Three-Month Real-Time Dengue Forecast Models: An Early Warning System for Outbreak Alerts and Policy Decision Support in Singapore

Abstract

Background:With its tropical rainforest climate, rapid urbanization, and changing demography and ecology, Singapore experiences endemic dengue; the last large outbreak in 2013 culminated in 22,170 cases. In the absence of a vaccine on the market, vector control is the key approach for prevention.Objectives:We sought to forecast the evolution of dengue epidemics in Singapore to provide early warning of outbreaks and to facilitate the public health response to moderate an impending outbreak.Methods:We developed a set of statistical models using least absolute shrinkage and selection operator (LASSO) methods to forecast the weekly incidence of dengue notifications over a 3-month time horizon. This forecasting tool used a variety of data streams and was updated weekly, including recent case data, meteorological data, vector surveillance data, and population-based national statistics. The forecasting methodology was compared with alternative approaches that have been proposed to model dengue case data (seasonal autoregressive integrated moving average and step-down linear regression) by fielding them on the 2013 dengue epidemic, the largest on record in Singapore.Results:Operationally useful forecasts were obtained at a 3-month lag using the LASSO-derived models. Based on the mean average percentage error, the LASSO approach provided more accurate forecasts than the other methods we assessed. We demonstrate its utility in Singapore’s dengue control program by providing a forecast of the 2013 outbreak for advance preparation of outbreak response.Conclusions:Statistical models built using machine learning methods such as LASSO have the potential to markedly improve forecasting techniques for recurrent infectious disease outbreaks such as dengue.Citation:Shi Y, Liu X, Kok SY, Rajarethinam J, Liang S, Yap G, Chong CS, Lee KS, Tan SS, Chin CK, Lo A, Kong W, Ng LC, Cook AR. 2016. Three-month real-time dengue forecast models: an early warning system for outbreak alerts and policy decision support in Singapore. Environ Health Perspect 124:1369–1375; http://dx.doi.org/10.1289/ehp.1509981

Niche Modeling of Dengue Fever Using Remotely Sensed Environmental Factors and Boosted Regression Trees

“…While precipitation has been considered significant in other studies [37], here precipitation did not have a significant relative influence. Stanforth et al [46] also detected a relatively low influence of precipitation.…”

“…Over-fitting can be seen as a failure of the algorithm to 'stop' once a balance between predictive performance and model fit has been optimized [33]. This and other reasons have made the BRT method useful in previous studies which successfully mapped DF, along with other vector-borne diseases, and the Aedes mosquito vector itself [4,16,35,36], usually in small select areas of countries such as Malaysia [18] and Singapore [37].…”

The utility of LASSO-based models for real time forecasts of endemic infectious diseases: A cross country comparison

“…The LASSO method has previously been used in dengue outbreak prediction in Singapore, where it is now routinely used to guide vector control policy [38]. The objective of this paper is to apply the LASSO method to infectious disease forecasting and assess more generally in which situations LASSO models will provide useful forecasts.…”

Predictive likelihood for coherent forecasting of count time series

“…The conditional distribution for each observation given past information on responses and past and possibly present covariates is described by a generalized linear model (GLM) distribution. In worldwide development of early-warning systems for vector-borne diseases such as dengue, the main aim is to forecast future disease counts by considering the effect of past disease outcomes and also different covariates such as climate conditions and social ecological conditions on disease counts (Lee et al, 2017;Shi et al, 2016). This type of flexible modeling framework includes regression models for count time series proposed by Zeger and Qaqish (1988), GLM time series models by Li (1994) and Fokianos and Kedem (2004), generalized linear autoregression models of Shephard (1995), GLARMA models of Davis et al (2003), and GARMA models of Benjamin et al (2003).…”