Predicting diarrhoea outbreaks with climate change

The prediction performance of all ML methods improved when the augmented data-sets were used for training, with the LSTM (LSTM method section) giving the overall best perfor- mance. This implies that a large training set size boosts the performance of most ML algo- rithms. We also surmise that the LSTM method performs better when the size of training data is large, perhaps the reason for its relatively poor performance in the first experiment where only real-world data with limited training set was used. A study conducted by [39] have shown that LSTM benefits from a large training set size. In addition, Another study by [14] reported that LSTMs are a state of the art for capturing the long-term dependencies specific to a given data-set thus their ability to learn patterns in sequential data with sufficient training size regardless of its noisy nature. Performance based on parameter tuning method With respect to the parameter tuning as a factor for task performance with the augmented data, we found that with the given grid-search parameters (Table 1), the average percentage increase in task performance of the CNN method was the least when compared to the other methods across individual provinces. The provincial instances such as in Gauteng, Eastern Cape, and Mpumalanga in Fig 8a (Provincial results of the ML methods with the augmented data-sets in Experiments II & III. figure) where SVM outperformed the CNN method is likely due to the parameter settings of CNN used during training. Therefore, we deduce that the

choice of parameters greatly affects the performance of deep learning models especially when applied to noisy and augmented data-sets. Thus, we setup a different experiment with REVAC tuning strategy. With the REVAC parameter tuning implementation, the CNN method gave the highest percentage increase in performance across each province. However, the LSTM method’s pre- diction performance was still better than the other methods for most provinces. However, the SVM demonstrated the least average percentage increase and the highest average percentage decrease across the provinces. Therefore, we can infer from these results that the REVAC parameter tuning is not ideal for the SVM method rather it is more suited to deep learning methods. A possible explanation maybe the low dimensional search space of parameters for the SVM method considering that an SVM’s (with RBF kernel) major parameters are gamma and C only. A study by [40] have found that predefining a search space especially for few parameters can be difficult. However, [22] reported that grid-search is better suited for low dimensional search space perhaps the reason for SVM’s satisfactory performance with grid- search tuning. In Table 5, we compared the performance of the results obtained when REVAC parameter tuning was used on the upward augmented data with the results of some existing models on diarrhoea outbreak prediction with different datasests [14, 15, 41]. Although our RMSE values appear lower, we note that the difference in the error values may be due to the type/size of the

dataset used in the different study as well the unit and scale of the dataset. Table 5. Root Mean Square Error (RMSE) performance comparison with the existing diarrhoea prediction studies. Study CNN LSTM SVM RF ARIMA our study 0.22 0.21 0.22 - - [14] - 1.43 - - 1.38 [15] - - 49.91 48.14 - [41] - - - 0.45 0.31 https://doi.org/10.1371/journal.pone.0262008.t005 PLOS ONE Predicting diarrhoea outbreaks with climate change PLOS ONE | https://doi.org/10.1371/journal.pone.0262008 April 19, 2022 14 / 18