Cognitive Distraction Detection Using Gaze and Pupil with an Interpretable Approach

Tamura, Kimimasa; Stent, Simon; Gideon, John; Shintani, Kohei; Rosman, Guy · 2025 · IEEE Intelligent Vehicles Symposium

DOI: 10.1109/iv64158.2025.11097826

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 addresses the challenge of detecting cognitive distraction (CD) in drivers, a significant cause of traffic accidents that is difficult to identify due to its subtle manifestations. While prior research often relied on basic statistical measures of gaze and pupil data, this work hypothesizes that capturing complex, nonlinear patterns and interactions between multiple modalities is essential for improved detection. The authors aim to determine whether interpretable machine learning models can match the performance of deep neural networks (DNNs) while offering better explainability for in-vehicle safety systems. The researchers utilized data from 52 participants in a driving simulator, who performed two types of cognitive distraction tasks: a verbal n-back task and a statement task. Data collection involved a Tobii Pro Spark eye tracker to record gaze coordinates and pupil diameter, alongside steering wheel angle inputs. The study employed extensive feature engineering, extracting raw time-series signals and derived metrics such as fixation-saccade ratio, gaze entropy, and gaze depth. These signals were processed using comprehensive time-series feature extraction libraries (tsfresh) and SARIMAX models to generate high-dimensional feature vectors. The experimental design included a between-subjects split to ensure generalization, with models trained on one task subset and tested on both the same and different tasks to evaluate cross-task generalization. The results demonstrate that tree-based ensemble methods, particularly CatBoost, achieved detection performance comparable to or exceeding that of DNNs like Transformers and TabNet, with ROC AUC scores reaching 0.931 for the n-back task. Linear models performed significantly worse. Feature analysis revealed that combining gaze, pupil, and derived physiological signals boosted detection accuracy, with vertical gaze movements, baseline pupil size, and minimum gaze distance identified as key indicators of distraction. Furthermore, models trained on the n-back task generalized effectively to the statement task, indicating that the models learned general features of cognitive load rather than task-specific artifacts. The significance of this work lies in validating that interpretable, tree-based models are viable alternatives to "black-box" DNNs for cognitive distraction detection, offering high accuracy without sacrificing explainability. The findings suggest that integrating multiple modalities and sophisticated feature engineering is crucial for capturing the subtle signs of CD. By releasing their code and preprocessed data, the authors provide a foundation for future research, emphasizing that functional safety and interpretability can be maintained alongside high detection performance in driver monitoring systems.

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. Discovered via author_sweep_intake on 2026-05-28.

StageOutcomeToolModelPromptAttemptsCompleted
discover success 2 2026-05-28
archive success canonical_url 7 2026-06-09
extract success cached 2 2026-06-10
clean success clean 1 2026-06-04
chunk success chunk 1 2026-06-04
embed success embed Qwen/Qwen3-Embedding-8B 1 2026-06-04
enrich success 1 2026-05-28
promote success 1 2026-06-04
summarize success llm qwen3.6-27b-prismaquant summ-v5 1 2026-06-10
tag success vector_similarity 15 2026-06-11
verify success 1 2026-06-10

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

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

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).