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

Pogliano, Marco; Colavincenzo, Manuel; Martorana, Stefano; Guglieri, Giorgio; Demarchi, Danilo · 2025 · Crossref

DOI: 10.54941/ahfe1006885

archive: archived pipeline: cataloged

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This study addresses the challenge of real-time mental workload (MWL) monitoring in safety-critical environments, such as avionics and healthcare, where accurate cognitive state estimation is vital for adaptive automation. While physiological monitoring is a promising approach, most existing methods rely on multimodal signals that require invasive or burdensome preparation. This paper investigates whether MWL can be reliably predicted using solely non-contact ocular signals, which are easily captured via wearable eye-tracker glasses or remote cameras. The goal is to develop a machine learning pipeline that classifies MWL states without the logistical constraints of traditional physiological sensors. The experimental design involved 28 participants (22 male, 6 female; mean age 29) who performed the Multi-Attribute Task Battery II (MATB-II), a multitasking test designed to induce varying cognitive loads through visuomotor, auditory, logical, and reflex tasks. A secondary arithmetic task was also included to broaden the workload spectrum. Ocular data were recorded using Tobii Pro Glasses 3 at 50 Hz. The protocol included a 5-minute baseline rest period followed by five MATB-II trials of increasing complexity, separated by 3-minute rest periods. During these rests, participants provided subjective MWL ratings using a custom six-point scale (A–F) adapted from the Bedford scale. These subjective ratings served as the ground truth labels. Ocular features extracted included blink duration and frequency, fixation metrics, saccade velocity and frequency, saccadic intrusions, and spectral power distributions (low-frequency and high-frequency bands) of the gaze signal. The six subjective ratings were grouped into four classes: Rest, Low (A–B), Moderate (C–D), and High (E–F). The data were normalized per subject to minimize inter-individual variability and split into training and test sets using a subject-wise 80/20 partition to prevent data leakage and ensure generalization to unseen individuals. A machine learning pipeline combined feature selection techniques (SelectKBest, L1-based importance, Random Forest importance) with classification algorithms (Logistic Regression, Random Forest, SVM, XGBoost, etc.), optimized via 5-fold stratified cross-validation. The best-performing model achieved a macro F1-score of 0.77. Performance was highest for the Rest (F1 = 0.89) and Low MWL (F1 = 0.87) classes, with perfect precision for Rest. Performance degraded for Moderate (F1 = 0.71) and High (F1 = 0.62) MWL, indicating confusion between adjacent higher workload levels. An empirical online validation on a subset of 5 participants confirmed that the model could track MWL fluctuations in real-time using sliding windows, correctly identifying high-demand segments, though with some latency in state transition detection. The findings demonstrate that ocular signals alone contain sufficient information to distinguish between distinct MWL states, particularly low and high workload, using a non-invasive, wearable-compatible approach. The study highlights that while the model generalizes well to new subjects, differentiating intermediate workload levels remains challenging due to subjective variability in self-reporting and the granularity of the MATB-II task. This work underscores the feasibility of ocular-based MWL monitoring for real-time applications like pilot monitoring and adaptive automation, while identifying the need for more objective labeling methods and higher temporal density of assessments to improve accuracy in the moderate-to-high workload range.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.8-27b-gittensor on 2026-08-23; verification: pending re-verification.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• workload measurement
• drowsiness detection algorithms
• neuro workload indices
• gaze based attention detection

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Empirical Findings: physiological data
• Methodological Resource: metric or index, measurement protocol