The following method was used to create the maps:

1. We obtained district level diarrheal disease data for 75 districts of Nepal (July 2002-June 2014) from the Health Management Information System (HMIS: http://dohs.gov.np/). Likewise, data for Taiwan (2008-2019) was obtained from Taiwan Centers for Disease Control (https://www.cdc.gov.tw/),  and data for Vietnam (2000-2015) was obtained from Vietnam General Department of Preventive Medicine (VGDPM), Ministry of Health (http://www.vncdc.gov.vn). The temporal resolution of the data varied from weekly scale for Taiwan to monthly scale for Nepal and Vietnam.

2. We also obtained meteorological data for Taiwan, Nepal, and Vietnam from Global Historical Climatology Network (GHCN) through the National Climate Data Center (NCDC) data portal (https://www.ncdc.noaa.gov). Additional meteorological data for Nepal were obtained from the Department of Hydrology and Meteorology, Government of Nepal. Since some of the locations in Vietnam did not have weather station data, we obtained gridded ECMWF-ERA5 reanalysis products with 0.25 x 0.25-degree spatial resolution to be used for analysis of Vietnam data. All daily data compiled from multiple sources were grouped by month for Nepal and Vietnam, and by week for Taiwan, to match the temporal resolution of the outcome data.

3. To predict diarrhea rates for each region at a given time, we trained a shallow time-series neural network using past diarrhea rates, phases of ENSO, and weather information for that region.  In Shallow Neural Networks, the neurons in the input layer are connected to the neuron that produces the prediction through a second set of neurons referred to as the hidden layer. The term “shallow” refers to the fact that there is one single hidden layer of neurons as opposed to deep-learning neural networks that have more than one hidden layer. The “time-series” expression has to do with the capability of the software which implements these networks to include delayed values of the predictors as input parameters. In this case, we used a delay of 4 for all the independent variables

4. For each region in each country, we ran neural network analyses consisting of weather data (temperature, precipitation, and phases of ENSO as measured by the Oceanic Niño Index (ONI)) and average diarrhea disease rate for the same week (Taiwan) or month (Nepal/Vietnam). For each location, we used all available diarrheal disease data to calculate thresholds for classifying disease rate as low (0-25th percentile), medium (25th -75th percentile), high (75th-95th percentile and extreme (> 95th percentile). In the maps each region is painted with a color that represents the disease rate (green: low, yellow: medium; orange: high, red: extreme) predicted by the neural network for that given time period.