Machine Learning–Based Prediction of Structural MRI Measures from Military Environmental Exposures in Veterans
Jasmin M. Francisco
Mentor: Dr. Peter Bayley, PhD, War Related Illness and Injury Study Center (WRIISC), U.S. Department of Veterans Affairs; Department of Psychiatry and Behavioral Sciences, Stanford University School of Medicine; Dr. Ansgar Furst, PhD, War Related Illness and Injury Study Center (WRIISC), U.S. Department of Veterans Affairs; Department of Psychiatry and Behavioral Sciences and Department of Neurology and Neurological Sciences, Stanford University School of Medicine.
Date/Time: August 21st, 2025 at 1:45 PM.
Abstract: Approximately 250,000 Veterans of the 1991 Gulf War experience Gulf War Illness (GWI), a condition marked by chronic health symptoms such as fatigue, musculoskeletal pain, neurological complaints, respiratory problems, gastrointestinal issues, and dermatologic conditions. Despite decades of research, these symptoms often persist, leading to poorer overall health and a higher burden of chronic medical conditions compared with non-deployed Veterans. Because the underlying biological mechanisms of GWI remain unclear, advanced tools are needed to move beyond symptom-based descriptions. Magnetic resonance imaging (MRI), combined with detailed exposure histories, offers a promising approach for uncovering objective neural markers of GWI that could shed light on disease mechanisms and support the development of targeted diagnostic and therapeutic strategies. Identifying such biomarkers remains a critical step toward improving care for affected Veterans.
MRI data (n = 286) from Gulf War Veterans evaluated at the VA War Related Illness and Injury Study Center (WRIISC) were processed with FreeSurfer and analyzed to assess whether environmental exposures could predict structural brain variation. For each MRI-derived metric, the single most correlated exposure was selected as the predictor, and linear regression models with 10-fold cross-validation were applied. The highest-performing model predicted cortical thickness in the right medial orbitofrontal cortex based on chemical nerve agent exposure, but the predictive power was weak (mean R² = 0.072), and dropped further after adjusting for age, sex, and traumatic brain injury (mean R² = 0.017). Across more than 7,000 MRI–exposure prediction tests, linear models showed weak explained variance (mean R² ≈ 0.10). Although some models produced modest error sizes, the variance explained was minimal, underscoring the limited predictive value of these models
These findings suggest that predicting brain structure from single-exposure models is highly challenging, particularly in the context of Gulf War Illness, where the impact of military exposures may be minimal compared with the broad range of chronic symptoms and comorbidities affecting Veterans. The difficulty may reflect both the complexity of GWI and the relatively small contribution of individual exposures to brain variation. Future research should incorporate additional covariates known to influence brain changes, include longitudinal MRI data, and recruit healthy control groups to strengthen biomarker discovery and improve diagnostic and therapeutic strategies.