out using a multilayer perceptron and a convolutional neural network. The constructed models were evaluated on electrocorticographic signals from a test set of 2160 samples, which were recorded after administration of compound RU-1205. The accuracy of the models ranged from 71–84 %. Moreover, the accuracy scores for all three models were statistically significantly different from the random classifier (p 0,05). The study showed that artificial neural networks, including multilayer perceptron and convolutional neural networks, can be effectively applied to classify electrocorticographic signals and determine the anticonvulsant activity of various compounds.