Mapping war trauma: A machine learning approach to predict mental health impacts in Ukraine

Safiyeh Tayebi; Ayse Sert Oti; Hossein Fathollahian; Ubydul Haque

doi:10.1016/j.ssmph.2025.101879

Mapping war trauma: A machine learning approach to predict mental health impacts in Ukraine

SSM Popul Health. 2025 Nov 15:32:101879. doi: 10.1016/j.ssmph.2025.101879. eCollection 2025 Dec.

Authors

Safiyeh Tayebi¹, Ayse Sert Oti², Hossein Fathollahian³, Ubydul Haque^{4

5}

Affiliations

¹ Department of Geography, Rutgers, The State University of New Jersey, 54 Joyce Kilmer, USA.
² Public Informatics, Rutgers Edward J. Bloustein School of Planning and Public Policy, Rutgers, the State University of New Jersey, 33 Livingston Ave, New Brunswick, USA.
³ Electronic Visualization Laboratory (EVL), Department of Computer Science, University of Illinois Chicago, Chicago, IL, USA.
⁴ Rutgers Global Health Institute, New Brunswick, NJ, USA.
⁵ Department of Biostatistics and Epidemiology, School of Public Health, Rutgers University, Piscataway, NJ, USA.

Abstract

This paper presents a predictive modeling system based on the use of GeoAI to estimate mental health outcomes of wartime in Ukrainian cities, utilizing spatially linked data on the environment, infrastructure, and conflict. Six self-reported psychological outcomes, symptoms of post-traumatic stress disorder, or PSD, anxiety, depression, insomnia, loneliness, and sleep duration, were collected in surveys throughout the war and analyzed in the context of more than 30 spatial predictors: cold exposure, access to heating, power outages, housing insulation, and city-level frequency of drone, missile, artillery, and shelling attacks. Predictor datasets that are derived from incident tracking, World Health Organization, and humanitarian reporting systems, and environmental indicators derived from surveys, which are harmonized using a spatial data integration protocol. In realizing the GeoAI concept, we developed a machine learning pipeline utilizing Ordinary Least Squares, Lasso, Random Forest, Gradient Boosting, and Extreme Gradient Boosting. All models were trained and tested on spatially independent training and testing splits in order to preserve the generalization properties of the models. XG Boost is also shown to be effective, with R2 values often exceeding 0.74 and MAPE values typically less than 7.2 %. Feature importance analysis revealed that key drivers of being exposed to prolonged cold, inadequate insulation, and exposure to drones or artillery were found to be dominant drivers of psychological distress. This GeoAI framework combines the strength of geospatial analytics with artificial intelligence to give precise and high-resolution location-based forecasting of mental health burdens in conflict settings. The method offers a flexible tool for a proactive humanitarian response that can target mental health services spatially based on predictions of mental health vulnerability, in contrast to retrospective clinical information, for relief agencies and public health planners. This work is a step towards incorporating GeoAI in the field of crisis epidemiology, demonstrating the ability to use GeoAI in real-time, place-based mental health planning in war-affected areas.

Keywords: Environmental stressors; GeoAI (geospatial artificial intelligence); Predictive modeling; Spatial machine learning; Wartime mental health.