Brain-computer interfaces translate brain activity into digital commands, offering new possibilities for controlling prosthetic limbs, wheelchairs, robotic systems, or computer cursors through thought alone.
A new study by Zhenis Otarbay of Nazarbayev University and Abzal Kyzyrkanov of Astana IT University addresses a key challenge in this field: making EEG-based brain-computer interfaces work reliably across different users. EEG signals are noisy and vary from person to person, which often requires long individual calibration.
The researchers proposed a transfer learning framework called ConvoReleNet, designed for motor imagery EEG classification. Instead of training a separate model from scratch for each user, the system learns from data across multiple subjects and adapts to new users more efficiently.
Key findings
- The study introduces ConvoReleNet, a transfer learning framework for motor imagery EEG classification.
- The model aims to improve subject-independent brain-computer interfaces, reducing the need for extensive calibration for each user.
- Accuracy improved from 72.22% to 79.44% on the BNCI IV-2a dataset.
- Accuracy improved from 75.10% to 83.85% on the BNCI IV-2b dataset.
- The framework reduced inter-subject variability, making performance more stable across users.
- The results support the development of more practical brain-computer interfaces for rehabilitation and assistive technologies.
Why it matters
Brain-computer interfaces have strong potential in healthcare, especially for people who have lost motor function due to neurological injury or disease. However, many current systems require lengthy calibration and do not perform equally well for all users.
By improving classification accuracy and reducing variability across subjects, this study helps address one of the central barriers to real-world BCI deployment. The proposed approach could contribute to future assistive systems that are easier to set up, more reliable, and more suitable for clinical and rehabilitation environments.