Long-Term Intracortical Cursor BCI (Card et al., 2026)
Measured by Card, Singer-Clark, Peracha, Stavisky, Brandman et al. · Nature Medicine (2026)
Inputs
The measured or assumed values behind the calculations, each with its source.
- B = 2.90 bits/s
- Mean achieved grid-task bitrate (2.90 +/- 0.16 bits/s; peak 3.16) for the neural cursor on the main 14x14 grid (0.77-inch tiles; Methods). It is the field-standard achieved-bitrate metric (log2(N-1) per cued selection). The authors compare it to arrays in hand motor cortex, while T15's arrays were in ventral precentral (speech) cortex. This is not a Fitts throughput. A separate larger-tile 6x6 condition gave 1.67 bits/s with the neural cursor.
- method = log2(N-1)/selection, net of errors
- Standard achieved-bitrate for 2D target acquisition, reported directly by the authors and explicitly benchmarked against prior intracortical cursor work. This paper uses log2(N−1) per net-correct selection (grid-task refs 9,18,34-37, in the Nuyujukian et al. 2015 lineage); the 2015 source paper's Eq. 1 uses log2(N).
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.
-
Grid geometry -> movement difficulty
14x14 grid, 0.77-inch tiles: W = S/14; random targets -> mean amplitude A ~= 0.52*S, so A/W ~= 7.28. ID = log2(A/W + 1) = log2(8.28) ~= 3.05 bits/movement.
S = board span; 0.52 is the mean distance between two random points on a square. The target is cued, so each correct selection is a pointing movement carrying the Fitts index of difficulty (~3.05 bits), not the log2(195) ~= 7.61 bits the achieved-bitrate metric credits. The discount ratio ID/log2(N-1) is 0.401 here, 0.398 at 6x6 and 0.414 at 30x30, so it is nearly grid-size independent.
-
Re-credit the reported achieved bitrate onto the Fitts basis
ITR = B * ID / log2(N-1) = 2.90 * 3.05 / 7.61 = 1.17 bits/s
The paper reports only the achieved bitrate B = 2.90 bits/s and no independent grid selection timing, so the Fitts figure is that same measured throughput re-credited per movement at the Fitts ID instead of log2(N-1) -- a unit re-crediting of the reported score, not an independent measurement. Same correction applied to Pandarinath's ReFIT grid and Neuralink's Webgrid, placing all three intracortical cursors on the mouse's Fitts basis (mouse 4.5, stylus 4.9, trackball 3.3 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
- Continuous control space
- Size
- Continuous
- Prior
- Uniform: all actions assumed equally likely
- Notes
- The same implanted arrays used for long-term speech decoding also drove a 2D neural cursor in a structured grid-task benchmark. This entry is the cursor-control channel only; the speech/text-output application is split into the companion Card 2026 speech entry. The authors' headline uses the field-standard achieved-bitrate (log2(N−1) per cued selection), while the atlas pointing convention uses a Fitts-style movement difficulty on the same grid, so the reference number here is a Fitts throughput for comparability with the mouse and the other intracortical cursors (Pandarinath grid, Neuralink Webgrid); the as-reported achieved bitrate is kept as a secondary figure.
Comparability The strictest bound here is Fitts throughput: the index of difficulty, log₂(A/W + 1), per movement. Directly comparable to the other continuous-pointing entries (mouse, trackball, stylus, gaze and the cursor BCIs). Set against the text entries (keyboards, spellers, speech) it crosses methods: both report bits/s, but one measures movement difficulty and the other text information, so compare within the family first.
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 metric
14x14 grid -> N = 196 targets; this paper credits log2(N-1) = log2(195) ~= 7.61 bits per net-correct cued selection (Card et al. 2026, grid-task refs 9,18,34-37, in the Nuyujukian et al. 2015 lineage; the 2015 original uses log2(N)).
-
Authors' reported score (taken as reported, not re-derived)
B = 2.90 +/- 0.16 bits/s on the 14x14 grid (0.77 in tiles); peak 3.16.
Reported in the cursor results (on the same task gaze control reached 0.80 bits/s; a coarser 6x6 grid gave 1.67 bits/s neural and 2.59 bits/s gaze). The paper gives only the bitrate, no separate grid selections/s, so there is no more-primary quantity to derive it from: it is author-reported. Dividing by the 7.61-bit credit implies ~0.381 selections/s, but that rate is a consequence of the score, not an independent measurement.
Source
- Authors
- Card, Singer-Clark, Peracha, Stavisky, Brandman et al.
- Publication
- Nature Medicine, 2026
- Reference
- Nature Medicine full text
- Reference
- Open-access preprint (bioRxiv)