Identification of cortical sources contributing to BCI classification

• Applicant: Professor Dr. Thorsten Zander
• Subject Area: Human Cognitive and Systems Neuroscience ; Human Factors, Ergonomics, Human-Machine Systems ; Image and Language Processing, Computer Graphics and Visualisation, Human Computer Interaction, Ubiquitous and Wearable Computing
• Term: since 2024
• Project identifier: Deutsche Forschungsgemeinschaft (DFG) - Project number 541907312

A Brain-Computer Interface (BCI) is a neurotechnological system that provides a direct communication pathway from a human brain to a computer. BCIs process brain activity, convert it into distinguishable features, and employ machine learning to construct a classifier which distinguishes cortical correlates of the user’s internal cognitive states. However, the machine learning algorithms employed by BCIs are usually data-driven and obscure, leading to a lack of explainability: the extent to which the internal mechanics of the algorithm can be explained in human terms is small. Furthermore, because these data-driven approaches are agnostic to the neurophysiological basis of the data, it is unclear to what extent the resulting classifiers can reliably identify the brain activity which is correlated to the considered cognitive states, and to what extent they rely on uncorrelated cortical or noncortical (artifactual) activity. A mismatch between the cortical regions neurophysiologically known to generate the correlated brain activity and the sources actually contributing to classification can lead to suboptimal, non-robust, or potentially dangerous technology. In this proposal, we present a method to validate BCI classifiers by identifying which neurophysiological processes contribute to classification. The validation method builds upon the general principle of a visualization tool that visually indicates which sources are relevant for a BCI classifier. The main goals of this proposal are: 1. To transform the visualization tool into an objective validation method; 2. To identify the factors which influence the validation method’s output; 3. to quantify the output of the validation method in a way that objectively reflects a classifier’s brain focus, i.e. the proportion of cortical sources contributing to classification, and the validation method’s confidence in this metric; 4. to demonstrate the knowledge gain compared to traditional classifier analysis methods. We plan to create a validation method which originates from the visualization tool, by shifting the focus to quantifiable classifier validation, and implementing it into an open-source toolbox. This will result in an analysis framework that produces a new, neurophysiologically relevant measure of validity, increases robustness of classifiers and adds explainability to the field of BCI.

• DFG Programme: Research Grants