Environmental and Clinical Predictors Forecast Malaria
Environmental and Clinical Predictors Forecast Malaria
The Abuja Declaration noted the importance of accurate disease prediction for targeting and evaluating control measures. For forecasting models to be useful for clinical and public health decision-making, models must produce accurate forecasts. This study examined various predictors across six different settings in Uganda and consistently found that both environmental and clinical predictors were necessary to achieve the highest possible predictive power. This is the first study that examines clinical predictors, other than malaria cases, in combination with environmental predictors for forecasting malaria. Future forecasting work should consider clinical predictors given the likelihood of their relevance in different endemic settings.
Incorporating clinical predictors such as anti-malarial treatment, the proportion of individuals screened for malaria, and the number of malaria negative individuals, produced models with the best predictive power across a range of settings in Uganda and across forecast horizons. In addition, rainfall, temperature, and EVI were also identified as necessary for several of the models in terms of achieving the greatest predictive ability. The accuracy of the models varied widely between the sites, with models at some sites (e.g., Kamwezi) influenced by low and zero counts in the response series, leading to large relative error measures (200%).
It is not known if the observed cases were incident or recrudescent. Inclusion of recrudescent cases in the outcome series would weaken the predictive ability of environmental covariates, which have a stronger relationship with incident cases, although inclusion of recrudescent cases may strengthen the predictive ability of certain treatment predictors. There are different ways in which measurement error could have influenced the findings. Remote sensing data was used in lieu of ground observations due to data availability, and these remote sensing observations are subject to measurement error. The treatment data were based upon prescriptions and not on dispensed anti-malarial medication or treatment taken by the patient, which may have introduced noise into the series, and facility-level factors likely influenced the accuracy of the observed counts of confirmed malaria. Finally, incorporate other predictors, such as humidity and intervention data (e.g., insecticide-treated nets, indoor residual spraying), were not included which may further improve the forecasting accuracy. The Aduku region, for example, has been subject to rounds of indoor residual spraying, which likely accounts for some of the unexplained variation. All of these factors have likely resulted in measurement error, increasing the noise of the different series and decreasing their ability to predict malaria.
The models were not developed to explain causal relationships but were developed with the goal of achieving the highest predictive power. Consequently, multicollinearity was present between various predictor series and influenced which predictors and respective lags were included in the final models. The biological interpretation of specific lags and combination of predictors is therefore limited.
There are different potential users of malaria forecasts. Health facilities could use the forecasts to plan for patient visits, for example, in ensuring that sufficient diagnostic and treatment materials are available. Policy-makers and those involved with malaria control strategy planning could use the information to understand the burden of malaria in a particular location for the coming year, to inform the procurement of anti-malarials and diagnostic equipment, and also in informing malaria control strategy, such as targeting intervention efforts. With the increasing availability of electronic medical records and electronic systems, clinical predictors could be collected and analysed in real-time in conjunction with remote sensing data, if meteorological data are not an option. Using malaria forecasting models in practice would also allow us to understand how accurate a model needs to be, in order to be useful. Potential barriers to the utility of the models include the supply chain management approach, if supply decisions are made at the national level through a national store ('push' system) versus at the health-facility level as well as a lack of resources required to guide community-tailored prevention measures.
Discussion
The Abuja Declaration noted the importance of accurate disease prediction for targeting and evaluating control measures. For forecasting models to be useful for clinical and public health decision-making, models must produce accurate forecasts. This study examined various predictors across six different settings in Uganda and consistently found that both environmental and clinical predictors were necessary to achieve the highest possible predictive power. This is the first study that examines clinical predictors, other than malaria cases, in combination with environmental predictors for forecasting malaria. Future forecasting work should consider clinical predictors given the likelihood of their relevance in different endemic settings.
Incorporating clinical predictors such as anti-malarial treatment, the proportion of individuals screened for malaria, and the number of malaria negative individuals, produced models with the best predictive power across a range of settings in Uganda and across forecast horizons. In addition, rainfall, temperature, and EVI were also identified as necessary for several of the models in terms of achieving the greatest predictive ability. The accuracy of the models varied widely between the sites, with models at some sites (e.g., Kamwezi) influenced by low and zero counts in the response series, leading to large relative error measures (200%).
It is not known if the observed cases were incident or recrudescent. Inclusion of recrudescent cases in the outcome series would weaken the predictive ability of environmental covariates, which have a stronger relationship with incident cases, although inclusion of recrudescent cases may strengthen the predictive ability of certain treatment predictors. There are different ways in which measurement error could have influenced the findings. Remote sensing data was used in lieu of ground observations due to data availability, and these remote sensing observations are subject to measurement error. The treatment data were based upon prescriptions and not on dispensed anti-malarial medication or treatment taken by the patient, which may have introduced noise into the series, and facility-level factors likely influenced the accuracy of the observed counts of confirmed malaria. Finally, incorporate other predictors, such as humidity and intervention data (e.g., insecticide-treated nets, indoor residual spraying), were not included which may further improve the forecasting accuracy. The Aduku region, for example, has been subject to rounds of indoor residual spraying, which likely accounts for some of the unexplained variation. All of these factors have likely resulted in measurement error, increasing the noise of the different series and decreasing their ability to predict malaria.
The models were not developed to explain causal relationships but were developed with the goal of achieving the highest predictive power. Consequently, multicollinearity was present between various predictor series and influenced which predictors and respective lags were included in the final models. The biological interpretation of specific lags and combination of predictors is therefore limited.
There are different potential users of malaria forecasts. Health facilities could use the forecasts to plan for patient visits, for example, in ensuring that sufficient diagnostic and treatment materials are available. Policy-makers and those involved with malaria control strategy planning could use the information to understand the burden of malaria in a particular location for the coming year, to inform the procurement of anti-malarials and diagnostic equipment, and also in informing malaria control strategy, such as targeting intervention efforts. With the increasing availability of electronic medical records and electronic systems, clinical predictors could be collected and analysed in real-time in conjunction with remote sensing data, if meteorological data are not an option. Using malaria forecasting models in practice would also allow us to understand how accurate a model needs to be, in order to be useful. Potential barriers to the utility of the models include the supply chain management approach, if supply decisions are made at the national level through a national store ('push' system) versus at the health-facility level as well as a lack of resources required to guide community-tailored prevention measures.