A Malaria early warning alert application for health officials (Imalaria)

Abstract

Malaria is a critical public health issue that must not be overlooked since it affects majority of the world’s population. The goal of the study was to determine how to use confirmed malaria case data, and current climatic conditions to forecast malaria epidemics and offer timely information to health officials. This research will help to inform and build a malaria early warning alert application that can offer public health experts with early warning signs and prediction of malaria outbreaks. The first step was selection of the variables that has a significant impact on malaria transmission. From literature review, malaria incidences, rainfall/precipitation, humidity, and temperature were selected. Data collection was done where malaria data was collected from Kenya Health Information System (KHIS) and incidence rates calculated based on cases and population. Weather data was also collected from historical data API accessed through this link: https://open-meteo.com/. Hourly data collected was aggregated and averaged for the month derived. Selection of the best machine learning algorithm for the malaria early warning alert system was conducted and six models including Decision Tree Regressor, Linear Regression, Random Forest Regressor, Ada Boost Regressor, Gradient Boosting Regressor, SVR and decision tree were tested. Random forest Regressor which had the best score was selected. Training of the model on the collected data to make predictions and provide alerts to health officials at facility level was done using 70% of the data. Evaluation of the performance of the model using metrics such as Mean Absolute Error (MAE), Mean Squared Error (MSE) and Root Mean Squared Error RMSE was also conducted using 30% of the data. Deployment of the model was done using TensorFlow's Saved Model format which was integrated into the React application to create the malaria early warning alert system. The UNHCR model for emergency response was adopted to calculate the alert/epidemic threshold. This model considers the threshold for a malaria outbreak to be 1.5 times the baseline over the previous three weeks. Alert (or ‘epidemic threshold’) and action thresholds were used to provide information to the public health staff. The rest of the data was labelled as normal threshold.