Topics and Methods of the Brain Data Science Group

In order to isolate individual brain processes with high SNR and to enable neurophysiologically meaningful interpretations, it is usually indispensable to transform EEG/MEG sensor readings to their underlying cortical current generators (sources) before proceeding with further analysis. This inverse problem, however, does not have a unique solution unless prior knowledge on the structure of the sources is introduced. We have previously successfully used structured sparsity constraints to encode a preference for physiologically plausible source configurations (Haufe et al, 2008, 2011; Castaño-Candamil et al., 2015). Currently, we are working on methods to improve source estimations by automatically learning the spatio-temporal structure of the sources from observations in a hierarchical Bayesian framework using efficient non-linear optimization. We moreover explore ways to automatically learn the mapping from sensor to source using state-of-the-art supervised deep learning techniques based on carefully designed synthetic training data.

EEG/MEG source modeling requires the availability of an electromagnetic volume conductor model of the head, which is computed from a structural magnetic resonance image (MRI). We have published a highly detailed FEM model of the average human head (Huang et al., 2016), which can serve as a "gold standard" for source reconstruction if no individual MR image is available. Next, we plan to generate similarly detailed models for individual subjects. We further plan to use imputation methods from deep learning to enable the generation of accurate head models from de-identified MRI data as commonly found in large public data repositories.

Technical (e.g., line noise) and physiological (heart beat, eye movements) artifacts are ubiquituous in EEG and MEG data. To remove contaminations with such signals, we have built a quasi-automatic preprocessing pipeline for EEG data. As part of this procedure, independent component analysis and supervised classifaction is used to automatically identify and remove artifactual signal components (Winkler et al., 2011).

The high temporal resolution of EEG/MEG enables the study of  functional brain connectivity at different frequencies and according to different postulated mechanisms of brain communication. However, established FC metrics are prone to spurious detection of connectivity (Nolte et al, 2004). We have proposed time reversal to overcome this problem for linear connectivity metrics such as Granger causality (Haufe et al., 2013; Winkler et al., 2016), which has been widely acknowledged as the most effective way to estimate directed connectivity in the presence of correlated signals/noise by our peers. We are now interested in achieving for the first time a similar robustification for important non-linear connectivity metrics (e.g. amplitude-amplitude and phase-amplitude coupling), which have been postulated to reflect likely mechanisms of brain communication. If successful, our approaches would thus enable the non-invasive assessment of these types of brain interactions as potential biomarkers to study health and disease. We also study to what extent topological properties of connectivity graphs, a widely used family of summary statistics thought to provide a global view of brain efficiency, can at all be reconstructed from EEG/MEG data.

BSS techniques are powerful unsupervised machine learning algorithms that can decompose the observed data into latent signal components based on assumptions on their statistical properties. As such, they are a widely used alternative to inverse solutions based on a physical head model for identifying the activity of underlying brain sources from EEG/MEG measurements. We have already developed BSS methods to extract specific oscillatory activity as well as sources with linear connectivity structure from EEG/MEG data with high SNR (e.g. Haufe et al., 2010; Dähne et al., 2014). We are now interested in developing methods allowing the investigation of non-linear coupling between sources, as well as in fusing BSS methods with inverse solutions using state-space approaches in order to combine accurate estimation of sources times courses with precise localization of these activities.

The wealth of information contained in EEG/MEG functional connectivity data (e.g., voxel x voxel x frequency arrays) makes it a challenging task to identify the relevant patterns for example associated with a pathological state. In particular, we face the multiple testing problem when trying to identify significant effects (e.g., differences between patient groups) using mass-univariate statistics. One the other hand, machine learning models trained to achieve accurate predictions on training data are prone to overfit on test data due to the high dimensionality of these connectomes. To overcome these difficulties, the braindata group develops robust statistical tests and advanced machine learning algorithms. One particular strategy we are pursuing is to design kernel machines with domain-specific low-dimensional similarity metrics to obtain robust predictions.

A crucial aspect when using machine learning in clinical contexts (e.g. when predicting clinical variables from neuroimaging data) is that the physician needs to understand the basis of the algorithm’s decision. We have demonstrated in a widely acknowledged work (Haufe et al., 2014) that the common interpretation of the weights of machine learning models is misleading. In an extreme example this could mislead a neurosurgeon to cut a healthy brain area. We have pointed out a way to obtain correct interpretations from linear machine learning models, and we are working towards explainable state-of-the-art deep learning architectures and kernel machines. 

Since our data analysis pipelines are complex, a core principle of our work is the careful validation of our methods using synthetic and real data. To this end, we have created a simulation environment based on a realistic volume conductor model of the head (Huang et al., 2016, Haufe and Ewald, 2016). We also develop benchmarks for the research community.

We collaborate in-house with the group of Prof. Kühn as well as with the group of Prof. Isaias at UK Würzburg to study alterations of brain connectivity in movement disorders. Currently, we study the disturbance of information flow between cortex and basal ganglia in dystonic patients as well as in patients with Parkinson's disease.

We are also interested in how brain communication is shaped during development, and how it is altered in various psychiatric developmental  diseases such as ADHD. Here, we collaborate with colleagues in Zurich and New York City.

We collaborate with Prof. Schneck of Columbia University on neural correlates of complicated grief in participants who lost a beloved person due to suicide. To address this question we employ fMRI measurement in combination with a multivariate decoding approach (Schneck et al., 2017).

Over the years, we have contributed continuously to the field of brain-computer interfacing (BCI, see, e.g., Blankertz et al., 2011, Vidaurre et al., 2019). The primary goal of BCIs is to enable patients with severe motor deficits to communicate exclusively through their brain activity.