Predicting diarrhoea outbreaks with climate change

complexity of SVM on the other hand is On3 [37]. This shows that computational complexity of each model is different hence training and test time will also differ. In other to address this, the average run-time of each model was computed for this experiment. (see the S1 Appendix section more details). After concluding the first experiment, we measured the degree of importance of each cli- mate variable to the best performing diarrhoea prediction method in a specific province by conducting a sensitivity analysis [38]. We adopted the Backward stepwise method [38] in which we measured the effect of one variable at a time while keeping the other variables fixed. Sensitivity is then measured by observing changes in the RMSE error of the given method based on the omission of a certain variable. The larger the increase in RMSE, the higher the importance of the omitted variable. The second experiment (Experiment II) was conducted to determine the effect of augmented training and testing data as well as the effect of a larger training data size on the prediction performance of the three ML methods. The data-sets used in this experiment were combinations of the synthetic and real-world data-set, that is, the upward and downward augmented data in each province. Predictions by each ML method were made with each input data-set separately for each province. The data preprocessing steps and the parameters selected by the grid-search tuning in the first experiment were maintained for each ML method with regards to a specific province.

The third experiment (Experiment III) was performed to determine the effect of REVAC parameter tuning on the prediction performance of the three ML methods with the upward and downward augmented data. The major difference between the second and third experi- ment is the method used for tuning the parameters of each ML method. For all the prediction tasks carried out in the third experiment, data preprocessing steps taken for the three ML methods were the same as the previous experiments. However, the parameter values of each ML method were tuned with REVAC tuning method. Once the REVAC parameter tuning tasks were completed for each ML method, the fittest set of parameter values for each province were used to carry out final predictions. Results Table 3 represents the average RMSE for predictions made with real-world data in all prov- inces. We observed that the high performance of the CNN method was closely followed by the LSTM method. SVM on the other hand showed the poorest performance. Table 3 also showed that the CNN method had the least overall RMSE average of 31.55% while LSTM and SVM averages were 32.91% and 33.89% respectively. We can infer from these results that the RMSE errors are lower for the deep learning methods (CNN & LSTM). Fig 6 shows that the use of augmented data greatly improved the performance of the three ML methods in each province. Predictions for Limpopo province show the highest improve- ment with over 50% increase for each ML methods when both upward and downward aug-

mented data were used for predictions. However, over most provinces, the percentage increase Table 3. Root Mean Square Error (RMSE) averages for predictions using real-world data. ML method RMSE Convolutional Neural Network (CNN) 31.55% Long-term Short Term Memory (LSTM) 32.91% Support Vector Machine (SVM) 33.89% Standard Deviation 0.008 https://doi.org/10.1371/journal.pone.0262008.t003 PLOS ONE Predicting diarrhoea outbreaks with climate change PLOS ONE | https://doi.org/10.1371/journal.pone.0262008 April 19, 2022 10 / 18