Rong Chen

Predictive models of autism spectrum disorder based on brain regional cortical thickness

Autism spectrum disorder (ASD) is a neurodevelopmental disorder with a wide phenotypic range, often affecting personality and communication. Previous voxel-based morphometry (VBM) studies of ASD have identified both gray- and white-matter volume changes. However, the cerebral cortex is a 2-D sheet with a highly folded and curved geometry, which VBM cannot directly measure. Surface-based morphometry (SBM) has the advantage of being able to measure cortical surface features, such as thickness. The goals of this study were twofold: to construct diagnostic models for ASD, based on regional thickness measurements extracted from SBM, and to compare these models to diagnostic models based on volumetric morphometry. Our study included 22 subjects with ASD (mean age 9.2+/-2.1 years) and 16 volunteer controls (mean age 10.0+/-1.9 years). Using SBM, we obtained regional cortical thicknesses for 66 brain structures for each subject. In addition, we obtained volumes for the same 66 structures for these subjects. To generate diagnostic models, we employed four machine-learning techniques: support vector machines (SVMs), multilayer perceptrons (MLPs), functional trees (FTs), and logistic model trees (LMTs). We found that thickness-based diagnostic models were superior to those based on regional volumes. For thickness-based classification, LMT achieved the best classification performance, with accuracy=87%, area under the receiver operating characteristic (ROC) curve (AUC)=0.93, sensitivity=95%, and specificity=75%. For volume-based classification, LMT achieved the highest accuracy, with accuracy=74%, AUC=0.77, sensitivity=77%, and specificity=69%. The thickness-based diagnostic model generated by LMT included 7 structures. Relative to controls, children with ASD had decreased cortical thickness in the left and right pars triangularis, left medial orbitofrontal gyrus, left parahippocampal gyrus, and left frontal pole, and increased cortical thickness in the left caudal anterior cingulate and left precuneus. Overall, thickness-based classification outperformed volume-based classification across a variety of classification methods.

Collective Mining of Bayesian Networks from Distributed Heterogeneous Data

We present a collective approach to learning a Bayesian network from distributed heterogeneous data. In this approach, we first learn a local Bayesian network at each site using the local data. Then each site identifies the observations that are most likely to be evidence of coupling between local and non-local variables and transmits a subset of these observations to a central site. Another Bayesian network is learnt at the central site using the data transmitted from the local site. The local and central Bayesian networks are combined to obtain a collective Bayesian network, which models the entire data. Experimental results and theoretical justification that demonstrate the feasibility of our approach are presented.

Structural MRI in Autism Spectrum Disorder

Magnetic resonance (MR) examination provides a powerful tool for investigating brain structural changes in children with autism spectrum disorder (ASD). We review recent advances in the understanding of structural MR correlates of ASD. We summarize findings from studies based on voxel-based morphometry, surface-based morphometry, tensor-based morphometry, and diffusion-tensor imaging. Finally, we discuss diagnostic models of ASD based on MR-derived features.

Spatially Compact Neural Clusters in the Dorsal Striatum Encode Locomotion Relevant Information

An influential striatal model postulates that neural activities in the striatal direct and indirect pathways promote and inhibit movement, respectively. Normal behavior requires coordinated activity in the direct pathway to facilitate intended locomotion and indirect pathway to inhibit unwanted locomotion. In this striatal model, neuronal population activity is assumed to encode locomotion relevant information. Here, we propose a novel encoding mechanism for the dorsal striatum. We identified spatially compact neural clusters in both the direct and indirect pathways. Detailed characterization revealed similar cluster organization between the direct and indirect pathways, and cluster activities from both pathways were correlated with mouse locomotion velocities. Using machine-learning algorithms, cluster activities could be used to decode locomotion relevant behavioral states and locomotion velocity. We propose that neural clusters in the dorsal striatum encode locomotion relevant information and that coordinated activities of direct and indirect pathway neural clusters are required for normal striatal controlled behavior. VIDEO ABSTRACT.

Network analysis of mild cognitive impairment.

We present a network analysis of a cross-sectional study of mild cognitive impairment (MCI). Network analysis, as opposed to univariate analysis, accounts for interactions among brain structures in explaining a clinical outcome. In this context, we analyze structural magnetic resonance (MR) data based on a Bayesian network representation of variables in the problem domain. The Bayesian network resulting from this analysis reveals complex, nonlinear multivariate associations among morphological changes in the left hippocampus and in the right thalamus and the presence of mild cognitive impairment. This Bayesian network could be used to predict the presence of mild cognitive impairment from structural MR scans.

Distributed Web mining using Bayesian networks from multiple data streams

We present a collective approach to mining Bayesian networks from distributed heterogenous Web-log data streams. In this approach we first learn a local Bayesian network at each site using the local data. Then each site identifies the observations that are most likely to be evidence of coupling between local and non-local variables and transmits a subset of these observations to a central site. Another Bayesian network is learnt at the central site using the data transmitted from the local site. The local and central Bayesian networks are combined to obtain a collective Bayesian network that models the entire data. We applied this technique to mining multiple data streams, where data centralization is difficult because of large response time and scalability issues. Experimental results and theoretical justification that demonstrate the feasibility of our approach are presented.

Machine-learning techniques for building a diagnostic model for very mild dementia.

Many researchers have sought to construct diagnostic models to differentiate individuals with very mild dementia (VMD) from healthy elderly people, based on structural magnetic-resonance (MR) images. These models have, for the most part, been based on discriminant analysis or logistic regression, with few reports of alternative approaches. To determine the relative strengths of different approaches to analyzing structural MR data to distinguish people with VMD from normal elderly control subjects, we evaluated seven different classification approaches, each of which we used to generate a diagnostic model from a training data set acquired from 83 subjects (33 VMD and 50 control). We then evaluated each diagnostic model using an independent data set acquired from 30 subjects (13 VMD and 17 controls). We found that there were significant performance differences across these seven diagnostic models. Relative to the diagnostic models generated by discriminant analysis and logistic regression, the diagnostic models generated by other high-performance diagnostic-model-generation algorithms manifested increased generalizability when diagnostic models were generated from all atlas structures.

Graphical-model-based morphometric analysis

We propose a novel method for voxel-based morphometry (VBM), which we call graphical-model-based morphometric analysis (GAMMA), to identify morphological abnormalities automatically, and to find complex probabilistic associations among voxels in magnetic-resonance images and clinical variables. GAMMA is a fully automatic, nonparametric morphometric-analysis algorithm, with high sensitivity and specificity. It uses a Bayesian network to represent the associations among voxels and the function variable, and uses a contextual-clustering method based on a Markov random field to find clusters in which all voxels have similar associations with the function variable. We use loopy belief propagation to infer the unobserved label field and belief map. As opposed to voxel-based morphometric methods based on general linear models, GAMMA is capable of identifying nonlinear associations among the function variable and voxels. Compared with our previous approach, a Bayesian morphometry algorithm, GAMMA has greater sensitivity, specificity, and computational efficiency.

Dynamic Bayesian network modeling for longitudinal brain morphometry.

Identifying interactions among brain regions from structural magnetic-resonance images presents one of the major challenges in computational neuroanatomy. We propose a Bayesian data-mining approach to the detection of longitudinal morphological changes in the human brain. Our method uses a dynamic Bayesian network to represent evolving inter-regional dependencies. The major advantage of dynamic Bayesian network modeling is that it can represent complicated interactions among temporal processes. We validated our approach by analyzing a simulated atrophy study, and found that this approach requires only a small number of samples to detect the ground-truth temporal model. We further applied dynamic Bayesian network modeling to a longitudinal study of normal aging and mild cognitive impairment--the Baltimore Longitudinal Study of Aging. We found that interactions among regional volume-change rates for the mild cognitive impairment group are different from those for the normal-aging group.

Brain Morphometry and Intelligence Quotient Measurements in Children With Sickle Cell Disease

Objective: To verify the hypothesis that volume of regional gray matter accounts substantially for variability in intelligence quotient (IQ) score among children with sickle cell disease, who have no magnetic resonance visible infarcts. Methods: We studied 31 children with sickle cell disease, homozygous for hemoglobin S, with no history of stroke, no magnetic resonance signal-intensity abnormality, and transcranial Doppler velocities <170 cm/sec, with a T1-weighted magnetic resonance sequence and the Kaufman Brief Intelligence Test. On the basis of Kaufman Brief Intelligence Test, we classified these children into 2 groups: high and low IQ based on a median split. We then used an automated and novel Bayesian voxel-based morphometry technique, called Graphical-Model-Based Multivariate Analysis (GAMMA), to assess the probabilistic association between IQ score and regional gray matter volume. Results: GAMMA found 1 region linking low IQ with smaller cortical gray matter volume. In comparison with the children in the high-IQ group, children in the low-IQ group had smaller regional gray matter volume in both frontal lobes, both temporal lobes, and both parietal lobes. Conclusions: In children with sickle cell disease, we found a linear association between IQ and regional gray matter volume. This finding suggests that some variance in intellectual ability in children with sickle cell disease is accounted for by regional variability of gray matter volume, which is independent of neuroradiological evidence of infarct.