Abstract
Gaze-based keyboards offer a flexible human-computer interaction (HCI) method for both disabled and able-bodied individuals. However, these systems are often error-prone due to eye-tracking misinterpretations and calibration drifts, particularly in "fast mode." This paper presents a novel bimodal error detection system (EDS) that aggregates decisions from two distinct data streams: gaze-related measurements and electroencephalography (EEG). By detecting Error-Related Potentials (ErrPs) and specific eye-transition patterns, the system can automatically flag and correct typos. Experimental results from 10 participants demonstrate that this hybrid approach significantly reduces typesetting time and improves overall system utility.
Introduction
Gaze-based keyboards are essential communication tools for individuals with severe motor disabilities, such as those with ALS. Despite their utility, they suffer from the "Midas touch" problem—where accidental selections occur due to lingering gaze—and calibration inaccuracies. In high-speed typing scenarios, the traditional "detect-react-correct" cycle, where a user must manually navigate to a backspace key after realizing a mistake, becomes a major bottleneck. This research aims to automate error correction by identifying the user's internal realization of an error the moment it happens.
The Bimodal Error Detection System
The proposed system utilizes a hybrid Brain-Computer Interface (hBCI) architecture to process two independent signals simultaneously.
1. Neural Signal Processing (EEG)
The system monitors for Error-Related Potentials (ErrPs), which are neurophysiological signatures appearing 300–400ms after a subject perceives an error. To extract these signals from the noisy EEG data (often contaminated by eye-movement artifacts), the researchers employed Discriminant Spatial Patterns (DSP) for spatial filtering. This ensures that the ErrP is isolated as a distinct event rather than a byproduct of gaze shifts.
2. Gaze Trajectory Analysis
Beyond simple coordinates, the system analyzes the dynamics of eye movements using Hjorth Descriptors (Activity, Mobility, and Complexity). These parameters describe the speed and acceleration of the gaze. When a selection is correct, the eyes move predictably toward the next target; when an error occurs, the gaze often lingers or exhibits "micro-adjustments" as the user reacts to the mistake.
3. Late Fusion Classification
Separate Support Vector Machine (SVM) classifiers were trained for the EEG and gaze streams. The system uses a "late fusion" approach, combining the confidence scores of both models to make a final determination. This method leverages the high specificity of EEG (reducing false deletions) and the high sensitivity of gaze patterns (catching more typos).
Experimental Methodology
The study involved 10 participants who typed 20 pre-memorized sentences. The hardware setup included a 64-channel EEG and a 30Hz eye-tracker.
To address the class imbalance—since typos occur less frequently than correct entries—the researchers utilized the Synthetic Minority Oversampling Technique (SMOTE) to generate synthetic error samples for training. A fixed dwell time of 0.5 seconds was used to simulate a "fast mode" typing environment.
Key Results and Discussion
The hybrid system demonstrated a clear advantage over traditional gaze-only keyboards:
- Efficiency Gains: The system achieved a 9.3% increase in typing speed, saving an average of 2.7 seconds per sentence.
- Classification Accuracy: The late fusion model reached a sensitivity of approximately 86% and a specificity of 83%.
- Utility Metric: The authors prioritized a "Utility metric" that penalizes false positives (deleting a correct letter) more heavily than false negatives, as the former is more disruptive to the user experience.
Crucially, the study confirmed that ErrPs can be robustly detected even during active eye-tracking, debunking concerns that eye artifacts would render the neural signal unusable. Users who were naturally more error-prone saw the greatest benefit, suggesting the system is highly effective for those who struggle with eye-tracker calibration.
Conclusion and Future Work
While the results are promising, the study's small sample size of able-bodied participants necessitates further validation with the target demographic of individuals with severe motor impairments. Future research should explore the integration of linguistic auto-correct features and the transition from research-grade 64-channel EEG to consumer-grade wearables to make this technology accessible for daily use.
Summary updated