Predicting diarrhoea outbreaks with climate change

Methods Study population This study focused on the nine South African Provinces which are: Western Cape, Eastern Cape, Northern Cape, North West, Free State, Limpopo, KwaZulu Natal, Gauteng, and Mpu- malanga. Most provinces in South Africa experience rainfall in the summer with the excep- tion of Western Cape. Western Cape has a Mediterranean climate that receives rainfall during winter with an average annual rainfall of 515mm. Provinces such as KwaZulu Natal, Free State and Mpumalanga experience the highest annual rainfall rate which is between 800–1054mm while Eastern Cape, Limpopo, Gauteng, Northern Cape, and North West prov- ince receive an annual rainfall that is between 400–600mm. In terms of temperature condi- tions, Limpopo, Northern Cape, Mpumalanga and North West provinces usually record the highest temperature with annual averages between 27.1–30˚C while the least annual average temperatures which are between 22.1–23.3˚C are usually recorded for Western and Eastern Cape provinces. Datasets The datasets used for all experiments consists of nine features categorized into two data sub- sets: Diarrhoea and a set of eight climate features. For each province, daily sales records of Loperamide, an anti-diarrhoea compound that has been evaluated in the treatment of patients with chronic non-specific diarrhoea in South Africa and other parts of the world was obtained from Clicks Group Limited, South Africa (https:// www.clicksgroup.co.za/). The data contains a 10-year period of total number of loperamide

purchased between November 2008 and March 2018. This data was used as a proxy for diar- rhoea cases in the region. In this study, the number of diarrhoea cases per day for a specific province was computed as the number of loperamide sales per day associated with the prov- ince. Six-hourly data on Maximum temperature, Minimum temperature, Air temperature, Spe- cific humidity, Potential evaporation rate, Precipitation rate, Surface pressure, and Wind velocity climate factors for each South African province between the period of November 2008 and October 2019 were obtained from the National centres for Atmospheric Research and Atmospheric Prediction. Please see (https://psl.noaa.gov/). Generative Adversarial Networks (GANs) [25] were used to generate 20, 000 synthetic time- series samples with 24 time-steps each for the diarrhoea and eight climate data in each prov- ince. Data augmentation was performed to have sufficient data for making predictions, where synthetic data was augmented with the real-world data-sets in two ways: upward augmentation and downward augmentation. When the data-sets were augmented upwards, the training set included a combination of the real-world and synthetic samples, but the test set included only the synthetic data-sets and when the data-sets were augmented downwards, the training set included mainly the synthetic data-sets and the test set included the real-world data-set. Tech- nical details on GAN implementation can be seen in S1 Appendix. The violin plots in Fig 1 show the distribution of the augmented dataset used in the study

for each province. The distribution of the diarrhoea case variable (loperamide) is similar across Western Cape, KwaZulu Natal and Gauteng with Western Cape having the highest spread of cases among all provinces. The distribution of the pressure variable is shown to be symmetric across all provinces, meaning that its values occur at regular frequencies while the precipitation variable is positively skewed thus, the mean value for each province is greater than the median. The distribution of the other climate variables is shown to be approximately identical across provinces. PLOS ONE Predicting diarrhoea outbreaks with climate change PLOS ONE | https://doi.org/10.1371/journal.pone.0262008 April 19, 2022 4 / 18