What question did this study set out to answer?

The research aims to evaluate the efficiency of binary classification of motor images for human-machine interfaces using neural networks combined with optimization algorithms.

April 10, 2026

Efficiency of binary classification of motor images in human-machine interfaces using neural networks and optimization algorithms

Key Points

The research aims to evaluate the efficiency of binary classification of motor images for human-machine interfaces using neural networks combined with optimization algorithms.
Developed a neuroheadset with dry electrodes and wireless data transmission.
Analyzed various classifiers, including Rosenblatt perceptron and convolutional neural networks.
Used optimization algorithms like adaptive moment estimation and Levenberg-Marquardt.
Measured classification performance using Accuracy, Precision, Recall, and F1-score metrics.
Achieved a classification success rate of approximately 66% using the Accuracy metric.
The convolutional neural network combined with the adaptive moment estimation algorithm provided the best performance.
Results exceeded typical benchmarks for mobile brain-computer interfaces operating in real-time.

Abstract

A compact hardware and software system based on a neuroheadset with dry electrodes and wireless data transmission developed. The system designed to generate control inputs from the operator while simultaneously monitoring their functional state in real time and integrated into complex technological systems. The results of a study on the effectiveness of binary classification of operator motor patterns using several classifiers operating in combination with different optimization algorithms presented. The following classifiers analyzed Rosenblatt perceptron, linear discriminant analysis, and convolutional neural network. A classifier architecture based on a ResNet-type convolutional neural network consisting of eighteen repeating macrolayers is proposed. Using the Accuracy, Precision, Recall, and F1-score metrics, we analyzed the impact of various optimization algorithms (adaptive moment estimation, Levenberg-Marquardt with the proposed upgrade, stochastic gradient descent, and Broyden-Fletcher-Goldfarb-Shanno) on classification results. The best online performance demonstrated by a combination of a convolutional neural network-based classifier and the adaptive moment estimation algorithm. The classification success rate using the Accuracy metric was approximately 66%. The obtained results found to exceed typical results for mobile handheld brain-computer interfaces operating in real-time (online).

Bookmark

View Full Paper