Spatially-informed Disease Prediction with Structured Sparse Support Vector Machine

Background — Human Connectome and Brain Disorders

Advanced neuroimaging methods such as diffusion and functional MRI (fMRI) have demonstrated the human brain operates as a large-scale complex network known as the "human connectome." Abundant amount of research evidences suggest that many major brain disorders are linked with aberrations in the network topology of the brain, which has led to a large interest in establishing an objective, connectivity-based biomarker of psychiatric disorders using neuroimaging techniques.

Spatially-informed classifier with Structured Sparsity

Motivated by these findings, this work focuses on the supervised learning problem of binary classification, where the goal is to predict the psychiatric disorder status of an individual using functional connectomes (FC's), which are high-dimensional correlation maps derived from resting-state fMRI. In brief, FC's are generated by parcellating the brain into hundreds of distinct regions, and computing cross-correlation matrices across time.

Construction of the functional connectomes (FC's) using the time-series signal from resting-state fMRI. The lower-triangular portion of the correlation-matrix serves as the feature vector used for classification.

However, FC's are known to be quite noisy and also reside in a high dimensional space, which introduces critical statistical and computational challenges and also complicates model interpretation. In order to design a robust classifier that reflects the underlying biological mechanism of the disorder of interest, it is important to account for any structure in the data that are known a priori. In this work, we introduce a structured sparse variant of the well-known Support Vector Machine (SVM) classifier using regularization techniques. Specifically, in contrast to existing approaches, our method explicitly accounts for the 6-D spatial structure of the FC's (defined by pairs of points in 3-D space) using either the GraphNet or the fused Lasso.

Experiments on a large schizophrenia dataset demonstrates the utility of our proposed method, where we demontrate that our method can achieve reliable classification accuracy (72%) and also recover systematic network patterns, as illustrated in the figure below. Notice the prominent involvement of lateral prefrontal regions in connections within frontoparietal network and in connections between frontoparietal network and default network. Interestingly, reduced connectivity between these two networks is among the most commonly observed findings in connectivity research in schizophrenia.

Relevant Publications

T. Watanabe, D. Kessler, C. Scott, M. Angstadt, C. Sripada, "Disease Prediction based on Functional Connectomes using a Scalable and Spatially-Informed Support Vector Machine,'' NeuroImage, vol. 96, no. 4, pp. 183-202, 2014. [ Arxiv] [ code]

T. Watanabe, C. Scott, D. Kessler, M. Angstadt, C. Sripada, "Scalable Fused Lasso SVM for Connectome-based Disease Classification,'' IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), Florence, Italy, 2014.

C. Sripada, D. Kessler, T. Watanabe, R. Welsh, Y. Fang, M. Angstadt, S. Taylor, C. Scott, "Whole-brain connectomic analysis of 145 resting state scans reveals network neurosignatures of schizophrenia,'' Biological Psychiatry, vol. 73, pp. 37-38S, 2013.