USING NETWORK-BASED FUNCTIONAL CONNECTIVITY TO ESTIMATE COGNITIVE LOAD AND ADJUST BRAIN COMPUTER INTERFACE AUTONOMY

Sammanfattning

Cognitive processes such as attention and memory are supported by coordinated activity across macro-scale brain networks. Understanding how these networks interact under varying cognitive demands is essential for both neuroscience research and the application of these networks in technologies such as BCI. However, these networks have largely been studied using fMRI, and their use in BCI is still limited. In this project, the functional connectivity of two macro-scale neural networks — the DMN and the CEN — was investigated under high- and low-cognitive states. The analysis was conducted on an EEG dataset comprising 20 subjects and 64 electrodes, using a 10-20 placement system. The preprocessing pipeline included bandpass filtering, artifact subspace reconstruction, independent component analysis for eye artifact removal, and a surface Laplacian filter to reduce volume conduction effects. Functional connectivity was then extracted using complex Morlet wavelets and quantified with dwPLI, and significant connections were identified through surrogate permutation testing. A two-way ANOVA revealed no statistically significant differences between the two networks; however, the dwPLI metric showed statistically significant differences between low and high cognitive loads and was therefore selected as an input feature for classification. Three ANN models were trained on these features, and predictions from the models were made through weighted voting. A DUPLEX algorithm was used to split the data into training and test sets, and the system was evaluated using a local client-server protocol simulating an online BCI datastream. A class-conditional closed-loop adaptive autonomy controller was constructed around the classifier, adjusting the level of system support based on the subject's performance. The system achieved a baseline accuracy of 77.0% without the adaptive controller active and two further tests were conducted: one configured to increase overall accuracy, reaching 84.23%, and one configured to increase task difficulty, resulting in a decreased overall accuracy to 53.81%. The results indicate that while the two networks could not be significantly distinguished from one another — likely due to spatial overlap between their electrode regions — both networks contributed sufficient information to enable cognitive load classification. Source estimation is identified as a promising direction for future work, where a recommended starting point would be to try using BrainStorm, as it has been used in other studies and may partially solve the spatial overlap limitation and allow a significant network-level differentiation. Overall, the findings demonstrate a viable approach to develop a BCI systems that adapt in real time to the user's performance and cognitive demand, enabling subject-specific adjustment of task difficulty. This makes the system more personalized than a static BCI system and could increase a subject's overall learning rate while reducing frustration.