Leveraging Feature Importance Methods to Characterize Explosive Ordnance Disposal Veterans' Health Profiles.

INTRODUCTION: Explosive Ordnance Disposal (EOD) Veterans have complex exposure histories that may contribute to chronic symptom burden. Identifying health features that are distinct to EOD Veterans may help describe this complex, chronic symptom burden. A machine learning approach to identify the unique health symptoms that can discriminate EOD and non-EOD Veterans who participated in the WRIISC (War Related Illness and Injury Study Center) clinical program will be presented to demonstrate these analytic methods. This approach orders clinical data to characterize EOD Veterans' health profiles to prioritize Veterans that may require additional clinical services, research opportunities, and focused education. MATERIALS AND METHODS: Clinical intake data from a battery of validated questionnaires and clinical measures acquired demographics, military occupational and deployment history, medical history, exposure concerns, and symptoms that were analyzed from EOD (n = 52) and non-EOD (n = 794) Veterans. Features were selected from these data using the Boruta algorithm, then used to train a balanced random forest classifier. Gini importances were then extracted from this model. RESULTS: Training of the balanced random forest classifier required 75% of the selected clinical data to achieve a Receiver Operator Curve, Area Under the Curve (ROC AUC) score of 0.93 on the remaining 25% of data. In total, 21 of the 143 features from the Clinical intake packet were selected using the Boruta algorithm. CONCLUSIONS: Characterizing EOD health profiles using aggregated clinical information supports efficient delivery of VA services. The results specify combat exposure, physical functioning, and TBI symptom burden rank the highest as vital measures to the classification of EOD Veterans. Results indicate attention should be given to combat exposure and to physical functioning when considering complex military exposure cases. These methods can be applied towards other complex multifactorial exposures and novel emerging threat concerns to monitor health trends and guide quality improvement activities.