Effect of time length of data analysis on the accuracy of mental workload estimation during automobile driving with eye-movement

CHIHARA, Takanori; SAKAMOTO, Jiro · 2020 · Crossref

DOI: 10.1299/jsmedsd.2020.30.1404

archive: archived pipeline: cataloged verified

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

Summary

This study investigates the optimal time window for calculating eye and head movement parameters to estimate mental workload (MWL) during automobile driving. The research is motivated by the need to monitor driver MWL to prevent inattentive driving, a leading cause of traffic accidents. While previous work identified specific eye and head movement metrics useful for MWL estimation, the relationship between the duration of data analysis and estimation accuracy remained unexamined. Longer analysis windows may reduce noise but decrease temporal resolution, creating a trade-off that requires optimization for effective real-time monitoring. The experiment involved 12 healthy students (ages 21–22) performing driving tasks on a simulator while concurrently engaging in N-back tasks to manipulate cognitive load. Five conditions were tested: no N-back task (None), 0-back, 1-back, 2-back, and 3-back. Eye and head movements were captured using an image sensor at approximately 21.8 Hz. Four parameters were derived: standard deviation of gaze angle, standard deviation of eyeball rotation angle, the ratio of head movement to gaze movement, and blink frequency. These parameters were calculated using sliding time windows of 30, 60, 90, 120, and 150 seconds. A one-class support vector machine (OCSVM) was trained on "None" condition data to detect anomalies indicative of high MWL. Performance was evaluated using the area under the curve (AUC) for binary classification between None and 3-back conditions, the percentage of data classified as anomalous, and the correlation between the distance from the decision boundary and subjective MWL scores (NASA-TLX Adaptive Weighted Workload scores). Results indicated that the 30-second time window yielded significantly lower AUC values compared to longer windows, demonstrating inferior discrimination performance. The proportion of data classified as anomalous increased monotonically with MWL difficulty across all windows, but longer windows (120s and 150s) resulted in nearly 100% anomaly detection for lower-load conditions (1-back and 2-back), reducing sensitivity to gradations in workload. The correlation coefficient between subjective MWL and the OCSVM decision boundary distance increased monotonically from 30s to 120s, peaking at 120s, before decreasing at 150s. This decline at 150s suggests that excessive window length obscures variations in workload intensity. The study concludes that a time window of 60 to 120 seconds is appropriate for MWL estimation using eye and head movement data. This range balances the need for sufficient data to reduce noise and improve classification accuracy with the requirement for adequate temporal resolution to distinguish between varying levels of cognitive load. These findings provide specific guidelines for designing real-time driver monitoring systems that utilize machine learning-based anomaly detection.

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 qwen3.6-27b-nvidia on 2026-08-10; verification: verified.

Topics

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

• workload measurement

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: behavioral performance data, physiological data
• Methodological Resource: tool software