Predicting diarrhoea outbreaks with climate change

thus, a fixed random seed (https://www.tensorflow.org/) was set for all experiments. For all ML methods, we kept some parameters fixed (based on parameter values established in previ- ous related work [14, 15]), while others were tuned. See Table 1 for the list of tuned parameters. Determining the optimal ML parameters One of the major factors that influence the performance of ML methods is the configuration settings of its parameters. Thus, in this study we used grid-search tuning [22] and REVAC [27] parameter tuning methods to find optimal parameter values for all ML methods. Both parame- ter tuning technique select a combination of possible parameter values from a range of values specified by a user. See Table 1 for the list and range of parameters values specified for each ML method. The deep learning parameters that were not specified used the default values of the Keras package. The grid-search method was implemented with the python Scikit-Learn package (https://scikit-learn.org/) while REVAC tuning was designed based on the methodol- ogy used by Nannen & Eiben [27]. REVAC was implemented at a layer that aids in searching for optimal parameter values for an ML algorithm trying to solve the problem of predicting daily diarrhoea cases. See S1 Appendix for technical details on REVAC implementation. The parameter tuning of each ML method was implemented separately for each province. Experiments setup Table 2 gives an overview of the experiments conducted for this study and Fig 5 presents the

overall pipeline used to predict daily diarrhoea cases in our experiments. Since this is a regres- sion task, the input data were all in numerical format. Previous studies such as [12, 15] have shown that the basic form of feature engineering applied to a time series prediction task is tak- ing past observations into consideration. Although, deep learning methods are known for automatic feature engineering [17], we applied this feature engineering step across all models for consistency. This approach is also consistent with previous works such as [14, 15, 33]. To make forecasts on the possible number of daily diarrhoea cases, we considered past observa- tions (lags) in all our methods because patterns of the past are likely to be repeated in the Table 1. Experiment parameters and corresponding value ranges. ML method Parameter Parameter range SVM C [1, 100] Gamma [0.001, 0.1] LSTM Dropout rate [0.1, 0.2, 1.0] LSTM layers [1, 2, 3] Neurons [6, 12, 16, 18, 24, 28, 32, 50, 64, 100] Batch size [4, 16, 18, 32, 64] Learning rate [0.001, 0.01] Epochs size [40, 50, 60, 70, 100, 120, 150, 200] CNN Pool size [1, 2] Convolutional layers [1, 2, 3] Kernel size [1, 2, 3] Batch size [4, 16, 18, 32, 64] Learning rate [0.001, 0.01] Epochs size [40, 50, 60, 70, 100, 120, 150, 200] Filter size [6, 12, 16, 18, 24, 28, 32, 64] https://doi.org/10.1371/journal.pone.0262008.t001 PLOS ONE Predicting diarrhoea outbreaks with climate change PLOS ONE | https://doi.org/10.1371/journal.pone.0262008 April 19, 2022 8 / 18