Looking back to move forward: Bridging the gap: From large-scale aggregation to individual prediction

Munich Psychiatry Lecture Series | MPLS

  • Date: Mar 12, 2019
  • Time: 03:00 PM - 04:00 PM (Local Time Germany)
  • Speaker: Dr. Simon B. Eickhoff
  • Institute of Neuroscience and Medicine (INM-7) | Forschungszentrum Jülich
  • Location: Max Planck Institute of Psychiatry
  • Room: Lecture Hall
  • Host: International Max Planck Research School for Translational Psychiatry (IMPRS-TP)
  • Contact: imprs-tp@psych.mpg.de
Looking back to move forward: Bridging the gap: From large-scale aggregation to individual prediction
Over the last two decades, neuroimaging has provided ample knowledge on the structure, function and connectivity of the human brain as well as the aberrations thereof in patients with neurological and psychiatric disorders. In this context, the long predominant paradigm has been to compare (mean) local volume or activity between groups, or to correlate these to behavioral phenotypes. Such approach, however, is intrinsically limited in terms of possible insight into inter-individual differences and application in clinical practice. Recently, the increasing availability of large cohort data and tools for multivariate statistical learning, allowing the prediction of individual cognitive or clinical phenotypes in new subjects, have started a revolution in imaging neuroscience.

The transformation of systems neuroscience into a big data discipline poses a lot of new challenges related to data processing, workflow management and the need for high-performance computing. Yet, one of the most critical aspects is the still sub-optimal relationship between the extremely wide feature-space formed by neuroimaging data and the comparably low number of subjects. In this talk I will argue, however, that this is only true when approaching neuroimaging machine-learning in a naïve fashion, i.e., when ignoring the large body of existing work on human brain mapping. The task-dependent recruitment of distributed networks provide the fundamental basis for cognitive information processing. Importantly, these networks can now be mapped in a highly robust fashion by integrating information on thousands of individual subjects. Such integrated knowledge then can then provide critical a priori information for dimensionality reduction and feature selection, aiding the development of machine-learning models on smaller but deeper characterized datasets.

This approach allows to leverage knowledge on human brain organization for inference on cognitive and social traits in previously unseen individual subjects or the prediction of diagnoses and subtype in individual patients with, e.g., Parkinson's disease or Schizophrenia. Providing a bidirectional translation, such application will in turn provide information on the respective a priori nodes and networks. These developments will thus open up the possibility for a deeper understanding of inter-individual variability and the development of individualized healthcare while at the same time contributing to a better understanding of the human brain.

Go to Editor View