Professional Documents
Culture Documents
Outline
Brain connectivity networks and real-world complex networks A probabilistic model of the human brain connectom from zero and one to a probabilistic model using network characteristics as potential biomarkers Evaluation and results Discussion and future work
complex networks
Many real-world complex networks including the brain networks have similar topological features: small-world: high clustering coefficient, but relatively low characteristic path length
(Watts and Strogatz, 1998)
a probabilistic model
Existing connectivity studies assume temporal stability of two ROIs, as they compute the correlation coefficient using the entire timecourses We need to consider the temporal fluctuations of the connectivities between ROIs, as hinted by recent neuroscience studies (Stamoulis et. al., 2010 and Whitlow et. al., 2011) Therefore, we propose to consider the frequency of functional connections between brain regions over time and regard the frequent connections as strong and important to the operation of the overall brain network
7
a probabilistic model
s1-2 = 4/8 = 0.5 se(i,j) = |e(i,j)|/|TW| 0 1 0
{
1 0 1 1 0
frequency graph
b) strong-edge graph
a) conventional graph
9
10
evaluation methods
We study the subjects of Major Depression Disorder (MDD) 19 health women with no MDD history in the control group; 19 women with MDD (6 with current MDD, 13 with MDD history) as the diseased group; MDD was characterized with the Structured Clinical Interview for DSM-IV Axis I Disorders (SCID). Both groups underwent 7.2 mins resting-state fMRI scan. We selected 11 ROIs of the default mode network (DMN) based on previous neuroscience studies. PCC, rPG, lPG, SCC, rPC, lPC, rITG, lITG, rSFG, lSFG, MPC
11
network measures
For each subject's brain connectivity network, we compute the following network indices: degree (or strength in a weighted graph), betweenness centrality, closeness centrality, local and global clustering coefficient, and global characteristic path length.
12
SVM classifiers
We use the network features to build Support Vector Machines (SVMs) for both the conventional and our probabilistic model. Standard SVM techniques have been considered: scaling, grid-based kernel parameter search and feature selection using F-test. Evaluation methods: Bootstrap the sample 1000 times and during each iteration: 2/3 is used for training, and the rest1/3 is kept for testing. We measure the prediction accuracy and the ROC area under the curve (AUC) at each iteration and take the average as the final result.
13
results
vs.
b) the strong-edge model a) the conventional model
(Accuracy: 76%, AUC: 0.87)
14
discussion
A number of free variables can influence the accuracy of brain network models and consequently affect the classifiers' performance. In the conventional method: the choice of the correlation coefficient threshold T that determines whether we shall draw an edge between two voxels (ROIs); determined by the targeted density, and we choose (0.37 ~ 0.5) followed existing studies. In the strong-edge (probabilistic) model: the choice of the density value is relaxed and we can choose a high density value (0.76); however, we introduce the Sminsup that determines the cut-off frequency for defining an edge to be strong (less important for the classifier)
15
16
Q & A?
jbian@uams.edu
17