Abstract. This paper presents a Bayesian method for constructing probabilistic networks from databases. In particular, we focus on constructing Bayesian belief networks. Potential applications include computer-assisted hypothesis testing, automated scientific discovery, and automated construction of probabifistic expert systems. We extend the basic method to handle missing data and hidden (latent) variables. We show how to perform probahilistic inference by averaging over the inferences of multiple belief networks. Results are presented of a preliminary evaluation of an algorithm for constructing a belief network from a database of cases. Finally, we relate the methods in this paper to previous work, and we discuss open problems.Keywords. probabilistic networks, Bayesian belief networks, machine learning, induction I n t r o d u c t i o nIn this paper, we present a Bayesian method for constructing a probabilistic network from a database of records, which we call cases. Once constructed, such a network can provide insight into probabilistic dependencies that exist among the variables in the database. One application is the automated discovery of dependency relationships. The computer program searches for a probabilistic-network structure that has a high posterior probability given the database, and outputs the structure and its probability. A related task is computer-assisted hypothesis testing: The user enters a hypothetical structure of the dependency relationships among a set of variables, and the program calculates the probability of the structure given a database of cases on the variables.We can also construct a network and use it for computer-based diagnosis. For example, suppose we have a database in which a case contains data about the behavior of some system (i.e., findings). Suppose further that a case contains data about whether this particular behavior follows from proper system operation, or alternatively, is caused by one of several possible faults. Assume that the database contains many such cases from previous episodes of proper and faulty behavior. The method that we present in this paper can be used to construct from the database a probabilistic network that captures the probabilistic dependencies among findings and faults. Such a network then can be applied to classify future cases of system behavior by assigning a posterior probability to each of the possible faults and to the event "proper system operation." In this paper, we also shall discuss diagnostic inference that is based on combining the inferences of multiple alternative networks.

The site of lesion responsible for left hemispatial neglect after stroke has been intensely debated recently. Some studies provide evidence that right angular lesions are most likely to cause left neglect, whereas others indicate that right superior temporal lesions are most likely to cause neglect. We examine two potential accounts of the conflicting results: (1) neglect could result from cortical dysfunction beyond the structural lesion in some studies; and (2) different forms of neglect with separate neural correlates have been included in different proportions in separate studies. To evaluate these proposals, we studied 50 patients with acute right subcortical infarcts using tests of hemispatial neglect and magnetic resonance diffusion-weighted and perfusion-weighted imaging performed within 48 h of onset of symptoms. Left "allocentric" neglect (errors on the left sides of individual stimuli, regardless of location with respect to the viewer) was most strongly associated with hypoperfusion of right superior temporal gyrus (Fisher's exact test; p Ͻ 0.0001), whereas left "egocentric" neglect (errors on the left of the viewer) was most strongly associated with hypoperfusion of the right angular gyrus ( p Ͻ 0.0001). Patients without cortical hypoperfusion showed no hemispatial neglect. Because the patients did not have cortical infarcts, our data show that neglect can be caused by hypoperfused dysfunctional tissue not detectable by structural magnetic resonance imaging. Moreover, different forms of neglect were associated with different sites of cortical hypoperfusion. Results help explain conflicting results in the literature and contribute to the understanding of spatial attention and representation in the human brain.

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 volumebased 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.Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers

There is evidence for different levels of visuospatial processing with their own frames of reference: viewer-centered, stimulus-centered, and object-centered. The neural locus of these levels can be explored by examining lesion location in subjects with unilateral spatial neglect (USN) manifest in these reference frames. Most studies regarding the neural locus of USN have treated it as a homogenous syndrome, resulting in conflicting results. In order to further explore the neural locus of visuospatial processes differentiated by frame of reference, we presented a battery of tests to 171 subjects within 48 hr after right supratentorial ischemic stroke before possible structural and/or functional reorganization. The battery included MR perfusion weighted imaging (which shows hypoperfused regions that may be dysfunctional), diffusion weighted imaging (which reveals areas of infarct or dense ischemia shortly after stroke onset), and tests designed to disambiguate between various types of neglect. Results were consistent with a dorsal/ventral stream distinction in egocentric/allocentric processing. We provide evidence that portions of the dorsal stream of visual processing, including the right supramarginal gyrus, are involved in spatial encoding in egocentric coordinates, whereas parts of the ventral stream (including the posterior inferior temporal gyrus) are involved in allocentric encoding.

This paper presents a deformable model for automatically segmenting brain structures from volumetric magnetic resonance (MR) images and obtaining point correspondences, using geometric and statistical information in a hierarchical scheme. Geometric information is embedded into the model via a set of affine-invariant attribute vectors, each of which characterizes the geometric structure around a point of the model from a local to a global scale. The attribute vectors, in conjunction with the deformation mechanism of the model, warranty that the model not only deforms to nearby edges, as is customary in most deformable surface models, but also that it determines point correspondences based on geometric similarity at different scales. The proposed model is adaptive in that it initially focuses on the most reliable structures of interest, and gradually shifts focus to other structures as those become closer to their respective targets and, therefore, more reliable. The proposed techniques have been used to segment boundaries of the ventricles, the caudate nucleus, and the lenticular nucleus from volumetric MR images.