Logo of the Physikalisch-Technische Bundesanstalt

“Probabilistic Source Modeling, Simulation, Validation and Interpretation in Functional Neuroimaging“

Kolloquium der Abteilung 8

Data analysis pipelines in neuroimaging studies involving EEG/MEG measurements can be very complex. For example, in order to relate functional interactions between brain areas to cognitive or clinical variables, one needs to solve the EEG/MEG inverse problem, estimate functional interactions between the reconstructed source time series, and apply mass-univariate statistical tests or multivariate machine learning models for prediction. Each of these steps may introduce variability and biases, and cause misinterpretations, all of which would add up in the final result. In this talk I will introduce the EEG inverse problem and outline penalized maximum-likelihood and empirical Bayesian techniques to deal with its non-uniqueness. I will further argue that spatially correlated noise is an essential feature of EEG/MEG and other neuroimaging data and will demonstrate that it can lead to incorrect assessment of interactions between time series and the misinterpretation of parameters learned by a machine learning model. Remedies to these problems for simple yet relevant cases (linear dynamics, linear multivariate models) will be proposed. Finally, simulations and benchmarks will be highlighted as important validation tools that can accompany or even drive theoretical developments.