TRCA SSVEP Speller (Nakanishi et al., 2018)
Measured by Nakanishi, Wang, Chen, Wei, Chuang, Jung & Wang · IEEE Trans. Biomed. Eng. 65(1) (2018)
Inputs
The measured or assumed values behind the calculations, each with its source.
- N = 40
- 40-target frequency-phase-coded speller (same stimulus set as Chen et al. 2015); the advance is the task-related component analysis (TRCA) decoder, not the target count.
- rate = 75 char/min
- Cue-guided spelling rate: 0.3 s stimulation + 0.5 s gaze shift = 0.8 s/selection → ~75 selections/min.
- P = 0.8983
- Online cue-guided selection accuracy at the 0.3 s data length.
- H = 1.0 bits/char
- English-text entropy (Shannon); the same ~1 bit/char standard applied to QWERTY and every other text entry.
- ITR_reported = 325.33 bits/s
- Authors' online cue-guided Wolpaw ITR (free-spelling 3.31 bits/s): the highest ITR reported for an EEG BCI at the time. Counts log2(N) per selection.
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.
-
Correct characters per minute
~75 selections/min (0.3 s flicker + 0.5 s gaze shift) × 89.83% ≈ 67 correct char/min
Each selection emits one character; this is the rate of correct English text produced in the cue-guided task (free-spelling is lower).
-
Bits per character
H(English) ≈ 1.0 bit/char (Shannon)
-
Information transfer rate
67 char/min × 1.0 bit/char ÷ 60 s/min = 1.12 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
- 40 distinguishable actions
- Prior
- Context-conditioned: likelihoods depend on prior actions
- Notes
- Same 40-target SSVEP paradigm as Chen et al. 2015; the decoder classifies which flicker frequency the user gazes at (a classifier, not Fitts pointing). The realized output is English text, so the reference uses character-entropy (~1 bit/char) like every other text entry; the authors' 5.42 bits/s Wolpaw figure counts log2(N) per selection and is kept as a secondary classifier metric. The faster decoder (0.8 s/selection vs ~1 s in Chen 2015) is why the atlas text-throughput estimate edges above Chen's.
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 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.
-
Bits per selection (Wolpaw, N = 40 at 89.83%)
B = log2(40) + 0.8983·log2(0.8983) + 0.1017·log2(0.1017/39) ≈ 4.31 bits/selection
Online cue-guided accuracy 89.83%. This counts log2(N) per selection, the classifier metric, not the 1 bit/char the atlas-ranked text figure holds every text entry to.
-
Selections per second
0.3 s flicker + 0.5 s gaze shift = 0.8 s/selection → 1.25 selections/s
-
Information transfer rate
ITR = 4.31 × 1.25 ≈ 5.39 bits/s
Reproduces the authors' reported online cue-guided ITR (325.33 bit/min ≈ 5.42 bits/s, the small gap being rounding). This was the record EEG-BCI ITR at publication.
-
Achieved-bitrate credit per net-correct selection
N = 40 → log2(N − 1) = log2(39) = 5.29 bits per net-correct selection (field-standard achieved bitrate, e.g. Webgrid; Nuyujukian 2015, which introduced the metric, used log2(N)).
-
Net-correct selection rate
net-correct = 2P − 1 = 2(0.8983) − 1 = 0.797 of selections. At 1.25 selections/s (0.3 s flicker + 0.5 s gaze shift) → 0.797 × 1.25 = 0.996 correct/s.
A speller error commits the wrong character rather than timing out, so incorrect = 1 − P. Same N (40), online cue-guided accuracy (89.83%) and selection rate (1.25/s) as the entry's Wolpaw calc. Netting each wrong selection against a correct one (2P − 1) drops the achieved figure a little below the 5.39 bits/s Wolpaw ITR; the faster TRCA decoder is why it edges above Chen 2015.
-
Achieved bitrate
5.29 bits × 0.996 correct/s = 5.26 bits/s.
Source
- Authors
- Nakanishi, Wang, Chen, Wei, Chuang, Jung & Wang
- Publication
- IEEE Trans. Biomed. Eng. 65(1), 2018