Модель глубокого обучения для автоматизированной интерпретации медицинских электрофизиологических данных

Информатика и автоматизация (Труды СПИИРАН). 2026. Т. 25. № 3. С. 720–750.

Lebedev O. B., Шмелева А. Г., Гежа Н. С.

This paper describes the development of a neural network model for automated analysis of medical data in electrophysiology based on deep learning methods. The relevance of this work stems from the growing need to improve the objectivity, speed, and accuracy of processing complex spatiotemporal signals, such as ECG or EEG. Convolutional neural networks (CNNs), which extract spatial features, and temporal convolutional neural networks (TCNs), which specialize in identifying temporal dependencies, were used as analysis tools. The integration of these architectures enabled the identification and generalization of complex patterns in signals and, subsequently, their classification with high accuracy. This paper presents the architecture of the proposed hybrid neural network model developed for time series processing and provides a rationale for the choice of key hyperparameters. To optimize the model, a strategy for adaptively changing the learning rate based on the value of the loss function on the validation dataset was implemented. An early training stopping mechanism was implemented to prevent redundant computations and to control overfitting in the event of no decrease in the error value on the validation data over a specified number of epochs. During training, model parameters were recorded after each epoch, including the epoch number, loss function value, and optimizer parameters. This allows for the reconstruction of training from any epoch and a detailed analysis of the model's behavior at various stages of training. Application of the developed analysis model enables deeper investigation of electrophysiological data, improving the objectivity of diagnostics. The methods presented in this paper can be scaled to other types of biomedical data. The practical significance of this work lies in the potential for integrating the proposed model into biomedical systems as a tool for supporting medical decision-making for automated pathology detection and reducing the workload on staff.

Keywords: машинное обучение, нейронные сети, neural networks, electrophysiology, электрофизиология, глубокое обучение, medical data, медицинские данные, machine learning, deep learning