With an ever changing, fast paced, global initiative, the U.S. must adapt to the increasing operational tempo of modern warfare. The need to rapidly adapt to changing conditions has contributed to the growing number of Soldiers experiencing fatigue-related injuries. These injuries cost the military billions of dollars annually, while impacting training opportunities and reducing the number of combat-ready Soldiers that support critical missions. Existing methods of injury prevention rely on Soldiers’ self-reporting their fatigue state which can be less reliable compared to objective physiological, cognitive, and physical indicators of fatigue. Additionally, lack of methods for early warning and prediction of emergent fatigue conditions hinders the military’s ability to proactively intervene and reduce the likelihood of injury. Motivated by these challenges, a machine learning framework was developed to predict fatigue with individual multimodal data collected as part of the US Army supported, MASTR-E (Measuring and Advancing Soldier Tactical Readiness and Effectiveness) program. Through the MASTR-E program, this study leveraged a subset of data collected from over 200 Soldiers during a 72hour training exercise. Collected data included baseline physical and cognitive status, demographics, standard surveys tracking information across health, physical, social-emotional, and cognitive domains, biomarker data from blood and saliva samples, and physiological and kinematic data from wearable sensors. Using the available data, a multimodal machine learning framework was developed to process data and perform supervised learning to predict individual fatigue up to 60 minutes forward in time. Fatigue states are labeled based on acute to chronic workload ratio (ACWR) along with training impulse, which enables accurate prediction of excessive fatigue conditions that can lead to injuries. The framework enables tradeoff analysis of different data types and prediction horizons for enhanced decision support. Demonstration of the framework achieved 70% balanced accuracy when predicting 30 minutes forward, with explainable AI embedded to provide data driven actions to mitigate injuries and better assess soldier performance.
Keywords
AI;ANALYTICS;DEEP LEARNING;FRAMEWORK;HUMAN PERFORMANCE;INTEGRATION;MACHINE LEARNING;WEARABLE DEVICES
Additional Keywords