Forecast of dengue incidence using temperature and rainfall

Abstract

Introduction: An accurate early warning system to predict impending epidemics enhances the effectiveness of preventive measures against dengue fever. The aim of this study was to develop and validate a forecasting model that could predict dengue cases and provide timely early warning in Singapore.

Methodology and principal findings: We developed a time series Poisson multivariate regression model using weekly mean temperature and cumulative rainfall over the period 2000-2010. Weather data were modeled using piecewise linear spline functions. We analyzed various lag times between dengue and weather variables to identify the optimal dengue forecasting period. Autoregression, seasonality and trend were considered in the model. We validated the model by forecasting dengue cases for week 1 of 2011 up to week 16 of 2012 using weather data alone. Model selection and validation were based on Akaike's Information Criterion, standardized Root Mean Square Error, and residuals diagnoses. A Receiver Operating Characteristics curve was used to analyze the sensitivity of the forecast of epidemics. The optimal period for dengue forecast was 16 weeks. Our model forecasted correctly with errors of 0.3 and 0.32 of the standard deviation of reported cases during the model training and validation periods, respectively. It was sensitive enough to distinguish between outbreak and non-outbreak to a 96% (CI = 93-98%) in 2004-2010 and 98% (CI = 95%-100%) in 2011. The model predicted the outbreak in 2011 accurately with less than 3% possibility of false alarm.

Significance: We have developed a weather-based dengue forecasting model that allows warning 16 weeks in advance of dengue epidemics with high sensitivity and specificity. We demonstrate that models using temperature and rainfall could be simple, precise, and low cost tools for dengue forecasting which could be used to enhance decision making on the timing, scale of vector control operations, and utilization of limited resources.

Figure 1. Average weekly distribution of dengue cases, mean temperature, and cumulative rainfall.

Graphical presentation of lag relationship between weather predictors and dengue cases using weekly average over the period 2000–2010.

Figure 2. Fitted dengue cases versus reported dengue cases in 2000–2010.

Model-based predicted or fitted dengue cases were plotted against actual reported dengue cases during the model training period.

Figure 3. Forecasted dengue cases versus reported dengue cases in 2011–2012.

Weekly forecasted dengue cases compared with reported cases during the validation period from 2011 week 1 to 2012 week 16. Epidemic threshold was 191 cases for 2011 and 200 cases for 2012.

Figure 4. Analysis of sensitivity of model to detect reported dengue epidemics using ROC curves.

ROC curves in graph A and B show sensitivity of model to detect true outbreak with corresponding probability of false alarm in year 2004–2010 and 2011, respectively.