Investigation of the effectiveness of classifiers of motor images constructed using diffusion models of artificial neural networks
D.V. Zhuravlev, A.N. Golubinsky, Andrey Tolstykh, A.A. Reznichenko · Biomedical Radioelectronics · 2025
The classification of motor imagery based on electroencephalogram signals is a complex applied task. Especially when it is implemented in portable small-sized brain-computer interfaces. Currently, there is no single solution that allows the classification of motor imagery with sufficient reliability and accuracy so that this technology can be used everywhere. Among the numerous models of artificial neural networks used to classify motor imagery, a new and not sufficiently studied direction is the use of diffusion neural networks for such operations. In relation to the brain-computer interfaces, to conduct a study of the effectiveness of using various models of diffusion neural networks in the tasks of classifying motor imagery based on electroencephalogram signals. To show the versatility of diffusion networks for working with dictator-dependent and dictator-independent datasets. Also, to conduct a study of the effect of pre-filtering of signals on the quality of classification. The efficiency of classifying motor images of both real and imaginary hand movements was studied using classifiers based on models of diffusion neural networks TimeGrad, ScoreGrad and DVa. Using the standard metrics "accuracy" and ROC-AUC, an analysis of the effectiveness of classifications, both dictator-dependent and dictator-independent data sets, was carried out. The maximum value of the "accuracy" metric was 60.93%, and the ROC-AUC metric was 0.813832 for a third-party dataset. The effect of filtering electroencephalogram signals on classification accuracy is also studied. The efficiency of diffusion neural network models is compared with the classical convolutional neural network model. This study makes a significant practical contribution to the work on improving technologies in the field of developing brain-computer interfaces, as new options for constructing classifiers based on models of diffusion neural networks have been explored. Thanks to the conducted computational experiments and their analysis, a number of conclusions and recommendations were made that will allow us to develop better classifiers of motor images in brain-computer interfaces.