Machine Learning Estimation of Stress and Mental Workload from Physiological Signals with Real-Time Closed-Loop Flight Simulation Demonstration

Abstract

The accurate, real-time estimation of pilot stress and mental workload (MWL) has the potentialto improve aviation safety and enable adaptive control systems to lessen the load on pilots. While the correlation of physiological features to stress and MWL is well-established in the literature, the design and optimization of machine learning pipelines capable of translating physiological features into real-time cognitive state estimates remains an open research problem. This research operates on a dataset derived from six physiological modalities: cardiac activity (ECG), electrodermal activity (EDA), respiratory activity, skin temperature, brain activity (fNIRS), and eye movement. Stress and MWL are classified on a four-point scale, with 0 as rest, 1 as low stress/MWL, 2 as medium stress/MWL, and 3 as high stress/MWL. A range of machine learning models are benchmarked across multiple pipeline configurations, while examining the effect of preprocessing choices including feature scaling and dimensionality reduction on classification performance, with Balanced Random Forest classifiers performing the best. A novel two-stage classification architecture is introduced, decomposing the problem into a rest vs. cognitively altered binary classification followed by a focused classification within altered states, improving pipeline macro F1 score by up to 14.6%. The trained models are subsequently applied to a real-time, closed-loop F-35 flight simulation. Pilot physiological data is feature extracted and sent to the classifier. Live MWL inferences are visually displayed to the pilot, closing the loop and validating the feasibility of the proposed approach under a simulator environment.