Improved Pre-Surgical Brain Mapping via Unsupervised Learning
When an individual is evaluated for the surgical treatment of epilepsy, a massive amount of brain data is acquired to help the clinical team localize the seizure-generating brain regions as well as regions critical for functions such as language and motor control. These data include weeks of electroencephalographic (EEG) recordings, various functional and structural MRIs, FDG-PET scans, and direct cortical electrical stimulation among others. This mountain of information is conventionally visually analyzed by expert epileptologists in what can be a very time consuming process. I have been working on unsupervised machine learning procedures to help epileptologists more quickly and accurately identify pathological and function-critical brain regions. The figure above illustrates an example of defining functionally distinct cortical areas by grouping together areas that exhibit very similar resting-state fMRI activity. Although this patient was lying still with closed eyes when the fMRI data was acquired, the cluster analysis is able to accurately identify visual and hand sensorimotor areas.
Related papers: Yaffe et al. (2015)
Characterizing the Macro-Scale Functional Architecture of Human Cerebral Cortex
The unique data produced by epilepsy surgery evaluations also provides the most detailed window currently available on human brain function. We can use this window to characterize the large scale functional organization in cortex with remarkable precision. For example, with my colleagues in Ashesh Metha’s Laboratory for Multimodal Human Brain Mapping, I have characterized what types of EEG oscillations dominate various cortical areas (see left). Contrary to what is found in scalp EEG where alpha (8-12 Hz) activity is dominant, theta (4-8 Hz) activity is most prevalent.
Related papers: Groppe et al. (2013)
Understanding Seizure Propagation
The mechanisms via which seizures spread are remarkably poorly understood. In Chris Honey’s lab I have been researching what characteristics of a brain area’s pre-seizure dynamics make it more-or-less likely to be recruited by seizure activity. We hope that this work will lead to improved methods for reducing the spread of seizure activity and thus the impact of seizures on function.
Developing Novel Methods/Software for EEG Analysis
As part of my research, I have been developing novel methods and Matlab software for EEG analysis. For example, I reviewed the use of mass univariate hypothesis testing for EEG event-related potential (ERP) analysis and created a Matlab toolbox to encourage their application with computationally efficient statistical procedure and interactive graphics. My code page provides link to this and other projects.