Epidemic Incidence Modeling and Forecasting Geospatially Using Machine Learning

Abstract

AbstractDue to climate changes and other socio-economic and anthropogenic in the world there is surgency in the epidemics and pandemics. Malaria is one of the oldest pandemics and still causing numerous deaths. Normally there is surge in the cases once the rainy season comes and it continues little over the rainy season. Normally multi-regression analysis-based empirical models have used to model the incidence and the rainfall, temperature, and relative humidity. Though the models developed were useful for public health departments, they often gave less accurate results than desired, and a need was felt to explore data science for a non-parametric solution. In this study we employed machine learning-based approaches and evaluated the results. The data came from Sundargarh District of the province of Odisha, India. Climate data used in the study include (1) block-wise station rainfall (RF) observation data, (2) gridded surface temperature (T2) data, and (3) Relative humidity (RH) data. High resolution gridded (with a horizontal resolution of 0.1°) ECMWF reanalysis land (ERA5-Land) data were used for observed temperature and relative humidity. The rainfall data was sourced from the Special Relief Commissioners’ web portal in Odisha. Malaria incidence data were collected from the state's Directorate of Public Health. Multilayer Perceptron (MLP) and J48 algorithm of Machine Learning were used to successfully predict malaria incidents. The two techniques’ performance evaluation was done using the classifier accuracy, Root Mean Square Error (RMSE), Kappa, and ROC scores. Classifier techniques such as the tenfold cross-validation and Percentile split (66%) and user-defined test options were used to divide the data into a training set and testing set. The study found out that the prediction accuracy in tenfold cross-validation with J48 decision tree classifiers was generally higher than the MLP. The results were encouraging, especially for the potential utilization of the model to predict future malaria incidence with higher confidence with access to weather forecasts. It was established that there is a medium-level malaria incidence in Odisha during periods of low rainfall, medium temperature, and lower relative humidity. Medium rain, high temperature, and low relative humidity are also associated with medium cases of malaria. Also, high temperatures, higher rainfall, and low relative humidity are characterized by a medium level of malaria incidence. Most importantly, it was established that high malaria incidence is associated with high or medium rainfall, low or medium temperature, and medium or very high relative humidity.KeywordsMachine learningMultilayer perceptron (MLP)J48 algorithmSpatial epidemiologyGIS for healthClimateHealth