Predicting diarrhoea outbreaks with climate change

In Eq (1), xi is the actual value while yi is the predicted value and n is the total number of observations to be analysed. The ML method with the smallest RMSE error is considered to be the best performing method in terms of prediction accuracy. Configuration of ML methods This study adopted two popular deep learning methods namely CNNs, LSTMs and a tradi- tional ML method SVM for all experiments. These methods were chosen because of their suc- cess in time series predictive tasks such as [14, 33]. Asides the powerful feature representation capabilities of deep learning models, the LSTM network is a powerful technique for analyzing temporal data. While the existence of other traditional ML methods such as decision trees [34] and ARIMA [12] are known, SVM was chosen because it is a widely used nonlinear regression estimation technique [16]. In addition, our preliminary analysis showed that the chosen ML methods outperforms the decision trees (see the S1 Appendix section). The rest of this section provides details on how the chosen methods were implemented. CNN method. CNNs are a class of feed forward, deep neural network that consist of mul- tiple convolutional and activation layers, pooling layers, and a fully connected layer as shown in Fig 2. These layers are designed to perform specific tasks in order to extract important fea- tures from the input data. After several iterations of convolutions, node activations and pool- ing the final output is computed in the fully connected layer of the network. Our CNN method

was designed with 1D convolutions to match the sequential nature of our input data. LSTM method. LSTMs as shown in Fig 3 are examples of Neural Networks under the cat- egory of Recurrent Neural Networks (RNNs) that address the issue of exploding and vanishing gradients. They contain memory cells that maintain their state overtime. The memory cells are managed by gating units that control how it memorize, erase, and expose information. These gating units are called the input gate, forget gate and output gate respectively. SVM method. SVMs are mathematical models whose main function is to find hyper- planes capable of creating margins that separates data points in a high dimensional feature Fig 2. Basic architecture of the Convolutional Neural Network (CNN) with two convolution and pooling layers. https://doi.org/10.1371/journal.pone.0262008.g002 PLOS ONE Predicting diarrhoea outbreaks with climate change PLOS ONE | https://doi.org/10.1371/journal.pone.0262008 April 19, 2022 6 / 18