Work

A Subject-Independent Model for Motor Imagery Brain-Computer Interfaces, with Extensions to Emotion Recognition and Seizure Detection

Year: 2026
Type: dissertation
Abstract: Motor imagery brain-computer interfaces (MI BCIs) translate imagined movements into control signals. However, there are different challenges that limit their performance. Typical applications require per-subject calibration, which takes a lot of time. To address this issue, we developed a unified framework that uses raw electroencephalography data to predict the imagined movements. As a backbone, we used the EEGNeX model with modifications and provided two new architectures, ImprovedEEGNeX and EEGNeX_SI_v2, with a 12-metric evaluation and Streamlit toolbox for easier use and analysis. ImprovedEEGNeX is a heavy model that includes various modifications, such as knowledge distillation from a spectral teacher, Euclidean Alignment, multi-scale temporal convolutions, etc. EEGNeX_SI_v2 replaces spatial BatchNorm with GroupNorm to reduce subject leakage. Both models were tested on BCI Competition IV-2a, ImprovedEEGNeX achieves 79.2% subject-dependent and 72.0% subject-independent four-class accuracy, while EEGNeX_SI_v2 offers competitive subject-independent performance with under 40k parameters. The same backbone and tooling were extended to emotion recognition and seizure detection and for that, the EAV and Siena Scalp EEG datasets were used. For easier usage, a Streamlit GUI toolbox was created to help analyze, preprocess, reproduce the results, etc.
Language: English
Open Access status: green
OpenAlex ID: W7197289311
Cites: 18
Cited by:
Related to: 10