Accurate localization of brain activity recorded by M/EEG measurements to their generating brain structures is a prerequisite for investigating brain functional connectivity. The project team achieved to develop algorithms that perform substantially better than existing ones under low-noise conditions and are furthermore able to automatically learn the noise level and correlation structure from the data, further contributing to improved localization. The project team further developed robust measures of non-linear brain interactions. Novel methods were developed for the estimation of phase-amplitude coupling (PAC) as well as the estimation of transmission delays under mixed noise. PAC is a widely hypothesized mechanism of brain communication by which the amplitude of a fast oscillation is tied to the phase of a slow oscillation. Being able to estimate transmission delays between regions, on the other hand, is critical to identify mental and clinical state dependent deviations from the brains normal signal transmission pathways. The novel methods are theoretically and practically shown to be insensitive to inevitable mixed noise sources such physiological and technical artifacts, leading to dramatic improvements in estimation quality.
The developed source connectivity analysis pipelines were extensively validated. Best practice pipelines ere identified, confirming the superiority of the newly developed methods. These pipelines were integrated into open-source toolboxes for Matlab and Python, including ROIConnect, PyBispectra and the MNE multivariate connectivity toolbox.
Levodopa medication and deep brain stimulation (DBS) are effective treatments that substantially reduce symptoms in PD, but their mechanisms of action has not been well understood. We, therefore, investigated functional connectivity between cortical and subcortical regions of the motor network under medication and stimulation on/off conditions. The observed results suggest that medication and stimulation possess shared but also distinct mechanisms of action, differentially activating different cortico-basagal-ganglia pathways. Further work characterized functional during gait using the developed FC analysis pipelines and marks an important step towards adaptive DBS that can switch stimulation parameters depending on the movement pattern (e.g. walking vs. standing). These insights will inform future, improved, treatments, for example using adaptive stimulation.
The project employed machine learning (ML) methods to extract specific brain signals, for example those encoding movement intention, with high signal-to-noise ratio from the multivariate M/EEG data. A scientific focus was then put on the question of how to correctly interpret such multivariate models in neurophysiological terms. To this end, the project group showed that highly popular XAI methods can lead to substantial misunderstandings about the relationship between input features of an ML model (e.g. functional properties of brain areas) and the predicted target variable (e.g. the presence of a disorder), thus promoting incorrect conclusions about brain functioning and generally limiting the utility of such methods for purposes such as model or data validation, counterfactual reasoning or scientific discovery. Further project work devised controlled benchmark data to validate the performance of XAI algorithms in linear and non-linear settings. All benchmarks and metrics for explanation peformance developed within the project have been made openly available and will enable future studies to avoid misinterpretations of multivariate machine learning models.
