Student Conference Proceedings
Vol. 1 No. 1 (2025): Stud Conf Proc
https://doi.org/10.18416/SCP.2025.1947

Medical Informatics, ID 1947

Comparison of Models for EEG-based Cross-Subject Movement Prediction

Main Article Content

Laeticia Moosdorf (1) Study program Medical Informatics, Universität zu Lübeck, Lübeck, Germany; 2) Robotics Innovation Center, German Research Center for Artificial Intelligence, Bremen, Germany), Niklas Kueper (1) Robotics Innovation Center, German Research Center for Artificial Intelligence, Bremen, Germany; 2) Institute of Medical Technology Systems, Universität Duisburg-Essen, Duisburg, Germany), Heinz Handels (1) Institute of Medical Informatics, Universität zu Lübeck, Lübeck, Germany; 2) AI in Medical Image and Signal Processing, German Research Center for Artificial Intelligence, Lübeck, Germany), Elsa Andrea Kirchner (1) Institute of Medical Technology Systems, Universität Duisburg-Essen, Duisburg, Germany; 2) Robotics Innovation Center, German Research Center for Artificial Intelligence, Bremen, Germany)

Copyright (c) 2025 Laeticia Moosdorf, Niklas Kueper, Elsa Andrea Kirchner, Heinz Handels

Creative Commons License

This work is licensed under a Creative Commons Attribution 4.0 International License.

Abstract

Brain-computer interfaces for robotic-assisted rehabilitation based on EEG recordings can play an important role in the rehabilitation of stroke patients. To effectively classify movement intentions from EEG signals, machine learning models are needed that can work with noisy and heterogeneous data. This work investigates four neural networks, namely the MLP, TKAN, EEGNet, and SincEEGNet, in their ability to predict movement intentions in intra- and cross-subject classification. Additionally, the amount of calibration data for adapting the pre-trained models of the cross-subject task to data of the target subject was explored. The model performance for intra-subject classification is highest with the EEGNet, however, SincEEGNet performs best in cross-subject classification. Calibrating the pre-trained models results in a performance gain with the overall highest accuracy of 88.1 % with SincEEGNet. Those results motivate to apply such models for EEG classification across subjects to reduce calibration times.

Article Details