Human Thermal-Work Strain Performance Optimization from Wearable Physiological Sensors

of “Human Thermal-Work Strain Performance Optimization from Wearable Physiological Sensors” by Mark J. Buller, Ph.D., Brown University, May 2015. Hot environments pose a risk of heat illness for many emergency workers, athletes, and other professions especially when heavy workloads or protective clothing are necessary. Modern wearable physiological monitors may be able to mitigate risk of heat illness and improve performance if they are able to track health state and provide feedback to the user. However, effective algorithms and models to make use of wearable sensor information are lacking. We present two contributions: 1) a method for health state estimation of the latent human body core temperature from physiological sensors, and 2) models for policy estimation to provide automated advice to reduce thermal-work strain and improve physiological performance over a course of prescribed work. Continuous measurement of core body temperature, a requisite of thermal-work strain health state, has been an open physiology problem in the field. We show that the physiological dependencies of the human thermo-regulatory system can be cast into a dynamic Bayesian network model that allows us to estimate core body temperature from wearable physiological sensors. We effectively simplify this model to use only an input of heart rate which is collected by many commercial wearable sensor systems. This approach is validated across different combinations of temperature, hydration, clothing, and acclimation states, and shows similar comparison accuracy to accepted laboratory measures. We finally demonstrate the use and effectiveness of the algorithm from experimental trials during a first responder live training event. We also present a Markov decision process that uses health state estimates to optimize individual pacing strategies to reduce the overall level of thermal-work strain. We describe the estimation of real world activity objectives and thermal-work strain constraints as a reinforcement learning problem. Using a dynamical simulation of physiology, pacing estimates from this model are shown to reduce overall thermal-work strain. Our health state and policy estimation contributions were evaluated in the context of an implementation to compare human self-guided pace and policy guided pace. The results show that the policy allowed individuals to complete the task with meaningfully lower thermal-work strain. We demonstrate that real-time feedback from our model was able to match the thermoregulatory efficiency of a well-trained athlete. We envision the work in this dissertation will enable practical real-time monitoring systems that can improve human health through preventing thermal injury and use reinforcement learning to improve the physical performance of novice athletes and regular individuals. Human Thermal-Work Strain Performance Optimization from Wearable Physiological Sensors