An Explainable Machine Learning Framework for Predicting West Nile Virus and Uncovering Key Transmission Drivers
TL;DR
DVI models could be used complementarily by health authorities to aid their decision‐making process and to provide critical insights into important drivers that influence the transmission of WNV.
Abstract
This study introduces Disease Vector Intelligence (DVI), an explainable Machine Learning (ML) framework developed within the EYWA (EarlY WArning system for mosquito‐borne diseases) ecosystem. DVI integrates big Earth Observation (EO), socioeconomic, and estimated mosquito abundance data to predict the presence or absence of West Nile Virus (WNV) cases, expressed as a risk score at a high spatiotemporal resolution. The SHAP (SHapley Additive exPlanations) method is employed for a local‐level feature importance analysis to uncover key drivers with a high impact on the prediction. The framework consists of two ML models: one intended to be used as an Early Warning System (EWS) predicting at a regional scale, and the other identifying areas susceptible to WNV transmission at a 2 × 2 km2 grid scale, independent of human‐defined administrative boundaries. Both models were trained on 11 years of historical WNV case data and validated in Greece. SHAP analysis on both models revealed temperature as a crucial driver of WNV transmission. The fine spatial resolution of the second model uncovered micro‐scale key drivers, such as elevation and land cover type, that the lower resolution model missed. DVI models could be used complementarily by health authorities to aid their decision‐making process and to provide critical insights into important drivers that influence the transmission of WNV.