Result: Классификация фрагментов электроэнцефалограммы по степени патологической значимости с помощью последовательных нейронных сетей: Classification of Electroencephalogram Segments Based on Pathological Significance Using Sequential Neural Networks

Исследована возможность классификации фрагментов электроэнцефалограммы как элементов нормальной, пограничной и патологической электроэнцефалограммы при помощи последовательных нейронных сетей. Для этого создана новая обучающая база данных, содержащая по 500 фрагментов сигналов нормальной, пограничной и патологической ЭЭГ. С использованием разработанной базы данных поставлен ряд экспериментов по обучению последовательных нейронных сетей с различным количеством слоёв и числом нейронов в слоях, при этом наибольшая точность классификации фрагментов электроэнцефалограммы составила 80,1 %. Для обеспечения возможности запуска и дообучения разработанных нейронных сетей на различных электронных устройствах выполнена реализация алгоритма обратного распространения ошибки. Introduction. Electroencephalography (EEG) is the primary method for diagnosing the functional state of the human central nervous system. Many conditions, such as epilepsy, brain tumors, cranial and cerebral injuries, vascular diseases, psychiatric disorders, and various forms of encephalopathy, can be reflected in EEG recordings. Despite its widespread use, EEG analysis is still mostly conducted visually by experienced clinicians, which is time-consuming and subjective. The need for more efficient and reliable EEG analysis has led to the development of decision support systems that use artificial intelligence technologies for EEG pattern classification. Neural networks have proven to be particularly effective in achieving high classification accuracy in many applications. The aim of this paper is to enhance the efficiency of automatic electroencephalogram (EEG) classification based on pathological significance through the application of artificial intelligence methods, particularly sequential neural networks. To achieve this goal, the following objectives are addressed in the study: 1) a review of computational EEG analysis methods; 2) creation of a training database of EEG fragments, including a sufficient number of normal, borderline, and pathological EEG signals; 3) implementation of a neural network training algorithm; 4) examination of the effect of sequential neural network parameters on the accuracy of EEG fragment classification. Results. A new training database consisting of 500 EEG fragments, including normal, borderline, and pathological signals, was created. The database was used to train sequential neural networks with varying numbers of layers and neurons, resulting in the highest classification accuracy of 80.1 %. The error back propagation algorithm was implemented to allow the developed neural networks to be run and retrained on various electronic devices. The practical value of this work lies in the reduction of speculative operations performed by neurophysiologists and the provision of a more objective, qualitative, and faster intellectual EEG analysis when used in medical decision support systems.