gms | German Medical Science

GMDS 2013: 58. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e. V. (GMDS)

Deutsche Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie

01. - 05.09.2013, Lübeck

Advances and caveats in machine learning approacheson resting-state fMRI connectomes

Meeting Abstract

Search Medline for

  • Norman Scheel - University of Lübeck, Institute for Neuro- and Bioinformatic, Lübeck, DE
  • Amir Madany - University of Lübeck, Institute for Neuro- and Bioinformatic, Lübeck, DE

GMDS 2013. 58. Jahrestagung der Deutschen Gesellschaft für Medizinische Informatik, Biometrie und Epidemiologie e.V. (GMDS). Lübeck, 01.-05.09.2013. Düsseldorf: German Medical Science GMS Publishing House; 2013. DocAbstr.313

doi: 10.3205/13gmds313, urn:nbn:de:0183-13gmds3136

Published: August 27, 2013

© 2013 Scheel et al.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License ( You are free: to Share – to copy, distribute and transmit the work, provided the original author and source are credited.



Introduction: Blood oxygenation level dependent (BOLD) fluctuations measured by functional magnetic resonance imaging (fMRI) have been studied in various ways. As neuronal activity is mainly responsible for these fluctuations [1] it allows us to investigate the internal functions of the brain. Especially resting state fMRI has drawn more and more attention since its introduction in [2] as task based experiments are rather artificial and limited in their applications. Inferring information from the fMRI of the mind wandering brain may momentarily be the best way to grasp the complex processes of brain-function [3], [4], [5], [6]. One main approach to study resting-state brain connectivity are correlational matrices based on brain segmentations using diverse brain atlases or points of interest (POI's) and different correlation measures. These functional connectivity matrices can be interpreted in the way that they describe the functional interactions of brain regions. For further investigation of these matrices machine learning is recently emerging into the field of neuroscience.

Methods and Material: Already for task fMRI studies, machine learning, esp. Support Vector Machines (SVM's) have proven to be of great value [7], [8]. With the ADHD-200 challenge and the 1000 functional connectomes project [9] machine learning and pattern classification on resting state fMRI became very popular in clinical circumstances.Unfortunately, the complexity of the data as well as missing constraints in the experimental design leads to many problems and challenges for multivariate learning approaches. Here we want to discuss common culprits using a representative open access data set kindly provided by the Beijing Normal University, State Key Laboratory of Cognitive Neuroscience and Learning through the INDI 1000 Functional Connectomes Project. In their study they measured 44 students, age 23 ± 2.2 years, 22 female, 22 male, to assess the differences in resting state fMRI of subjects having their eyes open vs. eyes closed.

Results: Using Support Vector Machines we were able to find significant differences between eyes open and eyes closed conditions as well as differences between females and males. Further analysis then showed that most differences are more likely to be associated with subject motion as well as scanner session effects as certain long-range connections disappear when subject motion is ruled out before classification. Another phenomenon is the appearance of negative significant results, where cross-validation (CV) accuracies drop significantly below chance level.

Discussion: Confounds correlated with the group assignment may lead to false positive classifications and just as negative significance it indicates how sensitive multivariate learning is in high-dimensional low-sample size scenarios. As it is uncertain which aspects of data-set generation and preprocessing have an influence on the classification accuracies, there is only the possibility of reverse engineering to rule out as many confounding effects as possible. Even then one might not discover the “true” cause. On the example of the above data, we will discuss the potential advantages of multivariate rs-fMRI data analysis just as well as we will elucidate the caveats regarding interpretation and generalization of such results.


Heeger DJ, Ress D. What does fMRI tell us about neuronal activity? Nature reviews Neuroscience. 2002 Feb;3(2):142–51.
Biswal BB, Yetkin FZ, Haughton VM, Hyde JS. Functional connectivity in the motor cortex of resting human brain using echo-planar MRI. Magnetic Resonance in Medicine. 1995;34(4):537–541.
Fox MD, Snyder AZ, Vincent JL, Corbetta M, Van Essen DC, Raichle ME. The human brain is intrinsically organized into dynamic, anticorrelated functional networks. Proc Natl Acad Sci U S A. 2005 Jul;102(27):9673–9678.
Fox MD, Raichle ME. Spontaneous fluctuations in brain activity observed with functional magnetic resonance imaging. Nat Rev Neurosci. 2007 Sep;8(9):700–711.
Raichle ME. The Brain’s Dark Energy. Scientific American. 2010;302(3):44–49.
Zhang D, Raichle ME. Disease and the brain’s dark energy. Nature reviews. Neurology. 2010 Jan;6(1):15–28.
Mourão-Miranda J, Reynaud E, McGlone F, Calvert G, Brammer M. The impact of temporal compression and space selection on SVM analysis of single-subject and multi-subject fMRI data. Neuroimage. 2006;33(4):1055–1065.
Mourão-Miranda J, Friston KJ, Brammer M. Dynamic discrimination analysis: a spatial-temporal SVM”. Neuroimage. 2007;36(1):88–99.
HD-200 Consortium. The ADHD-200 Consortium: A Model to Advance the Translational Potential of Neuroimaging in Clinical Neuroscience. Front Syst Neurosci. 2012 Sep 5;6:62.