2014.ISDS.Abstracts.Final.pdf


ISDS Annual Conference Proceedings 2014. This is an Open Access article distributed under the terms of the Creative Commons Attribution-
Noncommercial 3.0 Unported License (http://creativecommons.org/licenses/by-nc/3.0/), permitting all non-commercial use, distribution, 
and reproduction in any medium, provided the original work is properly cited.

163

ISDS 2014 Conference Abstracts

Real-time Forecasting of the 2014 Dengue Fever Season 
in Thailand

Nicholas G. Reich1, 2, Krzysztof Sakrejda1, Stephen A. Lauer*1, Derek A. Cummings2, 
Paphanij Suangtho3, Soawapak Hinjoy2, 3, Sopon Iamsirithaworn3, Hannah Clapham2, 
Henrik Salje2 and Justin Lessler2

1Biostatistics, University of Massachusetts, Amherst, MA, USA; 2Johns Hopkins Bloomberg School of Public Health, Baltimore, MD, 
USA; 3Thai Ministry of Public Health and Bureau of Epidemiology, Nonthaburi, Thailand

Objective
To develop a statistical model for dengue fever surveillance that 

uses data from across Thailand to give early warning of developing 
epidemics.

Introduction
Dengue is a major cause of morbidity in Thailand. Annual 

outbreaks of varying sizes provide a particular challenge to the public 
health system because treatment of severe cases requires significant 
resources. Advanced warning of increases in incidence could help 
public health authorities allocate resources more effectively and 
mitigate the impact of epidemics.

Methods
Starting in November of 2013, we received biweekly dengue 

fever count data from the Thai Ministry of Public Health and Bureau 
of Epidemiology. After cleaning, this data was combined with 
historical case counts dating back to 1968 and seasonal trends for 
each province to build a series of statistical models, ranging from 
smoothed regression-based models to Gaussian process models. Once 
case counts are completed for a given biweek - which can take up to 
six months in rural provinces - each model can be cross-validated 
and compared, using methods such as mean average squared error 
(MASE).

Results
Thus far we have generated forecasts at 22 time points for a suite 

of different models. Since cases continue to be reported throughout 
2014, the evaluation of the prediction performance is ongoing.

Conclusions
One of the largest challenges from real-time reporting is dealing 

with incomplete case count data. We have found that we do not 
receive a full case count for a province in a given biweek for at least 
two months and often up to six months. Since this is the first year of 
collected data, we have found it difficult to decipher whether case 
count differences are attributable to reporting rates, seasonality, 
climate change, or other population factors. Through working on this 
project, we have learned that models built on historical data are not 
necessarily applicable in real-time and that novel methods must be 
employed to improve infectious disease predictions, especially in 
developing countries with partial reporting data.

Keywords
Infectious disease surveillance; Real-time forecasts; Statistical model 
selection; Dengue fever

*Stephen A. Lauer

E-mail: stephenalauer@gmail.com    

Online Journal of Public Health Informatics * ISSN 1947-2579 * http://ojphi.org * (1):e205, 201