Predicting malaria outbreaks using earth observation measurements and spatiotemporal deep learning modelling: a South Asian case study from 2000 to 2017

Malaria remains one the leading communicable causes of death. Approximately half of the world's population is considered at risk of infection, predominantly in African and South Asian countries. Although malaria is preventable, heterogeneity in sociodemographic and environmental risk factors over time and across diverse geographical and climatological regions make outbreak prediction challenging. Data-driven approaches accounting for spatiotemporal variability could offer potential for location-specific early warning tools for malaria. In this case study, we developed and internally validated a data fusion approach to predict malaria incidence in Pakistan, India, and Bangladesh using geo-referenced environmental factors. For 2000–17, district-level malaria incidence rates for each country were obtained from the US Agency for International Development's Demographic and Health Survey datasets. Environmental factors included average annual temperature, rainfall, and normalised difference vegetation index, obtained from the Advancing Research on Nutrition and Agriculture (known as AReNA) project conducted by the International Food Policy Research Institute in 2020. Data on night-time light intensity was derived from two satellites of the National Oceanic and Atmospheric Administration Defense Meteorological Satellite Program—Operational Linescan System: Nighttime Lights Time Series Version 4, and VIIRS Nighttime Day/Night Band Composites version 1. A multi-dimensional spatiotemporal long short-term memory (M-LSTM) model was developed using data from 2000–16 and internally validated for the year 2017. The M-LSTM model consisted of four hidden layers, each with 100 LSTM units; a fully connected layer was used, followed by linear regression, to predict the malaria incidence rate for 2017 using spatiotemporal partitioning. Model performance was measured using accuracy and root mean squared error. Country-specific models were produced for Bangladesh, India, and Pakistan. Bivariate geospatial heatmaps were produced for a qualitative comparison of univariate environmental factors with malaria rates. Malaria incidence was predicted with 80·6% accuracy in districts across Pakistan, 76·7% in districts across India, and 99·1% in districts across Bangladesh. The root mean squared error was 7 × 10–4 for Pakistan, 4·86 × 10–6 for India, and 1·32 × 10–5 for Bangladesh. Bivariate maps showed an inverse relationship between night-time lights and malaria rates; whereas high mal