Predicting Operator Workload from Oculometric Data in High-Demand Environments: A Case Study with MATB-II

Abstract

In an increasingly dynamic and fast-paced social context, many professionals are exposed to conditions of cognitive stress that often lead to burnout. Mental workload can be empirically described as the ratio between the available mental resources and the demands imposed by the task. Given this parameter's inherently subjective and non-directly measurable nature, its assessment can be carried out through three main approaches: subjective self-assessment, behavioral analysis, or monitoring of variations in physiological signals. When envisioning the development of a device capable of autonomously estimating an individual's mental workload to support their professional activity, it becomes clear that the self-assessment approach is inapplicable. While behavioral analysis holds interesting potential, it suffers from limitations related to the difficulty of generalizing across heterogeneous scenarios. In contrast, the physiological approach is the most promising, as it allows for monitoring independent of the task. Based on these considerations, the present work aims to develop a system capable of estimating various levels of mental workload using solely ocular signals that can be easily captured through wearable devices such as smart glasses or non-invasive optical devices, like remote cameras.An experimental campaign was conducted with 40 subjects, using the MATB-II (Multi-Attribute Task Battery II) test as the primary task. NASA designed this test to stimulate cognitive workload and multitasking abilities, simultaneously activating various mental processing pathways: visuomotor coordination via tracking, auditory reflex through radio communications, logical reasoning through pump settings, and visual reflex through buttons and control bar management. Due to this multi-level structure, the MATB-II provides complete stimulation of the cognitive spectrum of participants. To explore the full range of mental workload, from minimal to very high levels, a secondary task based on simple arithmetic operations (single-digit addition, subtraction, and multiplication) was integrated into the primary test, with four response options and a maximum execution time. Each experimental session was structured with an initial rest phase to acquire baseline values, followed by five MATB-II trials of increasing difficulty, each interspersed with three minutes of relaxation to restore baseline physiological conditions.From each of the six phases (including the rest phase), a set of features derived from the ocular signal was extracted, which were subsequently used as input for a machine learning training pipeline. The data from each subject were initially corrected against their baseline and then normalized via min-max transformation. The pre-processed data were analyzed to extract the most essential features and then fed into various machine learning classification models. The results show high predictive reliability, outlining promising scenarios for developing automated systems for estimating mental workload. Most tested models can easily separate the baseline and the lower workload state. The distinction between the higher mental workload classes is less evident but still statistically significant.An additional perspective, currently still exploratory, involves analyzing model performance in real-time contexts, using short time windows. This extension would make the system applicable in a wide range of cognitively intensive operational fields, including the automotive, aerospace, and medical sectors, contributing to developing intelligent technologies for continuously monitoring operators' mental states.

Keywords: Ocular Signal, Human-Machine Interface (HMI), Machine Learning (ML), MATB-II, Mental Workload (MWL)

DOI: 10.54941/ahfe1006885