Network-based Models for Mental State Prediction: from Functional Networks to Sparse MRFs
Irina Rish
Biomedical Computing Department
Computational Biology Center
IBM T. J. Watson Research Center


Abstract:

Traditional fMRI analysis is often focused on discovering spatial activation patterns using simple univariate linear approach (GLM) that treats brain voxels as independent variables. However, there is also a growing consensus in the field that such``phrenological'' approach can be somewhat limited; an alternative is to explore non-local, distributed activation patterns, by treating brain as an interaction network.

I will present our recent work on network-based approaches applied to several fMRI-related studies. First, I will discuss the importance of topological information extracted from brain's functional (correlation) networks for building predictive models of schizophrenia, a complex psychiatric disorder that has eluded a characterization in terms of local abnormalities of brain activity, and is hypothesized to affect the collective, ``emergent'' working of the brain. Our findings demonstrate clear advantage of topological features over both traditional region-of-interest (ROI) and local, task-specific linear activation analyzes, and suggest that schizophrenia is indeed associated with disruption of global brain properties related to its functioning as a network, which cannot be explained just by alteration of local activation patterns. Moreover, further exploitation of voxel interactions by sparse Markov Random Field (MRF) classifiers consistently outperforms linear methods, such as Gaussian Naive Bayes and SVM, attaining 88% accuracy (over about 55% random guess baseline), which is quite remarkable given that it is based on a single fMRI experiment using a simple auditory task.

In the second part of this talk, I will elaborate more on our approach to learning sparse Markov Random Field, and provide more examples of its applications to fMRI analysis, such as 93-95% accurate prediction of mental state associated with viewing a picture vs. a sentence based on the data from (Mitchell, 2004). Our method is a simple but effective greedy algorithm, called SINCO, for Sparse INverse COvariance selection, which is equivalent to structure recovery of a Markov Network over Gaussian variables. We demonstrate advantages of our method over several state-of-art approaches such as glasso (Friedman et al, 2007) and COVSEL (Banerjee et al, 2006), showing that in some regimes SINCO can be faster and more accurate (in terms of network structure reconstruction). Our method has an additional advantage of being easily parallelizable. The problem of regularization parameter(s) selection is addressed in a Bayesian way, by assuming a prior on the parameter(s) and by using MAP optimization to find both the inverse covariance matrix and the unknown parameters. Our general formulation extends prior art by allowing a vector of regularization parameters and is well-suited for learning structured graphs where the sparsity of nodes varies significantly (e.g., scale-free graphs often encountered in practice).


Speaker's Bio:

Irina Rish is a research staff member at the Biomedical Computing Department which is a part of the Computational Biology Center at the IBM T. J. Watson Research Center. She received an M.S. in applied mathematics from Moscow Gubkin Institute, Russia, and a Ph.D. in computer science from the University of California, Irvine. Dr. Rish's primary research interests are in the areas of probabilistic inference, statistical learning, and information theory, and their applications to large-scale data analysis problems in biology and neuroscience. Her current research focuses on applying machine-learning techniques to neuroscience, and particularly on statistical analysis of fMRI data using sparse regression, dimensionality reduction and graphical models. In the past, she has worked on efficient approximations of probabilistic inference in Bayesian networks, probabilistic diagnosis and experiment design, active learning, collaborative prediction, sparse regression and sparse matrix factorization, and their applications to autonomic computing, as a part of the Adventurous Research project on Self-Managing Computer Systems that she lead at IBM Watson (2003-2007). She has over 40 conference and journal publications on the above topics. Dr. Rish taught several machine learning courses at the Electrical Engineering and Computer Science departments of Columbia University as an adjunct professor, and co-organized several machine-learning workshops at ICML, ECML and NIPS conferences.