Disease Outcome Prediction Using Graph Auto-encodersnetworks.ece.mcgill.ca/sites/default/files/Disease_Outcome_Predicti… · Disease Outcome Prediction Using GAE Introduction Motivation

Disease Outcome Prediction Using GAE

Disease Outcome Prediction Using GraphAuto-encoders

Juliette Valenchon

Department of Electrical and Computer Engineering, McGill University, MontrealUnder the supervision of Mark Coates

February 13, 2019

1/19


Introduction

Motivation

Head and Neck Cancer1 Breast Cancer2 Cardiovascular disease3

Colorectal Cancer4Lung Cancer5

Alzheimer’s Disease6

1Reproduced from ”Head and Neck Cancer is not Just a Smoker’s Disease Anymore”, Mount Sinai News. 2 S. Roan,

”Early Stage Breast Cancer: Do You Really Need Your Lymph Nodes Removed?”, Everyday Health. 3 ”ConqueringCardiovascular Disease”, NIH. 4 ”Colorectal Cancers”, Dr. Fuhrman. 5 K. O’Sullican, ”New drug approved for advanced lungcancer by HSE”, The Irish Times. 6 M. Casalino, ”Alzheimer’s Association Offers Virtual Dementia Tour”, Patch

2

3

4

5

6

2/19


Introduction

Motivation

I Traditionally: risk calculator for possibility of disease development.I Framingham study: prediction for hospitalization for long-term

cardiovascular disease

Risk calculator

Machine learning algorithm

Patient information

Patient information

Accuracy of 56 %

Accuracy of 82 %

Figure 2: Comparison of a risk calculator and a machine learning algorithm7

7W. Dai et al., “Prediction of hospitalization due to heart diseases by supervised learning methods” Int. J. medical

informatics, vol. 84, no. 3, pp. 189–197, 2015.3/19


Introduction

Motivation

Mild Cognitive Impairment

(MCI): Clinical diagnosis

Possible Alzheimer’s Disease (AD) : irreversible disease, destroys brain cells

Dementia : 46.8 million in 2015

Stable

→ Early and accurate diagnosis for an early treatment to improve thequality of life for some time

4/19


Introduction

Goal

I Predict conversion from MCI to ADI Multimodal data with missing values

(a) MRI8(b) DTI9

(c) PET10

I Use characteristics of subjects

8Clinica developers, ”Volume pre-processing - Clinica Documentation”. 9 Rachel VanCott, ”NOVA — scienceNOW —

Diagnosing Damage image 3 — PBS”. 10 University of California - Berkeley, ”PET scans reveal key details of Alzheimer’sprotein growth in aging brains”

9

10

5/19


Introduction

Goal

I Predict conversion from MCI to ADI Multimodal data with missing values

(a) MRI8(b) DTI9

(c) PET10

I Use characteristics of subjects

8Clinica developers, ”Volume pre-processing - Clinica Documentation”. 9 Rachel VanCott, ”NOVA — scienceNOW —

Diagnosing Damage image 3 — PBS”. 10 University of California - Berkeley, ”PET scans reveal key details of Alzheimer’sprotein growth in aging brains”

9

10

5/19


Introduction

Problem formulation

→ Deal with missing values→ Perform classification

→ Matrix completion→ Label as feature

6/19


Introduction

Problem formulation

→ Deal with missing values→ Perform classification

→ Matrix completion→ Label as feature

6/19


Introduction

Matrix completion

I Recover missing values by solving optimization problemI Loss function :

l = ||Ω ∗ (M − M)||+ γlΩb(M, M) + β

q∑i=1

Wi (1)

n-1 features

m s

ubje

cts

M mxn Value available

Missing value

labels

7/19


Architecture

Graph methods for the prediction of MCI to AD conversion

Figure 5: Overview of the pipeline used for classification of population graphsusing Graph Convolutional Networks. Reproduced from Parisot et al. [1]

I No missing data

I One graph

8/19


Architecture

Graph methods for the prediction of MCI to AD conversion

Figure 6: Vivar et al. [2]

I Matrix completionI Missing dataI One graph

9/19


Architecture

A novel graph-based method for the prediction of MCI toAD conversion

Features

Subjects

MGAE

Features

Subjects

Bipartite graph

Figure 7: Proposed architecture

I Matrix completion : Van den Berg et al. [3]I Missing dataI Multiple graphs

10/19


Architecture

Defining the feature dependencies

60 80age

1.0

0.5

0.0

0.5

1.0

feat

ure

sex0.01.0

60 80age

1.0

0.5

0.0

0.5

1.0

feat

ure

sex0.01.0

Age-related features.

60 80age

1.0

0.5

0.0

0.5

1.0

feat

ure

sex0.01.0

60 80age

1.0

0.5

0.0

0.5

1.0fe

atur

e

sex0.01.0

Sex-related features.

60 80age

1.0

0.5

0.0

0.5

1.0

feat

ure

sex0.01.0

60 80age

1.0

0.5

0.0

0.5

1.0

feat

ure

sex0.01.0

Age & Sex features.

Figure 8: Relationships of age and sex (Men and Women) with six differentfeatures in the case of Alzheimer’s disease.

11/19


Architecture

Bipartite graph

1

2

i

m

1

2

j

j+1

n

Subjects FeaturesGroup 1Group kGroup K

M(1,1)

M(1,2)M(2,2)

M(1,j+1)

M(1,n)

M(2,n) M(2,j)

M(2,j+1)

M(i,j)

M(i,n)

M(m,n)

M(m,1)

M(i,1)

M(i,2)

I Relationship between a groupof subjects and a group offeatures

12/19


Architecture

Graph Auto-encoder

Subject i

Age-related featuresSex-related featuresAge & Sex-related features

Features

Feature j

Subjects

WomenMen

i

j

Subject embedding

Feature embedding

Graph encoder

Embeddings

DecoderM(i,j)

13/19


Implementation details

Datasets

TADPOLE dataset

I 779 subjects

I 564 features

I 21 % missing data

564 features

779

subj

ects

M 779x564 Value available

Missing value

Creation of a synthetic dataset

I 779 subjects

I 564 features

I No missing data

564 features

779

subj

ects

M 779x564 Value available

14/19



Creation of the synthetic dataset

FeatureAge, Sex or Age & Sex related

M(i , j) = mj fij + ij +εij +vj ∗yi (2)

fij = xi if age (3)

= si if sex (4)

= sixi if age & sex (5)

mj ∼ U [−m,m] (6)

ij ∼ U [a, b] (7)

εij ∼ N (0, σ) (8)

vj ∼ U [c , d ] (9)

15/19



Evaluation measure for performance

I Chosen Metrics: Integral of ROC : AUC (Area Under the Curve)I ROC measures the true positive rate, relative to the number of false

positivesI Integral of ROC ranges from 0 to 1, with 1 being the best

0.0 0.2 0.4 0.6 0.8 1.0False Positive Rate

0.0

0.2

0.4

0.6

0.8

1.0

True

Pos

itive

Rat

e

Perfect ROCRandom

Figure 9: Perfect and random ROC curve

16/19


Results

Results on the real dataset

Linear SVM sRGCNN GC-MC RFAlgorithm

0.4

0.5

0.6

0.7

0.8

AUC

Comparison of four algorithms

17/19


Results

Results on the synthetic dataset

Linear SVM GC-MC RFAlgorithm

0.60

0.65

0.70

0.75

0.80

0.85

0.90

0.95

AUC

Comparison of three algorithms

18/19


Results

Conclusion

I Better than baseline methods linear SVM and MLP

I Better performance than sRGNN by 2.9 %

I Random Forest performs better

Future work:

I Remove missing values in the dataset

19/19

Documents

Disease Outcome Prediction Using Graph Auto-encodersnetworks.ece.mcgill.ca/sites/default/files/Disease_Outcome_Predicti… · Disease Outcome Prediction Using GAE Introduction Motivation