Inferring Cognitive Strategies from EEG Data via Neuro-Cognitive Multilevel Causal Modeling
Abstract
We applied a novel framework – Neuro-Cognitive Multilevel Causal Modeling (NC-MCM) [1] – to human neuronal data recorded with EEG. The framework offers a promising approach to developing a unified theory of cognition, bridging the gap between Marr’s three levels of analysis – computational, algorithmic, and implementational [2].
Classical neuroscience typically operates at the implementational level, focusing on biophysical mechanisms of the brain, while artificial intelligence concentrates on the computational level, modeling input-output relations. The algorithmic level – detailing cognitive strategies and internal representations – remains underdeveloped. This gap underscores the need for an integrative, interdisciplinary approach that not only links these levels but also provides a shared conceptual and formal foundation for the study of cognition.
Recent advances in machine learning have enabled the decoding of cognitive phenomena, such as stimuli or behavioral responses, from neuronal data. Despite their utility, decoding methods are inherently correlational and offer limited explanatory power regarding the causal structure or functional significance of the identified representations. The NC-MCM addresses this limitation by offering a mathematically rigorous neuronal-decoding framework that allows causal reasoning. Prior work has demonstrated the model's utility in inferring causally meaningful cognitive states from neuronal data of the nematode C. elegans; however, it has not yet been validated on human data.
In this pilot study, we analyzed EEG data recorded from a human participant performing an audiovisual attention task. The paradigm consisted of different combinations of auditory and visual stimuli, prompting one of two button presses. The participant was further instructed to apply distinct strategies: either prioritizing visual or auditory stimuli. Bandpower features were extracted from this data and analyzed using logistic regression. Additionally, EEGNet, a convolutional neural network designed specifically for EEG analysis [3], was trained on the raw data.
Both models successfully decoded button presses as well as auditory and visual stimuli. The outcome of the NC-MCM in inferring the participant’s strategies was promising, but further analysis and tuning of the models and experimental paradigm are needed to statistically validate its potential. The final contribution of this project is an extension to the NC-MCM toolbox, which includes a feature extraction pipeline for EEG data, and a method for integrating EEGNet in the analysis. With these additions, future research can seamlessly apply the NC-MCM framework to EEG data, marking a crucial step toward confirming the model’s utility.
References
[1] M. Grosse-Wentrup, A. Kumar, A. Meunier, and M. Zimmer, “Neuro-cognitive multilevel causal modeling: A framework that bridges the explanatory gap between neuronal activity and cognition,” PLOS Computational Biology, vol. 20, no. 12, p. e1012674, Dec. 2024. doi: 10.1371/journal.pcbi.1012674.
[2] D. Marr, “Vision: A Computational Investigation into the Human Representation and Processing of Visual Information,“ Henry Holt and Co., Inc., 1982.
[3] V. J. Lawhern, A. J. Solon, N. R. Waytowich, S. M. Gordon, C. P. Hung, and B. J. Lance, “EEGNet: a compact convolutional neural network for EEG-based brain–computer interfaces,” Journal of Neural Engineering, vol. 15, no. 5, p. 056013, Jul. 2018. doi: 10.1088/1741-2552/aace8c.
Published
Issue
Section
License
Copyright (c) 2025 Leonie Gadner

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