Predicting diarrhoea outbreaks with climate change

future. We tested the predictions of the three ML methods with respect to four different lag periods from all input features. The lag periods we considered include a lag of one (1) day, lag of five (5) days, lag of two (2) weeks and a lag of three (3) weeks. For example, a lag of one day means that the predictions made by a method for the 6th of January 2018 was made with input variables (for all features) of the 5th of January 2018 while a lag of five days means predictions for the 1st of January 2018 was made with input variables (for all features) of the 1st to the 5th of January 2018. These specific lag periods were chosen since our preliminary analyses show that they produce more accurate predictions. Thereafter, optimal parameters were selected and we determined the best performing ML method by comparing the RMSE from the predictions made by the three ML algorithms (with respect to the four lag periods) in three different experiments in which for each ML method, predictions were repeated three times for each lag, across each province and the average RMSE result was computed. The first experiment (Experiment I) was implemented with the real-world case data which contained the diarrhoea cases and eight climate features. The objective was to determine which ML method performs best given the amount of data instances contained in the real- world data-set. In order to obtain optimal training parameter values for each ML method across each province, the grid-search method was used in this experiment. For most deep neu-

ral networks, such as the CNNs, the computational complexity, can be computed as O(n2) for both training and inference time, where n is the input dataset size [35]. However for networks that deal with sequential learning such as the LSTM, their learning complexity per time step is O(W), where W is the number of parameters in a standard network [36]. The computational Fig 5. Pipeline of the daily diarrhoea prediction model. https://doi.org/10.1371/journal.pone.0262008.g005 Table 2. Experiments overview. Experiment description Parameter tuning technique Datasets used Research objective (I) Predictions with real-world data Grid-search Real-world data Determine best predicting method given real-world data (II) Predictions with augmented data Grid-search Upward and downward augmented data Determine the effect of augmented data on predicting performance (III) Predictions with augmented data and REVAC parameters REVAC tuning Upward and downward augmented data Determine the impact of REVAC tuning on predicting performance https://doi.org/10.1371/journal.pone.0262008.t002 PLOS ONE Predicting diarrhoea outbreaks with climate change PLOS ONE | https://doi.org/10.1371/journal.pone.0262008 April 19, 2022 9 / 18