AI Improves Health Predictions in Poor Nations

In low- and middle-income countries, health systems often struggle with delayed and incomplete data, hindering effective disease monitoring and resource allocation. A new study demonstrates that artificial intelligence can enhance the accuracy of predicting key health outcomes by integrating diverse data sources like satellite imagery and mobile records, offering a practical tool for strengthening public health in resource-limited settings.

The research team developed a multi-source geospatial foundation model (Multi-GeoFM) that combines satellite imagery embeddings, mobile call detail records, and internet search data to predict 15 routine health program outputs in Malawi. This model improved predictions for 87% of indicators compared to traditional statistical methods, with notable gains in estimating population density, HIV viral load suppression, and malaria case rates.

To build the model, the researchers used three geospatial foundation models: the Population Dynamics Model (PDFM) for satellite imagery, Google's AlphaEarth for search and map data, and mobile call detail records for movement patterns. These inputs were processed through machine learning algorithms, specifically XGBoost, and compared against baseline geostatistical methods like inverse distance weighting and kriging. The analysis covered 552 health catchment areas in Malawi from January 2021 to May 2023, using an 80/20 training-test split and cross-validation to ensure reliability.

Results showed that the Multi-GeoFM achieved cross-validated R² values up to 0.63 for population density and test set R² values of 0.64 for HIV viral load suppression, 0.68 for malaria cases, and 0.55 for child vaccinations. Satellite-based embeddings, particularly from PDFM, were most effective for count-based targets like population and antenatal care access, while mobile data underperformed, likely due to low smartphone penetration. The integration of multiple data sources consistently yielded the most robust predictions, outperforming single-source models and traditional approaches in most scenarios.

This approach matters because it enables more timely and accurate health monitoring in regions where data gaps can delay responses to outbreaks and resource distribution. For example, better predictions of malaria cases could help target mosquito control efforts, while improved HIV metrics might optimize treatment programs. The method leverages existing digital infrastructure without requiring expensive new data collection, making it scalable for other low-resource contexts.

Limitations include the study's focus on Malawi, which may not generalize to other countries with different health systems, and dependence on routine health data that can be incomplete or inaccurate. Indicators with small sample sizes, such as tuberculosis and malnutrition, were poorly predicted, highlighting the need for sufficient underlying data. Future research should explore optimal data combinations and temporal forecasting to enhance utility.