EveLoad: Cognitive Workload Recognition from Event-Based Eye Movements

As rehabilitation and assistive interfaces increasingly rely on adaptive, closed-loop systems, there is a critical need for unobtrusive methods to monitor a user's cognitive state in real time. Traditional frame-based eye trackers often struggle with motion blur and limited temporal resolution during rapid eye movements. This paper investigates whether event cameras—which capture asynchronous, microsecond-level brightness changes—can provide a more robust and effective sensing modality for recognizing cognitive workload levels.

The authors created the EveLoad dataset, which consists of 7.5 hours of synchronized event-based eye-movement recordings from 20 participants. To ensure the model learns workload-related ocular dynamics rather than simple gaze-location patterns, the researchers employed a spatially constrained N-back-guided fixation paradigm. Participants performed tasks across six levels of difficulty, defined by varying memory demands (no-load, 0-back, 1-back) and stimulus presentation speeds (750 ms vs. 1500 ms). The team then developed a learning framework using an adapted ResNet18 backbone that processes stacked event frames to classify these six workload levels.

The experimental results demonstrate that the proposed event-based framework is highly effective, achieving an average subject-specific accuracy of 96.36% and a mixed-subject random split accuracy of 96.13%. These findings suggest that event cameras successfully capture fine-grained ocular dynamics—such as saccades and microsaccades—that correlate with cognitive effort, providing a promising, low-latency alternative to traditional frame-based eye tracking for future adaptive rehabilitation technologies.