Predicting clinical outcomes of chemical war veterans exposed to mustard gas using machine learning models

• Lung damage was the most prevalent severe complication (6,151 cases, 10%), followed by skin (0.7%, 439 cases) and eye injuries (0.3%, 193 cases), highlighting the disproportionate impact on respiratory health. • Neural networks achieved the highest predictive accuracy for mortality (AUC: 90%) and severity of lung (AUC: 93%), skin (AUC: 90%), and eye (AUC: 94%) complications, with lung and eye severity prediction reaching 92.4% accuracy. • Age at exposure and overall severity of organ damage were identified as critical distinguishing factors in clinical outcome predictions. • Machine learning models demonstrated robust performance, with mean precision, correctness, and coverage metrics exceeding 85%, underscoring their utility in clinical decision-making for chemical injury outcomes. During the eight-year Iran-Iraq war, Iran experienced 387 chemical attacks exposing approximately 100,000 individuals. This study focuses on all chemical veterans in the country specifically those exposed to mustard gas between 1982 and 1987. In modern medicine, comprehensive data collection across diverse diseases is critical. This research aims to predict clinical outcomes for mustard gas-exposed chemical veterans using machine learning models. Variables included age, gender, severity of organ damage (lungs, skin, and eyes), exposure level, exposure date and date of death. A comparative machine learning framework was applied to evaluate Decision Trees, Random Forests, Neural Networks, Support Vector Machines, Nave Bayes and Generalized Linear Models. Statistical analyses were performed using SPSS and R software. The study analyzed data from 64,190 chemical veterans (98% male; mean age, 56±8 years). The mean age at exposure was 23±8 years (range: 2–86). Severe complications were recorded in the lungs (10%, n=6,151), skin (0.7%, n=439), and eyes (0.3%, n=193). Predictive performance measured by AUC was highest for lung severity (94%) followed by eyes (94%), skin (90%) and mortality (90%). Model accuracies were 92.4% (lungs), 91.0% (skin), 92.4% (eyes) and 91.1% (mortality). Age at exposure and overall severity emerged as primary predictors. The neural network and random forest models exhibited superior performance in predicting mortality and severity of skin and lung conditions. With average accuracy, precision, recall and F1-scores all exceeding 85%, these findings demonstrate the high predictive capability of machine learning algorithms in forecasting clinical outcomes for chemical veterans exposed to mustard gas.