Prediction of Dengue Outbreaks Based on Disease Surveillance and Meteorological Data

Abstract

Research is needed to create early warnings of dengue outbreaks to inform stakeholders and control the disease. This analysis composes of a comparative set of prediction models including only meteorological variables; only lag variables of disease surveillance; as well as combinations of meteorological and lag disease surveillance variables. Generalized linear regression models were used to fit relationships between the predictor variables and the dengue surveillance data as outcome variable on the basis of data from 2001 to 2010. Data from 2011 to 2013 were used for external validation purposed of prediction accuracy of the model. Model fit were evaluated based on prediction performance in terms of detecting epidemics, and for number of predicted cases according to RMSE and SRMSE, as well as AIC. An optimal combination of meteorology and autoregressive lag terms of dengue counts in the past were identified best in predicting dengue incidence and the occurrence of dengue epidemics. Past data on disease surveillance, as predictor alone, visually gave reasonably accurate results for outbreak periods, but not for non-outbreaks periods. A combination of surveillance and meteorological data including lag patterns up to a few years in the past showed most predictive of dengue incidence and occurrence in Yogyakarta, Indonesia. The external validation showed poorer results than the internal validation, but still showed skill in detecting outbreaks up to two months ahead. Prior studies support the fact that past meteorology and surveillance data can be predictive of dengue. However, to a less extent has prior research shown how the longer-term past disease incidence data, up to years, can play a role in predicting outbreaks in the coming years, possibly indicating cross-immunity status of the population.

Fig 1. The Number of Dengue Cases in Yogyakarta Province, 2001–2010.

Fig 2. Time Series Graphs of Surveillance and Meteorological Data.

Fig 3. Association between Meteorological Variables and Dengue over Lag 0–3.

Solid lines represent relative risks of dengue cases and dotted lines depict the upper and lower limits of 95% confidence intervals.

Fig 4. Relationship between Autoregressive Lags and Dengue Counts.

Upper panel shows (a) relative risks of dengue cases as functions of dengue surveillance at 2-month lag times. Lower panel shows (b) the relation between case intensity and dengue risk categories at all lag months; and (c) the risk in each future month following an increase of 5 dengue cases in a specific month.

Fig 5. Monthly Observed and Predicted Dengue Cases from 2001–2010.

Black line represents observed dengue cases and red line represents predicted cases. The vertical axis shows dengue cases and the horizontal axis denotes time in month from January 2001 to December 2010.

Fig 6. Residual Diagnosis.

Upper panel shows (a) the residual histograms; and (b) the Q-Q plot for deviance residuals. Lower panel shows (c) the partial ACF plot; and (d) the relationship between reported and predicted cases.

Fig 7. Predicted Dengue Cases Versus Reported Dengue Cases in 2001–2013.

Monthly predicted dengue cases compared with reported cases during January 2001 to December 2013. Black line represents observed dengue cases, grey line represents the epidemic threshold, red line represents predicted cases using training dataset, and blue line represents predicted cases using the external validation dataset not used for model fitting.