Abstract

Eye-gaze typing systems that use dwell time as the selection mechanism - holding gaze on a key until it activates - have served motor-impaired users for two decades, but the standard calibration model assumes a stable user. For people with ALS, multiple sclerosis, or Parkinson’s disease, both gaze accuracy and voluntary dwell control degrade over months. Most commercial systems including Tobii Dynavox and Eyegaze Edge require a clinician-assisted recalibration session to recover performance after decline, creating a care dependency that interrupts daily communication. The question is whether continuous background adaptation can flatten that curve without introducing false activations.

The Calibration Drift Problem

A nine-point calibration grid maps the camera’s gaze vector to screen coordinates. The mapping degrades when the user’s head position shifts habitually - common as postural support changes with disease progression - or when pupil dilation varies due to medication. Residual error at the periphery of the calibration grid is higher than at the center, which matters because on-screen keyboards typically place infrequently used characters at the edges. Studies from the Augmentative and Alternative Communication lab at Penn State have documented accuracy drops of 1.5 to 2.5 degrees over four-week periods in ALS users without recalibration, enough to make adjacent keys unreliable.

Adaptive Dwell Thresholds

The dwell duration - typically set between 400 and 900 milliseconds - interacts with fatigue. Early in a session, shorter dwell times are tolerable. After twenty minutes of continuous use, involuntary fixations on non-target keys increase, and the false activation rate rises. A straightforward adaptive model tracks the ratio of activated keys to intended words using an N-gram language model as a proxy for intent. When the mismatch rises above a threshold, the system incrementally lengthens the dwell window and widens the dead zone around recently activated keys. This costs typing speed but recovers accuracy, and the tradeoff is user-configurable. The open-source ACAT framework (Assistive Context-Aware Toolkit), originally developed at Intel, provides a plugin architecture where such policies can be swapped without modifying the core input loop.

Continuous Background Recalibration

Rather than a discrete nine-point session, background recalibration accumulates implicit corrections during normal typing. When a user gazes at a key and then immediately triggers a backspace, the system infers a probable selection error and adjusts the local mapping in the vicinity of that key. This is a noisy signal - not every backspace means gaze drift - so updates are weighted by confidence and smoothed over a sliding window. The risk is drift amplification: if the initial mapping error causes corrections that pull the mapping in the wrong direction, subsequent typing degrades faster. Regularization toward the last explicit calibration prevents runaway drift, capping the background adaptation’s influence.

Integration with Prediction Engines

Word prediction reduces the number of individual key activations required, which matters more as gaze accuracy declines. A 10-word prediction bar that surfaces likely next words from a personalized language model lets users select full words with a single dwell rather than spelling letter by letter. The prediction vocabulary should update continuously from the user’s own output, not only from a static corpus, to reflect idiolect and domain-specific terms. WordQ and Communicator 6 both support personal vocabulary growth, though neither currently integrates with gaze calibration state to modulate prediction aggressiveness based on current accuracy estimates - an open design gap.