Improved Inference and Autotyping in EEG-based BCI Typing Systems
Andrew Fowler, Brian Roark, Umut Orhan, Deniz Erdogmus, Melanie Fried-Oken · 2013 · Proceedings of the 15th International ACM SIGACCESS Conference on Computers and Accessibility (Assets '13) · doi:10.1145/2513383.2513453
Summary
This paper presents improvements to the RSVP Keyboard, a brain-computer interface (BCI) typing system designed for people with severe motor disabilities, particularly those with locked-in syndrome (LIS) resulting from ALS or brain stem stroke who cannot use any volitional switch. The system works by presenting a rapid serial sequence of 28 symbols (26 letters, space, backspace) on screen at 2.5 Hz. When the user sees their target letter, their brain produces an involuntary P300 event-related potential approximately 300ms after stimulus onset, which is detected from EEG signals recorded by a 10-20 electrode cap. The system combines EEG classifier scores with a 6-gram character-level language model via Bayesian fusion to determine posterior probabilities for each symbol. When a symbol exceeds a decision threshold (currently 0.9), it is typed. The key limitation addressed is that the baseline system discards all EEG observations when moving to a new character position, losing valuable information about alternative letter sequences. The authors propose maintaining posterior probabilities across all possible context strings — including paths through the backspace symbol — throughout the typing session. This allows the system to efficiently recover from errors and provides a principled estimate of backspace probability (previously fixed at 0.05) based on the accumulated evidence for alternative strings.
Key findings
The improved inference algorithm achieved a 20% increase in simulated typing speed over the best current system configuration across multiple AUC levels (representing different EEG classifier accuracies). At AUC=0.90, the system improved from 1.84 to 1.48 sequences per letter (20% improvement), and results generalized across three different text corpora: personal emails from a person with LIS (20% improvement), New York Times newswire (19%), and a simulated AAC corpus (16%). All improvements were statistically significant at p < 0.001. The method also enabled autotyping — automatically typing high-probability letters without requiring EEG observations — which provided additional speed gains, particularly at higher AUC levels where 41% of symbols could be autotyped and 18% of autotyping events completed entire words. An emergent behavior called "autorevision" was observed: sequences of autotype and autodelete actions could produce complex corrections from a single ERP detection. The optimal decision threshold dropped from 0.9 to 0.5 with the improved algorithm, meaning it was more efficient to type quickly and correct errors than to wait for higher certainty. The backspace tradeoff analysis showed users could halve backspace frequency with only slight typing speed reduction by adjusting the decision threshold.
Relevance
This paper addresses one of the most challenging accessibility scenarios: providing communication capability to people who have lost all voluntary motor control. For individuals with locked-in syndrome, a BCI typing system may be their only means of composing text. The 20% speed improvement is significant in a domain where typing rates are measured in letters per minute rather than words per minute — at AUC=0.90, the improved system achieves about 4.5 letters per minute versus 3.6 for the baseline. The work demonstrates the critical role of language models and intelligent inference in assistive technology, showing that software improvements can substantially benefit users without any hardware changes. The principled treatment of backspace probability is particularly important because error correction is a major bottleneck in BCI typing — the baseline system's fixed backspace probability was too low to recover from errors at lower AUC levels. For AAC practitioners, this research underscores that text prediction and language modeling can dramatically improve communication speed for users of all assistive input methods, not just BCI systems.
Tags: brain-computer interface · BCI · EEG · text entry · locked-in syndrome · ALS · AAC · language models · P300 · assistive technology · motor disability