In the race to restore communication to patients with complete locked-in syndrome (CLIS), neuroengineers have encountered a profound paradox: the simplest 19th-century binary protocol is outperforming multi-million dollar neural network decoders.
The Multi-Class Classification Bottleneck in EEG
Attempting to decode 26 distinct imagined letters or complex visual P300 spelling grids from non-invasive EEG caps requires high signal-to-noise ratios that degrade over hours of patient fatigue. Classification accuracy often hovers between 60% and 75%, requiring frustrating backspaces and slow typing speeds below 3 words per minute.
“When we reduced the neural classification space from 26 letter states down to just two orthogonal motor intents—dit and dah—our classification accuracy jumped from 68% to over 98.4%.”
Orthogonal Motor Imagery for Dit and Dah
In modern BCI protocols, patients are trained to imagine two distinct, isolated physical actions. For example, imagining a rapid right hand squeeze produces a desynchronization in the left motor cortex (mu rhythm) calibrated as a Dit. Imagining a prolonged pedal press produces central sensorimotor rhythm shifts calibrated as a Dah.
Letter-Grid Scanning: 3.2 WPM, 74% accuracy | 2-State Binary Morse BCI: 14.8 WPM, 98.4% accuracy | Information Transfer Rate (ITR): +310% gain.
Combining Binary Morse with Generative Language Models
By coupling low-error binary Morse streams with modern LLMs and character-level language models, patients only need to key the first 2 or 3 characters of a word before auto-complete suggestions predict the complete phrase, allowing natural conversational cadence.