Overview

This release contains pre-processed neural data used to evaluate long-term intracortical microelectrode array recording and movement-decoding performance in participants enrolled in the BrainGate and BrainGate2 pilot clinical trials. The dataset is intended to support reproduction and extension of the central analyses from the associated manuscript, including array yield, decoding signal-to-noise ratio (dSNR), tuning stability, and electrode-count scaling analyses.

The release is organized into two main datasets:

• yield (yield_<participant_id>.tar.gz): recordings used to assess array yield.
• decoding (decoding_<participant_id>.tar.gz): closed-loop cursor-control recordings used to assess movement-intention decoding, tuning stability, and electrode-count scaling.

The data are provided as .mat files packaged by participant as .tar.gz archives.

Associated publication and code

Associated manuscript:

Hahn, N. V., Stein, E., BrainGate Consortium, Donoghue, J. P., Simeral, J. D., Hochberg, L. R., & Willett, F. R. Performance of intracortical microelectrode arrays in people with implanted brain-computer interfaces over a 20-year period.

Dataset DOI:

https://doi.org/10.5061/dryad.x0k6djj1h

Associated analysis code:

https://github.com/nptl-stanford/array-paper

Data summary

The study summarizes neural recording and decoding performance across the first 14 participants enrolled in the BrainGate and BrainGate2 clinical trials. Each participant had one or two 96-channel platinum-metalized Utah microelectrode arrays placed in cortex. Nineteen of the twenty arrays were placed in the hand area of motor cortex, and one was placed in the middle frontal gyrus.

The release contains two curated datasets:

The decoding dataset is smaller because not every session included BCI cursor control, and because high-bandwidth neural data were not consistently available or alignable for all participants.

File organization

Dryad displays the release as a flat list of uploaded files:

To extract all participant archives:

for archive in *.tar.gz; do
  tar -xzf "$archive"
done

To extract a specific participant's decoding and yield files:

tar -xzf decoding_T10.tar.gz
tar -xzf yield_T10.tar.gz

Each archive expands into the dataset and participant directory. For example,

decoding/T10/
  T10_day_294_decoding.mat
  T10_day_<post_implant_day>_decoding.mat
  ...

yield/T10/
  T10_day_<post_implant_day>_yield.mat
  ...

The .mat files can be loaded in MATLAB:

decoding_example = load("decoding/T10/T10_day_294_decoding.mat")
yield_example = load("yield/T5/T5_day_2218_yield.mat")

or Python:

from scipy.io import loadmat
decoding_example = loadmat(
    "decoding/T10/T10_day_294_decoding.mat",
    simplify_cells=True,
)

yield_example = loadmat(
    "yield/T5/T5_day_2218_yield.mat",
    simplify_cells=True,
)

File naming convention

Session files follow this naming pattern:

<participant_id>_day_<post_implant_day>_<dataset_type>.mat

where:

Decoding dataset

The decoding dataset contains pre-processed neural recordings from sessions in which participants attempted to move a computer cursor toward defined targets. These sessions can be used to evaluate decoding quality.

The decoding files include aligned neural features, cursor positions, target positions, block metadata, and trial metadata. Task designs varied across sessions and participants and included target-directed cursor-control tasks such as Radial-8, grid tasks, Fitts tasks, and related variations.

For standardization, all neural time-series features in this release are provided in 10 ms bins.

Decoding file contents

Each decoding .mat file contains the following top-level variables.

electrodes struct

blocks struct

trials struct

The trials structure contains movement-attempt ("go") periods identified from the task data (changes in target positions and/or task-state signals).

neural struct in decoding files

The neural struct contains 10 ms binned neural features with shape n_time_bins x n_electrodes.

Threshold crossing fields (tx_*) are given in units of counts per bin and spike-band power (sbp) is given in µV²

Note that analyses in the associated manuscript apply additional block-wise normalization before fitting decoding or encoding models.

Yield dataset

The yield dataset contains one neural recording from each included research session. These files were used to estimate array yield (the number of electrodes detecting spiking activity in an array) over time.

Yield files include electrode metadata, including impedance when available, firing rates across threshold settings used for determining spiking electrodes, mean spike waveform snippets for detected spiking electrodes, and binned threshold crossing features. No associated behavioral data are included in this dataset, and derived threshold crossings for dual-array participants may not be time aligned between arrays.

Yield file contents

Each yield .mat file contains the following top-level variables.

firing_rates struct

The firing_rates struct is a long-format table with one row per electrode-threshold combination.

Firing rates were calculated by counting threshold crossings within each complete, non-overlapping 10-second window, converting each count to a rate in hertz, and taking the median rate across windows.

spike_waveforms struct

Waveform snippets were computed from preprocessed voltage traces. Preprocessing included decimation from 30 kHz to 15 kHz, band-pass filtering from 250–5000 Hz, linear regression referencing, and thresholding at −4.5 times the robust standard deviation of the voltage signal to identify threshold-crossing times. Mean waveforms were calculated by averaging waveform snippets within non-overlapping 10-second windows and then taking the sample-wise median across window-level means. Waveforms are reported in microvolts and are included for electrodes with a robust −4.5 threshold-crossing rate of at least 2 Hz.

This waveform estimator differs from the standard mean across all snippets used in the associated paper. The median of 10-second window means was found to be less sensitive to transient noise and was therefore used for the released waveform summaries. Consequently, released mean waveforms may differ slightly from the waveform summaries used in the paper.

neural struct in yield files

The yield neural structure is organized first by array name and then by threshold-crossing feature:

neural.<array_name>.<threshold_feature>

For example a dual array file may contain:

neural.lateral.tx_4_5
neural.medial.tx_4_5

and each matrix has shape:

n_time_bins_for_array x n_electrodes_for_array

Arrays from the same yield session can have different numbers of time bins and should not be assumed to be synchronized.

Threshold crossing fields are counts per bin.

De-identification

Participant identifiers are de-identified study identifiers. Calendar dates and direct personal identifiers are not included in these files. Timing is reported as day relative to implantation (post_implant_day).

Licensing and reuse

Dryad data are released under the Creative Commons Zero (CC0) waiver. As a courtesy, users are requested to cite the dataset and associated manuscript when reusing these data.

Contact

For questions about the dataset, contact:

Nick Hahn, nickhahn26@gmail.com
Frank Willett, fwillett@stanford.edu

Version notes

Version 1.0

Initial Dryad release. Use the latest Dryad version for analysis, as later versions may include corrections, additional documentation, or file packaging updates.

Human subjects data

No personally identifiable information (PII) is contained within this data