The Application of Machine Learning to Identify Large Animal Blast Exposure Thresholds.

INTRODUCTION

The Office of Naval Research sponsored the Blast Load Assessment Sense and Test program to develop a rapid, in-field solution that could be used by team leaders, commanders, and medical personnel to make science-based stand-down decisions for service members exposed to blast overpressure. Toward this goal, the authors propose an ensemble approach based on machine learning (ML) methods to derive a threshold surface for potential neurological deficits that encompasses the intensity of the blast events, the number of exposures, and the period over which the exposures occurred. Because of collection challenges presented by human subjects, the authors utilized data representing a comprehensive set of measures, including structural, behavioral, and cellular changes, from preclinical large animal studies on minipig models. This article describes the development process used to procure the resulting methodology from these studies.

METHODS AND MATERIALS

Using an ensemble of ML methods applied to experimental data obtained from 71 Yucatan minipigs, the relationship between blast exposure and neurological deficits was delineated. Despite a relatively small sample size, ML methods with k-fold cross-validation (with k = 5) were justified because of the complexity of the dataset reflecting numerous nonlinear relationships between cellular, structural, and behavioral markers. Based on the physiological responses and environmental measures collected during the large animal study, two models were developed to investigate the relationship between multiple outcome measures and exposure to blast. The histological features model was trained on single-exposure animal data to predict a binary injury response (injured or not) using histological features. The environmental features model related the observed behavioral changes to the environmental parameters collected.

RESULTS

The histological features model predicted a binary injury outcome from cellular and physiological measurements. Features identified in developing this classification model showed some level of correlation to observed behavioral changes, suggesting that glial activation inflammation and neurodegenerative responses occur even at the lowest levels of blast exposures tested. The results of the environmental features model, which estimated injury risk from environmental blast exposure characteristics, suggested that the observed changes are not just a function of impulse but an average dynamic impulse rate. Noticeable behavioral deficits were observed at loading rates of 100 kPa (impulse/positive duration) or peak pressures of 300-350 kPa, with an approximate positive phase duration of 3.4 ms for single exposure. Based on this analysis, a 3D threshold surface was developed to characterize the potential risk of neurological deficits.

CONCLUSIONS

The ensemble approach facilitated the identification of a pattern of changes across multiple variables to predict the occurrence of changes in brain function. Many changes observed after blast exposure were subtle, making them difficult to measure in human subjects. ML methodologies applied to minipig data demonstrated the value of these techniques in analyzing complex datasets to complement human studies. Importantly, the threshold surface supports the development of science-based blast exposure guidelines.