Experimental study demonstrates high-accuracy speller decoding via pupillary light reflex in healthy adults, suggesting viable calibration-free communication interfaces.
Key Points
To develop and validate a calibration-free, 40-class speller system that decodes user intent from pupil size fluctuations induced by flickering visual stimuli using deep learning.
Enrolled 12 healthy participants to test time-division flickering visual stimuli across three inter-stimulus intervals (0.2, 0.3, and 0.4 seconds).
Applied a dynamic visibility-based feature extraction method paired with a temporal-spatial hybrid deep learning network.
Evaluated performance using nested cross-validation, leave-one-subject-out cross-validation, and real-time online testing.
The 0.3-second inter-stimulus interval achieved optimal performance, with nested cross-validation reaching 94.14 ± 1.15% accuracy and an information transfer rate of 67.32 ± 0.66 bits/min.
Leave-one-subject-out cross-validation yielded an accuracy of 90.58 ± 1.36% and an information transfer rate of 62.80 ± 1.73 bits/min.
Real-time online evaluation demonstrated an accuracy of 93.06 ± 0.99% and an information transfer rate of 55.12 ± 1.04 bits/min without requiring user-specific gaze calibration.