Heterogeneous Data Mining for Brain Disorder Identiﬁcationbcao1/doc/WCP_slides.pdf · • Brain Network Analysis • Multi-view Feature Analysis • Future Work 2 Davidson et al

Heterogeneous Data Mining for Brain Disorder Identification

Bokai Cao 04/07/2015

Outline• Introduction

• Tensor Imaging Analysis

• Brain Network Analysis

• Multi-view Feature Analysis

• Future Work

2

Davidson et al. Network discovery via constrained tensor analysis of fMRI data. In KDD, 2013.

Wee et al. Identification of MCI individuals using structural and functional connectivity networks. Neuroimage, 2012.

Ye et al. Heterogeneous data fusion for Alzheimer’s disease study. In KDD, 2008.





• Future Work

3

Introduction• Brain Disorders

• HIV, AD, ADHD

• Neuroimaging Techniques

• fMRI, DTI, PET, EEG

• Data Representations

• Tensor, graph, vector

4

What is the data about?

Where does the data come from?

What does the data look like?

Introduction• Brain Disorders

• HIV infection on brain

• Bipolar disorder

• Alzheimer's disease (AD)

• Attention deficit hyperactivity disorder (ADHD)

• Schizophrenia

5

Introduction• Neuroimaging Techniques

• Functional magnetic resonance imaging (fMRI)

• Diffusion tensor imaging (DTI)

• Positron emission tomography (PET)

• Electroencephalogram (EEG)

6

Introduction• Data Representations

• Tensor: raw images

• Graph: brain networks

• Vector: multi-view features

7

View3

View2

View1





• Future Work

8

Tensor Imaging Analysis• Tensor Data in Neuroimaging

9

vector matrix tensor1st-order

tensor2nd-order

tensor3rd-order

tensor

Xx X

A voxel is the smallest three-dimensional point volume referenced in a neuroimaging of the brain. Typically, 256*256*256.

Tensor Imaging Analysis• Brain Network Discovery

1. Node discovery

2. Edge discovery

3. Network verification

10


• Brain Network Discovery

• Tensor factorization

Tensor Imaging Analysis

11

≈ + ⋯+ +a1

c1b1

c2b2

a2

cRbR

aR X


≈ + ⋯+ +a1

c1b1

c2b2

a2

cRbR

aR Xanatomical structures

fMRI data

constraints

Tensor Imaging Analysis• Brain Network Discovery

• Contributions: (1) simultaneously discover nodes and edges (2) facilitate better understanding about interaction mechanism between brain regions

• Drawback: fail to leverage relationships between neuroimages and their associated labels

12


• Supervised Tensor Learning

• Classification (review)

features: <age, weight>

label: diseased (positive) or healthy (negative)


13

He et al. DuSK: A dual structure-preserving kernel for supervised tensor learning with applications to neuroimages. In SDM, 2014.

+ ++

++-

- --

-

-

age

weight?

?


• Tensor classification

1. High dimensionality


14


+ ++

++-

- --

-

-

age

weight

m=256*256*256=16,777,216

x1

x2

x3

xm

.

.

.




2. Structural complexity


15


+ ++

++-

- --

-

-




2. Structural complexity

3. Nonlinear separability


16


+ ++

++-

- --

-

- +-


• Support tensor machine


17

max↵1,··· ,↵n

nX

i=1

↵i �1

2

nX

i,j=1

↵i↵jyiyj h�(Xi),�(Xj)i

s.t.

nX

i=1

↵iyi = 0

0 ↵i C, 8i = 1, · · · , n.

where ↵i are the Lagrangian multipliers and h�(Xi),�(Xj)i are the inner productbetween the mapped tensors of Xi and Xj in the Hilbert space.

minW,b,⇠

1

2kWk2F + C

nX

i=1

⇠i

s.t. yi(hW,Xii+ b) � 1� ⇠i

⇠i � 0, 8i = 1, · · · , n.

where W can be regarded as the weight tensor of the separating hyperplanein the tensor product space RI1⇥···⇥Im , b is the bias, ⇠i is the error of the i-thtraining sample, and C is the trade-o↵ between the margin and empirical loss.The above optimization problem is a generalization of the standard SupportVector Machine (SVM) from vector data to tensor data.

primary form dual form


Tensor Imaging Analysis• Tensor Data in Neuroimaging


• Node discovery, edge discovery, network verification, tensor factorization


• Tensor classification, support tensor machine

18





• Future Work

19

Brain Network Analysis• Brain Network Data

20

• Nodes: brain regions e.g., insula, hippocampus, thalamus

• fMRI links: correlations between the functional activities of brain regions DTI links: number of neural fibers connecting different brain regions

Graph

• Kernel Learning on Graphs

Brain Network Analysis

21

+ ++

++-

- --

-

-

?

?


how to ?

Brain Network Analysis• Kernel Learning on Graphs

1. Extracting clustering coefficients

2. Selecting the most discriminative features

3. Constructing kernel matrices

4. Training classifiers

22



23

clustering coefficient of region 1


+ ++

++-

- --

-

-

.

.

.

clustering coefficient of region 2 clustering

coefficient of region 3

clustering coefficient of region m

classifiers


• Contributions: (1) integrate complementary DTI and fMRI (2) achieve 96% classification accuracy (3) select the most discriminative brain regions

• Drawbacks: (1) connectivity structures are blinded (2) interpretability is limited

24


Brain Network Analysis• Subgraph Pattern Mining

25

Kong et al. Discriminative feature selection for uncertain graph classification. In SDM, 2013.

+ Alzheimer's disease + Alzheimer's disease + Alzheimer's disease

- Normal - Normal - Normal

A discriminative subgraph pattern


• Binary links: {0,1} e.g. gSpan [Yan and Han 2002]

• fMRI links: [-1,1]

• DTI links: N (0,1,2,…,1e6,…)

26



• Mining uncertain graphs

27


Brain Network Analysis• Brain Network Data

• Kernel Learning on Graphs

• Extracting clustering coefficients

• Subgraph Pattern Mining

• Mining uncertain graphs

28





• Future Work

29

Multi-view Feature Analysis• Multi-view Data in Medical Studies

30

HIV/seronegative

View 1

View 3

View 2

View 6

View 4

View 5

Immunologic measures

Clinical measuresSerologic measures

MRI sequence B

MRI sequence A Cognitive measures

limited subjects available yet introducing a large number of measurements

multi-view learning + feature selection

Multi-view Feature Analysis• Modeling View Correlations

• Tensor and AAL features from MRI images

• Demographic information: age and gender

• Genetic information

31



32


++ +

++

--- ---

++

++

+

--- - -- + +

+

++-

- --

-

-

x2(1)

x1(1)x2(2)

x1(2)

x2(1)

x1(1)

x1(2)

x2(2)

concatenation


33


++ +

++

--- ---

++

++

+

--- - --

x2(1)

x1(1)x2(2)

x1(2)

multi-kernel learning

α1

α 2


• Contributions: (1) integrate different types of features (2) identify biomarkers (brain regions) from multiple data sources

• Drawbacks: fail to explicitly consider correlations between features

34


Multi-view Feature Analysis• Modeling Feature Correlations

• Vector-based method

35

View3

View2

View1 modeling feature selection

multi-view data

neglect multi-view knowledge accuracy dropsinformation loss

Cao et al. Tensor-based multi-view feature selection with applications to brain diseases. In ICDM, 2014.


• Tensor-based method

36

View3

View2


multi-view data

suffer from redundancy and irrelevance accuracy dropsnoise



• Dual method

37

View3

View2


multi-view data



• Wrapper-based feature selection

38

Feature selection

Classification algorithm

Feature set Hypothesis

Classification algorithm

Training set

Classification evaluation

Feature set

Test set

Training set

Estimated performance



39

View 1

View 2

View 3

xi(1)

xi(2)

xi(3)

Xi W

W:,...,:,iv ,:,...,: Xi

*

tensor productsupport tensor

machine feature evaluation

feature selection

iteration


Multi-view Feature Analysis• Multi-view Data in Medical Studies

• Modeling View Correlations

• Multi-kernel method

• Modeling Feature Correlations

• Vector-based method, tensor-based method, dual method

40





• Future Work

41

Future Work• Learning on Different Data Representations

42

neuroimaging experiments tensor data

brain network data

multi-view dataother sources, e.g., clinical and serologic experiments

heterogeneous data fusion

single modality

Future Work• Integrating Multiple Imaging Modalities

• fMRI: functional connections

• DTI: structural connections

43

multiple thresholds [Jie. et al. 2014], multi-spectrum [Wee et al. 2012]

multiple physiological parameters: fiber count, fractional anisotropy (FA), mean diffusivity (MD), and principal diffusivities [Wee et al. 2011]

Future Work• Mining Bioinformatics Information Networks

44

Future Work• BRAIN Initiative

President Obama is making over $300 million investments in the “BRAIN” Initiative to revolutionize our understanding of the human mind and uncover new ways to treat, prevent, and cure brain disorders like Alzheimer’s, schizophrenia, autism, epilepsy, and traumatic brain injury.

45

http://www.whitehouse.gov/BRAIN

http://www.whitehouse.gov/BRAIN

Thank you.!

Documents

Heterogeneous Data Mining for Brain Disorder Identiﬁcationbcao1/doc/WCP_slides.pdf · • Brain Network Analysis • Multi-view Feature Analysis • Future Work 2 Davidson et al