Weather-based prediction of Plasmodium falciparum malaria in epidemic-prone regions of Ethiopia II. Weather-based prediction systems perform comparably to early detection systems in identifying times for interventions

Abstract

Background: Timely and accurate information about the onset of malaria epidemics is essential for effective control activities in epidemic-prone regions. Early warning methods that provide earlier alerts (usually by the use of weather variables) may permit control measures to interrupt transmission earlier in the epidemic, perhaps at the expense of some level of accuracy.

Methods: Expected case numbers were modeled using a Poisson regression with lagged weather factors in a 4th-degree polynomial distributed lag model. For each week, the numbers of malaria cases were predicted using coefficients obtained using all years except that for which the prediction was being made. The effectiveness of alerts generated by the prediction system was compared against that of alerts based on observed cases. The usefulness of the prediction system was evaluated in cold and hot districts.

Results: The system predicts the overall pattern of cases well, yet underestimates the height of the largest peaks. Relative to alerts triggered by observed cases, the alerts triggered by the predicted number of cases performed slightly worse, within 5% of the detection system. The prediction-based alerts were able to prevent 10-25% more cases at a given sensitivity in cold districts than in hot ones.

Conclusions: The prediction of malaria cases using lagged weather performed well in identifying periods of increased malaria cases. Weather-derived predictions identified epidemics with reasonable accuracy and better timeliness than early detection systems; therefore, the prediction of malarial epidemics using weather is a plausible alternative to early detection systems.

Figure 1

Observed and predicted number of malaria cases with alerts triggered by mean plus 1.5 SD using predicted cases. The solid lines for observed cases and the dotted lines for predicted cases. The red marks are the timing of alerts triggered using predicted cases; their position along the y-axis does not have a meaning.

Figure 2

Comparing performance of prediction and detection systems. Percent of PPC by number of alerts per year for different algorithms. (a) and (c) were obtained from cases in excess of the weekly mean (low effectiveness) with window of effectiveness of 8 and 24 weeks respectively. (b) and (d) were obtained from cases in excess of the weekly mean minus one standard deviation (high effectiveness) for windows of eight & 24 weeks, respectively. The solid lines are for detection (Obs) and the dotted lines for prediction (Pred). MeanSD and Percentile represent threshold algorithms based on mean plus standard deviation and percentile, respectively.

Figure 3

Comparison of performance of prediction systems in cold and hot districts. Percent of PPC by number of alerts per year. PPC was obtained from cases in excess of the weekly mean (low effectiveness) with windows of effectiveness of eight weeks (a) and 24 weeks (b). The solid lines represent cold and the dotted lines hot districts.