Predicting Malaria Incidence Using Artifical Neural Networks and Disaggregation Regression

Abstract: Disaggregation modelling is a method of predicting disease risk at high resolution using aggregated response data. High resolution disease mapping is an important public health tool to aid the optimisation of resources, and is commonly used in assisting responses to diseases such as malaria. Current disaggregation regression methods are slow, inflexible, and do not easily allow non-linear terms. Neural networks may offer a solution to the limitations of current disaggregation methods. This project aimed to design a neural network which mimics the behaviour of disaggregation, then benchmark it against current methods for accuracy, flexibility and speed. Cross-validation and nested cross-validation tested neural networks against traditional disaggregation for accuracy and execution speed was measured. Neural networks did not improve on the accuracy of current disaggregation methods, although did see an improvement in execution time. The neural network models are more flexible and offer potential for further improvements on all metrics. The R package 'Kedis' (Keras-Disaggregation) is introduced as a user-friendly method of implementing neural network disaggregation models.