Predicting Malaria Incidence Through Machine Learning: A National-Level Time-Series Approach Using Malaria Atlas Data

Abstract

Accurate forecasting of parasitic disease epidemics is crucial to enable early public health responses, particularly in regions such as endemic malaria areas. The research work herein proposes a predictive model framework that integrates historical epidemiological data for predicting national malaria incidence rates using machine learning algorithms. Using data from the Malaria Atlas Project, including incidence rate, infection prevalence, and mortality rate from 2010 to 2022, we trained and tested different predictive models like Long Short-Term Memory (LSTM) networks, Random Forests, and Linear Regression. Results indicate that all models gave high predictive accuracy, with the best overall performance from Random Forest (R² = 0.981), then Linear Regression (R² = 0.978) and LSTM (R² = 0.972). At a country level, LSTM did an excellent job at picking up long-term incidence trends, particularly in Uganda and Kenya, but performed less consistently where there were changing patterns, as in Tanzania. The study identifies the importance of machine learning for disease surveillance and early warning and recommends future use of environmental and subnational data to improve spatial resolution and prediction accuracy.