Digital twin: A machine learning approach to predict individual stress levels in extreme environments

Abstract

Remote Health Monitoring (RHM) has the potential to increase operational safety in extreme environments. Negative stress exposure influences mission success or the short- and long-term health conditions of deployed personnel. To quantify negative stress, we introduce a washable smart textile with integrated sensors. Analyzing the transmitted sensor values, medical advisors monitor up to 72 sensor values in parallel in case of an average group size of eight people. In order to aggregate the amount of data, we propose a stress level scale that includes stress trends. To predict individual stress levels based on sensor data, environmental quantities and the individual physiological fingerprint, we train different machine learning models. To evaluate such models, we implement a data acquisition environment to label data snapshots. Therefore, we do not need to collect in-field data and expose humans to negative stress. Moreover, we can mock sensor failures and rare, but relevant, sensor value combinations that are difficult to acquire in real-world scenarios. Our evaluation environment identifies Random Forest Regressor from a set of 25 models to perform best to predict individual stress levels. This model performs 23.19 times better than a zero rule classifier to distinguish among nine stress levels for mission goal health condition and 10.50 times better for mission goal mission success. Finally, we present our current RHM user interface design. It addresses issues such as information overload, avatar sympathy and unnecessary navigation paths.