Artificial neural networks for infectious diarrhea prediction using meteorological factors in Shanghai Yongming Wang, Junzhong Gu and Zili Zhou Department.

Artificial neural networks for infectious diarrhea prediction using meteorological factors in Shanghai

Yongming Wang, Junzhong Gu and Zili ZhouDepartment of Computer Science & Technology, East China Normal University

Institute of Computer Applications, Shanghai, ChinaE-mail: [email protected]

http://www.ica.stc.sh.cn

6th International Conference on Software Engineering and Knowledge Engineering

SEKE 2014, Hyatt Regency, Vancouver, Canada

OUTLINES• Introduction

• Study area and dataset

• Prediction method and performance metrics

• Development of FFBPNN model

– input and output parameters

– Data pre-processing and post-processing

– Determination of optimum network and parameters

• Development of MLR model

• Experiments results and discussion

• Sensitivity analyses

• Conclusions

Introduction

As a kind of common and important infectious disease, infectious diarrhea has a serious threat to human health and leads to one billion disease episodes and 1.8 million deaths each year (WHO, 2008).

In Shanghai of China which is the biggest developing country, the incidence of infectious diarrhea has significant seasonality throughout the year and is particularly high in the summer and autumn of recent years.

Hence, a robust short-term forecasting model for infectious diarrhea incidence is necessary for decision-making in policy and public health.

Introduction

Infectious diseases have a closely relation with meteorological factors, such as temperature and rainfall, and can affect infectious diseases in a linear or nonlinear fashion. In recent years, there has been a large scientific and public debate on climate change and its direct as well as indirect effects on human health.

As far as we are concerned with the prediction of diarrhea diseases in literature, many forecasting models based on statistical methods for diarrhea diseases forecasting have been reported.

With regard to the fact that number of meteorological factor that effect infectious diarrhea are too much and the inter-relation among them is also very complicated, prediction models based on statistics methods may not be fully suitable for such type of problems.

Introduction

Nowadays, Artificial Neural Networks (ANNs) are considered to be one of the intelligent tools to understand the complex problems and have been widely used in the medical and health field. To the best knowledge of the authors, there is no works has been carried out to utilize the ANNs method in predicting diarrhea disease.

Contribution: Establish a new ANNs model (FFBPNN) to predict infectious diarrhea in Shanghai with a set of meteorological factors as predictors.

Study area and Dataset-Study area

Shanghai is located in the eastern part of China which is the largest developing country in the world, and the city has a mild subtropical climate with four distinct seasons and abundant rainfalls. It is the most populous city in China comprising urban/suburban districts and counties, with a total area of 6,340.5 square kilometers and had a population of more then 25.0 million by the end of 2013.

Study area and dataset-dataset

The infectious diarrhea cases for the period 2005.1.3-2009.1.4

0 50 100 150 200 2500

50

100

150

200

250

300

350

Time(week)

Weekly

num

ber of in

fectio

us dia

rrhea

Study area and dataset-dataset

The meteorological factors data for the period 2005.1.3-2009.1.4

2005 2006 2007 2008 2009 20100

5

10

15

20

25

30

35

40

Time(week)

Wee

kly

aver

age

max

imum

tem

pera

ture

(a) 2005 2006 2007 2008 2009 2010-5

0

5

10

15

20

25

30

Time(week)

Wee

kly

aver

age

min

imum

tem

pera

ture

(b) 2005 2006 2007 2008 2009 20100

5

10

15

20

25

30

35

Time(week)

Wee

kly

aver

age

tem

pera

ture

(c) 2005 2006 2007 2008 2009 201045

50

55

60

65

70

75

80

85

90

Time(week)

Wee

kly

aver

age

min

imum

rel

ativ

e hu

mid

ity

(d)

2005 2006 2007 2008 2009 201045

50

55

60

65

70

75

80

85

90

Time(week)

Wee

kly

aver

age

rela

tive

hum

idity

(e)2005 2006 2007 2008 2009 2010

995

1000

1005

1010

1015

1020

1025

1030

1035

Time(week)

Wee

kly

aver

age

atm

osph

eric

pre

ssur

e

(f) 2005 2006 2007 2008 2009 20100

2

4

6

8

10

12

Time(week)W

eekl

y av

erag

e su

nshi

ne d

urat

ion

(g) 2005 2006 2007 2008 2009 20101

2

3

4

5

6

7

Time(week)

Wee

kly

aver

age

win

d sp

eed

(h)

2005 2006 2007 2008 2009 20100

5

10

15

20

25

30

35

Time(week)

Wee

kly

aver

age

rain

fall

(i)

Method and performance metrics

The schematic flowchart of proposed method.

Data Collecting

Dataset

Data calculating

Data normalizing

Data gathering

Pre-processing

Models development

Models testing and comparing

Data m

ining

Prediction Model

Step 1: Data collection

Step 2: Data pre-processing

Step 3: Data mining

Method and performance metrics

Three layered feed-forward back-propagation artificial neural network model.

bbxvwwxfym

j

n

ijiijji

1 10 )()(

Method and performance metrics

The models with the smallest RMSE, MAE and MAPE and the largest R and R2 are considered to be the best models.

n

ttt yyn

MAE1

ˆ1

%1001

1

ˆ

n

tt

tt

yyy

nMAPE

n

tyy ttn

RMSE1

2

)ˆ(1

n

t

n

t

yy

yy

t

ttR

1

21

2

)(

)ˆ(1

n

tt

n

ttt

y

yyR

1

2

1

2

2

)(

)(1

ˆ

Development FFBPNN model

The FFBPNN modeling consists of two steps: --- Train the network using training dataset --- Model input and output parameters --- Data pre-processing and post-processing --- Determination of optimum network and parameters --- Test the network with testing dataset

Hidden neurons and network errors

Development FFBPNN model

Parameters FFBPNN

Number of input layer units 9

Number of hidden layer 1

Number of hidden layer units 4

Number of output layer units 1

Momentum rate 0.9

Learning rate 0.74

Error after learning 1e-6

Learning cycle 1500 epoch

Transfer function in hidden layer Tansig

Transfer function in output layer Purelin

Training function TRAINGDM

The optimum model architecture and parameters for the diarrhea prediction.

Development MLR model

Dependent variable : diarrhea number

Independent variables : meteorological factors

RWS

SDAPRH

RHTT

TWNID

avg

avgavg

avg

6048.17205.15

7734.50902.22993.0

6506.16208.2.815802

0.961917903.1972

minmin

max

Results and discussion

PECs

Models

FFBPNN MLR

Training Testing Training Testing

MAE 20.7628 27.7547 29.8077 35.3774

RMSE 28.3007 36.0526 39.3739 48.9395

MAPE(%) 27.27% 38.41% 43.37% 41.82%

R 0.8783 0.8490 0.8089 0.6968

R2 0.9213 0.9125 0.8811 0.8388

The reason of better performances of the FFBPNN model over MLR model may be attributed to the complex nonlinear relationship between infectious diseases and meteorological factors.

Results and discussion

0 20 40 60 80 100 120 140 1600

50

100

150

200

250

300

350

Time(week)

The

wee

kly

num

ber

of in

fect

ious

dia

rrhe

a

Actual

FFBPNN

(a)0 20 40 60 80 100 120 140 160

-50

0

50

100

150

200

250

300

350

Time(week)

The

wee

kly

num

ber

of in

fect

ious

dia

rrhe

a

Actual

MLR

Comparison curves plot of actual vs. predicted trends for training dataset

FFBPNN MLR

Results and discussion

Comparison scatter plot of actual vs. predicted values for training dataset

FFBPNN MLR

0 50 100 150 200 250 300 3500

50

100

150

200

250

300

Actual values

FF

BP

NN

pre

dict

ed v

alue

s

y=0.83+17

R2=0.9385

(b)

Results and discussion

Comparison curves plot of actual vs. predicted trends for testing dataset

FFBPNN MLR

0 10 20 30 40 50 600

50

100

150

200

250

300

Time(week)

The

wee

kly

num

ber

of in

fect

ious

dia

rrhe

a

Actual

FFBPNN

(c) 0 10 20 30 40 50 600

50

100

150

200

250

300

Time(week)

The

wee

kly

num

ber

of in

fect

ious

dia

rrhe

a

ActualMLR

(c)

Results and discussion

Comparison scatter plot of actual vs. predicted values for testing dataset

FFBPNN MLR

0 50 100 150 200 250 3000

50

100

150

200

250

Actual values

FF

BP

NN

pre

dict

ed v

alue

s

y=0.68x+28

R2=0.9125

(d) 0 50 100 150 200 250 3000

20

40

60

80

100

120

140

160

180

200

Actual values

MLR

pre

dict

ed v

alue

s

y=0.54x+39

R2=0.8388

(d)

Sensitivity analyses

Sensitivity analysis (Cosine Amplitude Method)

ANNs

Meteorological factor Infectious diarrhea

black-box

m

k

m

k

m

kjkikjkikij xxxxr

1 1 1

22/

Sensitivity analyses

Most effective meteorological factor : temperature

least effective meteorological factor :average rainfall

Conclusions1. The proposed method is more suitable for prediction infectious diarrhea then statistical methods MLR.

2. The feed-forward back-propagation neural network (FFBPNN) model with architecture 9-4-1 has the best accurate prediction results in prediction of the weekly number of infectious diarrhea.

3. most effective meteorological factor on the infectious diarrhea is weekly average temperature, whereas weekly average rainfall is the least effective parameter on the infectious diarrhea.

Therefore, this technique can be used to predict infectious diarrhea. The results can be used as a baseline against which to compare other prediction techniques in the future.