Classifying mental stress from eye tracking data: deep learning approaches for out-of-the-lab conditions

Clinical Snapshot

Key Findings

• P Value: Not reported

• Effect Size: Performance reported as reaching 'up to [Formula: see text] macro-averaged F1-score' under favourable conditions (goalkeeper dataset); specific numerical value not recoverable from abstract due to rendering artefact

• Primary Outcome: Binary or multi-class mental stress classification from unimodal eye-tracking time-series data using deep learning, evaluated by macro-averaged F1-score

• Nnt Or Sensitivity: Sensitivity, specificity, and AUC not reported in abstract; macro-averaged F1-score is the sole reported performance metric, with acknowledged substantial variation across datasets

• Confidence Interval: Not reported

Clinical Application

Eye-tracking hardware is increasingly available in consumer VR headsets (e.g., Meta Quest Pro, Apple Vision Pro), making the approach technically feasible at scale. However, clinical deployment requires: validated stress ground truth, regulatory-grade device accuracy, robust performance across diverse populations, and integration with clinical workflows. Current evidence is insufficient to support clinical implementation. No TGA-registered eye-tracking stress detection device exists in Australia at the time of appraisal. The RACGP does not currently endorse digital biomarker-based stress screening tools for primary care. PBS listing for any such technology is not applicable at this stage. Australian occupational health regulators (Safe Work Australia) may have future interest in passive stress monitoring technologies, but evidence thresholds for workplace deployment have not been met by this study. Research ethics frameworks under NHMRC guidelines would require substantially more robust validation before clinical translation. Potentially applicable to monitoring mental stress in occupational health, sports performance, or human-computer interaction contexts where wearable eye-tracking is feasible; not yet validated for clinical psychiatric or medical populations

Abstract

Eye-tracking signals such as pupil diameter and gaze behavior have been widely used for stress detection, yet most approaches rely on task-specific features, controlled laboratory settings, or multimodal sensor combinations, limiting scalability in less controlled environments. This work investigates whether unimodal eye-tracking time-series data can support task-agnostic stress detection beyond static laboratory tasks. We analyze stress classification across two complementary datasets: a virtual reality goalkeeper task with moderate visuomotor activity and stable recording conditions, and a virtual job interview dataset reflecting less controlled settings with uncalibrated signals. The results show that these signals alone contain informative patterns related to stress-associated autonomic and oculomotor responses. Under favorable conditions, performance reaches up to [Formula: see text] macro-averaged F1-score. At the same time, performance varies substantially across datasets, indicating that effective learning depends strongly on data quality, calibration, signal characteristics, and task design. Overall, the findings demonstrate the potential of unimodal eye tracking as a lower-burden alternative to more complex multimodal systems, while highlighting that reliable stress detection is fundamentally conditioned by the interplay of data, signal representation, and modeling approach.

References

1. 1.Laut, M., Dorschky, E., Richer, R., Rohleder, N., & Eskofier, B. M. (2026). Classifying mental stress from eye tracking data: deep learning approaches for out-of-the-lab conditions. Scientific Reports. https://doi.org/10.1038/d41586-021-00868-5 [Note: DOI as supplied by source metadata; appraisers note this DOI string does not resolve to the indexed article and should be verified against the publisher record prior to citation]