Reading boredom in the eyes

People are bad at noticing the moment they lose interest, and worse at reporting it accurately afterward. If eye behavior shifts before a person decides they're bored, it could serve as a continuous, unprompted measure of disengagement. That matters for any team that needs to know when attention slips: in video, advertising, training or interface design.

Each participant watched three 17-minute films in a VR headset: one chosen to bore, one chosen to engage, and the clinical procedure footage that the course's virtual learning environment exists to deliver. Every person saw every film, so each participant served as their own baseline.

In a five-minute break after each film, a short survey collected self-reported boredom and engagement. That gave two kinds of evidence on purpose: the eyes record what happens during viewing, before any judgment, and the survey collects the verdict afterward.

Signals recorded on every session:

• Eye tracking: gaze direction and eye closure
• Heart rate
• EEG
• Headset position and rotation
• Observation video of the participant
• Self-report surveys at each break

The study ran in two rounds under UC Irvine IRB exemption: eight participants first, then four more recorded at the eye tracker's full sampling rate.

The devices ran on separate clocks, so before any analysis I built a pipeline to put everything on one timeline.

• Detected eye-closure events and used them to synchronize the eye tracker with video, then cropped each session to its viewing window.
• Reverse-engineered the projection from 3D gaze vectors to video pixels by fitting it against an earlier overlay, with a median error of 7.7 pixels.
• Rendered 4K, 60 fps gaze-overlay videos and attention heatmaps for every session in the second round.
• Cut matching segments from the observation camera by cross-correlating its audio with the film soundtrack.

With a small sample and repeated measures, standard p-values overstate confidence. I tested each association within participants using Monte Carlo permutation tests, which hold up when the sample is small. A signal counted only if it reached significance. I didn't claim an effect for any channel that only moved in the expected direction.

The EEG showed no reliable effect. Two signals that a headset already records, where the eyes point and how much they close, were enough to follow boredom without asking anyone.

Headsets are expensive, so we also tested whether an ordinary laptop webcam could follow the eyes well enough. Four participants tracked a moving target on screen with their own webcams. Webcam gaze matched the target's position with an average cross-correlation of 0.87, at a lag of 0.03 seconds, and held between 0.80 and 0.91 across distance, glasses and lighting. That result is where FocusTracker started.

Eye behavior offers a disengagement signal that needs no interruption and no survey. For a product team, that means attention can be measured while content plays, compared across versions, and potentially used to adapt what is shown next.

I made the case for that application in a pitch on digital health:

• King, C. E., M. Lo, M. Das, and D. Salvo. "Assessment of Student Engagement in Virtual Reality Clinical Immersion Environments through Eye Tracking." Proceedings of the 2024 ASEE Pacific Southwest Conference, 2024.
• King, C. E., K. R. Feeney, Q. Tang, M. Das, and D. Salvo. "Work in Progress: Physiological Assessment of Learning in a Virtual Reality Clinical Immersion Environment." Proceedings of the 2023 ASEE Annual Conference, 2023.Best BME Division Work in Progress Award

The full analysis is in my dissertation, On Hacking the Soul (UC Irvine, 2026), currently under embargo.