BrainGate2 Cursor BCI (ReFIT-KF), text entry (Pandarinath et al., 2017)
Measured by Pandarinath, Nuyujukian, Henderson, Shenoy et al. · eLife 6:e18554 (2017)
Inputs
The measured or assumed values behind the calculations, each with its source.
- rate = 39.2 correct char/min
- Best sustained copy-typing rate: participant T5 on the OPTI-II onscreen keyboard, NO word-completion or prediction (T5 QWERTY 36.1; T6 OPTI-II 31.6; T7 ABCDEF 13.5). ≈7.8 wpm. This copy-typing of English sentences is the paper's real-world communication task.
- H = 1.0 bits/char
- English-text entropy (Shannon); participants copy-typed English sentences, so the same ~1 bit/char standard used for QWERTY, eye-typing and Morse applies.
- N = 28
- Keys on the OPTI-II onscreen keyboard, for the raw-key Wolpaw ceiling (uniform prior over the alphabet). This bounds the keyboard selection itself; the underlying cursor channel is the separate BrainGate2 grid entry.
- T_key = 1.531 s/key
- Key-selection interval for the Wolpaw ceiling: 60 / 39.2 char/min = 1.531 s. Accuracy is not reported separately (the 39.2 is correct char/min), so the bound is taken at perfect copy (P=1) as a strict ceiling.
Strictest ITR
Each scoring method is an upper bound on the channel, so the headline is the strictest (smallest) one for this entry. Use the score selector on the home page to view any single method across entries.
-
Characters per minute
39.2 correct char/min (T5, OPTI-II keyboard, no word prediction) ≈ 7.8 wpm
Copy-typing of English sentences in 2-minute blocks: the rate the user actually communicated. Word-completion was deliberately disabled, so this is the raw BCI typing rate; predictive text would raise it.
-
Bits per character
H(English) ≈ 1.0 bit/char (Shannon)
-
Information transfer rate
39.2 char/min × 1.0 bit/char ÷ 60 s/min = 0.65 bits/s
What counts as a bit depends on the action space. The number of distinguishable actions and how likely each one is are design choices of the task, not the sensing hardware. The same modality can present a fixed set of targets, a set pruned per step by a grammar or language model, or a continuous control space. Each of these changes how many actions are live and how the probability mass is spread, and therefore the information per selection. Read the action space below before comparing headline numbers across entries.
Action space
What the user can produce at each step, and how those options are distributed.
- Structure
- Fixed set of targets
- Size
- 28 distinguishable actions
- Prior
- Context-conditioned: likelihoods depend on prior actions
- Notes
- A cursor-driven onscreen keyboard (OPTI-II / QWERTY) from the BrainGate2 pilot clinical trial, with a ReFIT Kalman-filter decoder: the intracortical cursor selects keys one at a time to spell English text. The action set is the ~28 keys, but real English is non-uniform, so the realized information is the character-entropy of the text (~1 bit/char), not log2(keys). The underlying continuous-cursor channel, along with the separate 6×6 grid benchmark that measures its peak bitrate, is the companion entry (BrainGate2 Cursor BCI (ReFIT-KF), pointing). This entry is what the participant actually communicated; that one is the channel benchmark.
Comparability The strictest bound here is the Shannon entropy of the output text, under one predictor held constant across the whole atlas (≈1 bit per character). That shared predictor makes it directly comparable to every other text entry (keyboards, spellers, silent speech and speech BCIs) regardless of prior or vocabulary size. For most text interfaces it comes out tighter than the raw-selection bounds, but not always. Where a small vocabulary makes Wolpaw tighter, that wins instead. Any Fitts, Wolpaw or log₂(N) figure shown below is another bound on the same channel. Switch the home-page score selector to compare one across entries.
Other bounds considered for the headline
Also valid upper bounds for this entry and eligible to be the headline. They just came out looser than the strictest above. Pick any of these in the home-page score selector.
-
Bits per selection (Wolpaw formula)
B = log2(N) + P*log2(P) + (1-P)*log2((1-P)/(N-1)) = log2(28) + 1*log2(1) + 0*log2(0/27) = 4.807 bits / selection
Term 1 is the information if every choice were correct; terms 2-3 subtract the bits lost to the error rate, assumed spread evenly over the other N-1 targets.
-
Selections per second
T = 1.531 s/selection -> 1 / 1.531 = 0.653 selections/s
-
Information transfer rate
ITR = B * selections/s = 4.807 * 0.653 = 3.14 bits/s
Other score types
Bounds the atlas keeps out of the default strictest headline: as-reported figures, alternate task conditions, or raw-channel ceilings that shouldn't win the headline by default. Each still carries a score type, so the home-page selector ranks this entry on it when you choose that type. Read its derivation before comparing across entries.
-
Achieved-bitrate credit per net-correct key
N = 28 keys → log2(N − 1) = log2(27) = 4.75 bits per net-correct selection (field-standard achieved bitrate, e.g. Webgrid; Nuyujukian 2015, which introduced the metric, used log2(N)).
-
Net-correct key rate
Accuracy is not reported separately (39.2 is already correct char/min), so this is a perfect-copy ceiling: net-correct = 39.2 char/min = 0.653 correct/s (one key per 1.531 s).
Same N (28 keys) and key interval (1.531 s) as the entry's perfect-copy Wolpaw ceiling. With no error term the achieved and Wolpaw ceilings differ only by log2(N − 1) vs log2(N), so both land near 3.1 bits/s: the uniform-prior key channel, above the 0.65 bits/s Shannon headline. The underlying continuous-cursor channel is the companion BrainGate2 grid entry.
-
Achieved bitrate
4.75 bits × 0.653 correct/s = 3.11 bits/s.
Source
- Authors
- Pandarinath, Nuyujukian, Henderson, Shenoy et al.
- Publication
- eLife 6:e18554, 2017
- Paper
- 10.7554/eLife.18554
- Reference
- Open-access full text (eLife)