Dynamic causal modelling and model-based multivariate analyses

Decoding models, such as those underlying multivariate classification algorithms, have been increasingly used to infer cognitive or clinical brain states from measures of neural activity obtained by fMRI. Such classifiers have great potential in predicting the clinical diagnosis of an individual subject from functional brain data. The practicality of current classifiers, however, is restricted. First, selecting the most informative features from high-dimensional fMRI data is difficult, often leading to poor generalization performance. Second, popular methods such as support vector machines (SVMs) rarely afford mechanistic interpretability. We address both issues using a generative-embedding approach that incorporates neurobiologically interpretable generative models into discriminative classifiers.

Current publications

- K.H. Brodersen, T.M. Schofield, A.P. Leff, C.S. Ong, E.I. Lomakina, J.M. Buhmann, K.E. Stephan (2011). Generative embedding for model-based classification of fMRI data. PLoS Computational Biology, 7(6): e1002079.

- K.H. Brodersen, F. Haiss, C.S. Ong, F. Jung, M. Tittgemeyer, J.M. Buhmann, B. Weber, K.E. Stephan (2011). Model-based feature construction for multivariate decoding. NeuroImage, 56, 601-615.

- K.H. Brodersen, T.M. Schofield, A.P. Leff, C.S. Ong, E.I. Lomakina, J.M. Buhmann, K.E. Stephan (2011). Generative embedding for model-based classification of fMRI data. Oral presentation at Human Brain Mapping 2011, Quebec City, Canada. Awarded with a Trainee Abstract Award.

- K.H. Brodersen, F. Haiss, C.S. Ong, F. Jung, P. Allen, M. Tittgemeyer, J.M. Buhmann, P. McGuire, B. Weber, K.E. Stephan (2010). Model-based multivariate decoding and model selection. Presented at Human Brain Mapping, Barcelona, Spain. Awarded with a Trainee Abstract Award.

Gaussian processes for whole-brain feature selection and classification

Multivariate decoding approaches in neuroimaging attempt to infer hidden labels from distributed measures of brain activity. For functional magnetic resonance imaging (fMRI) in particular, multiple multivariate classification approaches have been proposed. Due to the extremely high dimensionality of fMRI data the key challenge for classification methods, with no satisfactory solutions so far, is feature selection, i.e., the question of how to identify jointly informative voxels while ignoring uninformative noise. We address this challenge by combining Gaussian processes with parametric permutation tests on feature weights. We illustrate the utility of our approach by applying it to fMRI data from decision-making paradigms.

Current publications

- E.I. Lomakina, K.H. Brodersen, T.E.J. Behrens, K.E. Stephan, J.M. Buhmann (2011). Gaussian processes for whole-brain feature selection and classification in fMRI. Interactive session at Human Brain Mapping 2011, Quebec City, Canada.

- K.H. Brodersen, C.-S. Lin, E.I. Lomakina, K.E. Stephan, K. Wiech, I. Tracey (2011). Multivariate decoding of perceptual decisions about pain. Interactive session at Human Brain Mapping 2011, Quebec City, Canada.

Performance evaluation in hierarchical classification settings

In neuroimaging, multivariate classification algorithms are used to predict cognitive or pathophysiological states from measurements of distributed brain activity. This idea has great potential for the design of brain-computer interfaces, in clinical diagnostics, and for cognitive neuroscience. A common index of how much information can be decoded from a particular observation of brain activity is the classification accuracy. The key statistical question is how the accuracy obtained for a particular group of subjects generalizes to the wider population. While procedures for this mixed-effects inference are well-established for mass-univariate analyses, multivariate classification has been limited to fixed-effects inference so far. We work towards the introduction of a full mixed-effects analysis for classification in group studies that resolves these limitations.

Current publications

- K.H. Brodersen, J.R. Chumbley, C. Mathys, J. Daunizeau, J.M. Buhmann, K.E. Stephan (2011). Mixed-effects inference on classification performance in group studies. Interactive session at Human Brain Mapping 2011, Quebec City, Canada.