Predicting diarrhoea outbreaks with climate change

in performance for predictions with the LSTM and SVM methods was more than the CNN method. Table 4 compares the performance of the overall predictions made by each ML method when the augmented data-sets were used based on the parameter tuning technique selected. By comparing the RMSE of the augmented data made with grid-search parameters in Table 3 against the RMSE of predictions made with the real-world data in Table 2. For each ML method, the predictions made with the augmented datasets yielded better and lower RMSE than their predictions with the real-world data-sets. Thus, we can infer that the amount of training data used for training, significantly affects the prediction performance of all the three ML methods. By comparing the average RMSE percentages across data-sets, it also shows that CNN outperformed the other methods when the real-world dataset was used alone while LSTM outperformed the other methods when either of the augmented datasets were used. Fig 7 shows the results when the parameters of the three ML methods were tuned with REVAC instead of grid-search. We found that the CNN method’s prediction results improved across all provinces. The highest percentage increase recorded for CNN was over 12% and the least increase was about 2.5%. The LSTM method’s performance also increased across most province, however, its predictive task performance declined in Limpopo, KwaZulu Natal and Free state provinces. Among the three methods, the SVM recorded the highest number of provinces that saw a decline in task performance. The average increase of SVM task perfor-

mance across all provinces was also the least. Fig 8 shows the provincial prediction results of the ML methods when augmented data was used for training. In Fig 6a, when grid-search parameters were used, the LSTM method out- performed all the other methods in most provinces with both augmented data-sets and was Fig 6. Percentage change in performance of each ML method for predictions in Experiment II. (a) & (b): Percentage change in performance of each ML method over each province when predictions were made with the (a) upward augmented data-set and (b) downward augmented data-set instead of the real-world data. High percentage RMSE indicates an improvement in performance and vice-versa. https://doi.org/10.1371/journal.pone.0262008.g006 Table 4. RMSE averages for REVAC and grid-search method parameter tuning. ML Method REVAC tuning Grid-search tuning Upward augmented data Downward augmented data Upward augmented data Downward augmented data CNN 22.07% 23.86% 23.11% 25.80% LSTM 21.60% 23.61% 21.93% 23.78% SVM 22.17% 27.30% 22.17% 27.97% Standard Deviation 0.003 0.021 0.006 0.134 https://doi.org/10.1371/journal.pone.0262008.t004 PLOS ONE Predicting diarrhoea outbreaks with climate change PLOS ONE | https://doi.org/10.1371/journal.pone.0262008 April 19, 2022 11 / 18