Predicting diarrhoea outbreaks with climate change

closely followed by the SVM except in Western Cape and KwaZulu Natal province where the CNN outperformed the SVM. When REVAC tuning parameters were used as shown in Fig 6b, the LSTM method still outperformed the other methods for most provinces and was closely followed by the CNN for most of the data-sets. However, in Gauteng province, the SVM out- performed the CNN. The results from the sensitivity study we conducted in Fig 9 shows that the relative impor- tance of each climate variable differs across provinces. For instance, over provinces such as Western Cape, Eastern Cape and Free State, the Pressure climate variable was the most sensi- tive when training any given diarrhoea outbreak prediction method. Whereas, in North West and Mpumalanga, Evaporation was the most sensitive climate variable. In Gauteng, Maximum Temperature was most important while in and KwaZulu Natal, Minimum Temperature was more sensitive. In Limpopo, Humidity was most sensitive variable while Wind speed was more important in the Northern Cape. Fig 7. Percentage change in performance of each ML method for predictions in Experiment III. (a) & (b): Percentage change in performance of each ML method over each province when predictions were made with the parameters from REVAC tuning instead of the grid-search parameters for (a) upward augmented data and (b) downward augmented data-set. High percentage RMSE indicates an improvement in task performance and vice-versa. https://doi.org/10.1371/journal.pone.0262008.g007

Fig 8. Provincial results of the ML methods with the augmented data-sets in Experiments II & III. (a) & (b): Results of the predictions with the augmented data-sets for each province (a) represents the results with grid-search tuned parameters and (b) represents the results with REVAC tuning parameters. Low RMSE averages indicate better task performance and vice-versa. https://doi.org/10.1371/journal.pone.0262008.g008 PLOS ONE Predicting diarrhoea outbreaks with climate change PLOS ONE | https://doi.org/10.1371/journal.pone.0262008 April 19, 2022 12 / 18