Predicting diarrhoea outbreaks with climate change

methods. ML methods are known for their ability to handle high-dimensional data and model complex predictive problems. Several supervised learning-based ML techniques such as Support Vector Machines (SVMs) [16] and Deep learning techniques such as Convolutional Neural Networks (CNNs) [17], Long Short-Term Memory Networks (LSTMs) [18] have been applied in medical research for devel- oping predictive and diagnostic models for various diseases [14, 15]. For example, CNNs have been used for the detection of Malaria parasite [19] and Tuberculosis diseases [20] in individu- als. LSTMs have also been used to predict the outbreak of diseases like Typhoid, Chicken Pox and Scarlet Fever [14]. SVMs were also used for Hepatitis disease detection [21]. These ML methods are widely used for modelling infectious diseases because of the numerous advantages they possess. For instance, CNNs are popular for their powerful feature extraction capabilities [17]. LSTMs are commonly used to handle sequential tasks such as time series forecasting because of their ability to capture long term dependencies [14]. SVMs are widely accepted for their ability to solve nonlinear regression estimation problems, their non-parametric nature enables them to represent complex and nonlinear functions easily [16]. Despite advances in a range of health-care applications using such predictive-based ML [14, 15, 21], there is a lack of research and data on the efficacy of such predictive ML methods for diarrhoea outbreak prediction in Sub-Saharan Africa. Additionally, the overall task perfor-

mance of ML algorithms, applied to many health-care applications and more broadly to any predictive classification task, largely depends on the manual tuning and calibration by algo- rithm designers and experimenters of methodological parameters over the course of several experimental trials [22, 23]. Such manual tuning is often ineffective and significantly limits the full potential of task performance achieved by the ML method, especially for high-dimen- sional, partially observable, noisy and complex task domains [22], as are typified by the nature of data-sets in many health-care applications including diarrhoea outbreak prediction. Task performance also largely depends on the amount of available training data [24], which is a sig- nificant challenge for most predictive ML in health-care applications due to the sensitive and controlled nature of health-care data-sets [25]. The inaccessibility of data adds to the difficulty of method comparison, accuracy, and the advancement of ML as a whole [24, 26]. The overall aim of this study is to ascertain the suitability of various ML methods given vari- ous climate factors and synthetic (generative) training data for accurately predicting diarrhoea outbreaks. Specifically, the study aims to elucidate what type of ML method is most appropri- ate when coupled with specific training and test data-sets (that is, specific climate variables, data-sparseness, data-noise and synthetic data compliment), in order to optimise prediction efficacy. Thus, we compared task-performance of three ML methods (CNNs, LSTMs and

SVMs) to ascertain the most suitable method for predicting future number of daily diarrhoea cases in nine South African provinces. The average predictive accuracy of each method was compared across multiple datasets and experiment replications. Given the sparse and noisy nature of the data-sets used for method training and testing, we necessarily augmented the available data (real-world data) with synthetic data generated using Generative Adversarial Networks (GANs). GANs were selected as they have been previously demonstrated as effective for generating different types of realistic data [24, 25]. Also, since there was a lack of previous research to guide parameter tuning and calibration for optimising such ML methods applied to diarrhoea outbreak prediction, we used the Relevance Estimation and Value Calibration (REVAC) method [27]. REVAC is an evolutionary algorithm design for meta-heuristic param- eter tuning, and as such was applied to optimise methodological parameters of the ML meth- ods used in this study. Previous work has demonstrated the effectiveness of REVAC for parameter tuning and attaining optimal algorithm performance across a range of complex, noisy and high-dimensional search spaces [28, 29]. PLOS ONE Predicting diarrhoea outbreaks with climate change PLOS ONE | https://doi.org/10.1371/journal.pone.0262008 April 19, 2022 3 / 18