Skip to main content

Large-scale neural recordings with single neuron resolution using Neuropixels probes in human cortex

Cite this dataset

Paulk, Angelique C et al. (2023). Large-scale neural recordings with single neuron resolution using Neuropixels probes in human cortex [Dataset]. Dryad.


Recent advances in multi-electrode array technology have made it possible to monitor large neuronal ensembles at cellular resolution in animal models. In humans, however, current approaches restrict recordings to few neurons per penetrating electrode or combine the signals of thousands of neurons in local field potential (LFP) recordings. Here, we describe a new probe variant and set of techniques which enable simultaneous recording from over 200 well-isolated cortical single units in human participants during intraoperative neurosurgical procedures using silicon Neuropixels probes. We characterized a diversity of extracellular waveforms with eight separable single unit classes, with differing firing rates, locations along the length of the electrode array, waveform spatial spread, and modulation by LFP events such as inter-ictal discharges and burst suppression. While some challenges remain in creating a turn-key recording system, high-density silicon arrays provide a path for studying human-specific cognitive processes and their dysfunction at unprecedented spatiotemporal resolution. 


Patients & clinical/research electrode placement

All patients voluntarily participated after informed consent according to guidelines as monitored by the Massachusetts General Brigham (previously Partners) Institutional Review Board (IRB) Massachusetts General Hospital (MGH). Participants were informed that participation in the experiment would not alter their clinical treatment in any way, and that they could withdraw at any time without jeopardizing their clinical care. Participants were not compensated monetarily for participating. Recordings in the operating room were acquired with 9 participants (mean= 59 years old, ranging from 34 to 75; 7 female) who were already scheduled for a craniotomy for concurrent clinical intraoperative neurophysiological monitoring or testing for mapping motor, language, and sensory regions and removal of tissue as a result of tumor or epilepsy or undergo intra-operative neurophysiology as part of their planned deep brain stimulator (DBS) placement. Prior to inserting the Neuropixels probe, a small superficial incision in the pia was done using an arachnoid surgical knife. The Neuropixels probe was inserted through this incision. Recordings were referenced to sterile ground and recording reference needle electrodes (Medtronic) placed in nearby muscle tissue (often scalp) as deemed safe by the neurosurgical team though a series of tests ground and reference tests were performed to identify the ideal combinations of ground and reference options.

Neuropixels recordings, data collection & analysis

            Neuropixels probes (NP v 1.0, version S, IMEC) sterilized with Ethylene Oxide (BioSeal) were connected to a 3B2 IMEC headstage wrapped in a sterile plastic bag and sealed using TegaDerm (3M) to keep the field sterile.  Neuropixels probes (NP v 1.0-S, IMEC) include an electrode shank (width: 70µm, length: 10 mm, thickness: 100µm) of 960 total sites laid out in a checkerboard pattern with contacts at ~18 µm site to site distances (16 µm (column), 20 µm (row)). Handling of the electrodes and the headstage from outside the sterile bag was all performed in sterile conditions in the operating room.  The headstage was connected via a multiplexed cable to a PXIe acquisition module card (IMEC), installed into a PXIe Chassis (PXIe-1071 chassis, National Instruments).  For the Neuropixels-1.0 probes as used in this study, the linear dynamic range of the Neuropixels amplifier is 10 mVpp. This range is digitized using a 10 bits Analog to Digital conversion. All Neuropixels recordings were performed using SpikeGLX Release v20201103-phase30 ( on a computer connected to the PXIe acquisition module recording the action potential band (AP, band-pass filtered from 0.3-10 kHz) sampled at 30 kHz and a local field potential band (LFP, band-pass filtered from 0.5-500 Hz), sampled at 2.5 kHz.  Since these Neuropixels probes enable 384 recording channels which can then be used to address 960 electrodes across the probe shank, we used the default electrode map which allowed us to record from the lower third of the probe (the most distal channels).

As the Neuropixels probe is stably rigid with regard to the channels relative to one another, we could estimate depth of the sites in the tissue along the electrode using the LFP signal. When some channels are outside the tissue, we found the LFP signal was relatively noisy and did not show differences between channels when outside of tissue. This electrophysiological marker, however, provided an upper boundary limit of how deep the electrode was in the tissue relative to the measurements on the electrode itself resulting in a depth relative to this upper boundary. However, we did not identify the depth as cortical depth as the Neuropixels probe could be inserted at different angles relative to the sulci and gyri.

            Synchronization was performed through two different approaches.  TTL triggers via a parallel port produced either during a task via MATLAB or custom code from a separate computer were sent to both the National Instruments and IMEC recording systems, via a parallel port system.  In addition, we used the TTL output to send the synchronization trigger via the SMA input to the IMEC PXIe acquisition module card to allow for added synchronizing triggers which were also recorded on an additional breakout analog and digital input/output board (BNC-2110, National Instruments) connected via a PXIe board (PXIe-6341 module, National Instruments). The TTL triggers were produced either during a task via MATLAB or custom code on the task computer.

Data collection and analysis were not performed blind to the conditions of the experiments.

Compensation for tissue movement and electrode alignment through time

            We found clear evidence of vertical tissue movement relative to the Neuropixels probe in the local field potential (LFP) recordings. To confirm that this was due to movement of the tissue as well as effects of heartbeat, we aligned the movement artifact to the heartbeat in time (this was possible thanks to audio tracking of the EKG in 2 participants’ cases).  We found the movement roughly matched this tracking. To confirm that the manual tracking could match the movement of the brain relative to the electrode, we performed tissue-level tracking of the video recordings of the case and found we could align the filmed movement of the brain pumping relative to the electrode, which was well visualized in the LFP band across channels as tracked through time. We tested several approaches to address this movement and correct for the alignment, including the Kilosort 3.0 drift adjustments and estimation ( and spike time-informed alignment approaches ( We chose to use the LFP-informed manual tracking as it was better-resolved in the time domain since the dynamic range of LFP allowed for per time step (0.0004 sec) alignment and interpolation. In contrast, the automatic approach depended on firing rate and arrival of spikes, which were sparse.

Manual tracking of movement using LFP signals

            The signal was first extracted from the binary files into local field potential (LFP, <500 Hz filtered data, sampled at 2500 Hz) and action potential (AP, >500 Hz filtered data, sampled at 30000 Hz) from SpikeGLX using MATLAB and available preprocessing code ( ). We inspected the data visually as well as examined the timeline of the recording to reject noisy time ranges (such as during insertion.) We then further examined the voltage deflections in the LFP for a prominent, bounded deflection in the voltage where we observed the voltage values shifting in unison which was consistently present throughout the recording. We attempted to use a number of algorithms to detect these shifts, but the multiple changes present (heartrate, slow and mid-range drifting, and other shifts) were not effectively tracked by these algorithms. Instead, to capture the displacement in the movement bands, we imported the LFP voltage as an .stl file from MATLAB into Blender ( ), a three-dimensional animation program which allowed for easier manual tracing compared to MATLAB. Using the surface voltage and the Grease Pencil feature, we traced the shifting band of negatively deflecting LFP throughout the recording at a resolution of 500 Hz. The line produced then was exported as a .csv file and imported into MATLAB, where it was compared with the LFP at higher resolution to check whether the manual tracing matched the LFP displacement.  This traced line information was upsampled to 2500 Hz to match the sampling frequency of the LFP channels (interp1, ‘makima’).

Preprocessing AP recordings

            Once we had the LFP baseline to track probe movement through time, we then applied analyses to the AP sampled band. To account for differences in the channels before aligning the data (as channels can have differences in impedance), we first detrended data (which removes best fitted line to each channel), calculated the median, and subtracted it from all channels. We then normalized the voltage signal across channels by multiplying each channel’s voltage time series by a normalization factor where Normalized data = Channel signal * (1/std) * 600. In this case, the std was the standard deviation of channel data without outliers, particularly epochs which were relatively quiet. We defined outliers as elements which were more than 1.5 interquartile ranges above the upper quartile or below the lower quartile of the distribution of voltage signals. Finally, we chose the value of 600 in the normalization to allow us to scale the data up to an int16 format for improved data resolution.

Alignment and interpolation of AP channels for manual registration

            To then re-align the AP channel data so as to offset the movement artifact, we upsampled the traced line to 30KHz to match the AP sampling rate (interp1, ‘makima’). We then, for each time bin, applied a spatial interpolation between channels vertically in two columns of the Neuropixels recording, resulting in a vertical spatial resolution of 1um. These steps resulted in a large, high resolution interpolated matrix that we could then follow through time. This let us compensate for the movement effects by resampling the voltage in space based on the manually registered movement trajectory described in “Manual Tracking of Movement using LFP signals”. 

            Specifically, for each time bin, we shifted the vertical channels vector up or down according the upsampled traced line, resulting in >450 ‘virtual channels’ that each contained voltage information putatively from a specific brain location. Finally, since the virtual channels on both ends (top and bottom of the shank) contained only partial data (due to brain movement relative to the electrode), we selected a subset of 384 virtual channels that contained the most continuous information throughout the recording (and did not shift channels into the edge), which could be inferred from the average channel offset.

Automatic tracking of movement using LFP signals

            We developed a more automatic pipeline, Decentralized Registration of Electrophysiological Data (DREDge;  and ) which relies on decentralized correlation approaches to track the movement from the neural recordings (both using spiking activity and LFP frequency ranges). We saved the tracked motion as well and have uploaded this information along with the interpolated data (see below).

Alignment and interpolation of AP channels for AUTOMATIC registration

After estimating the motion trace using DREDge (  and ), this displacement estimate can be used to apply motion correction to the underlying raw data using a variety of interpolation strategies. We used KiloSort 2.5’s kriging interpolation  enabling it to correct motion estimates sampled at arbitrary frequencies, rather than being locked to KiloSort’s temporal block length. Source code for this approach is available at .

Code involving manual alignment available on Github (  

Code involving preprocessing available on Github 

Code involving DREDge (Decentralized Registration of Electrophysiology Data) Motion registration code (python version used in this processing):

Usage notes

We report and load only the three successful recordings from the cortex of temporal and frontal lobes in patients undergoing brain tissue resection to treat epilepsy (Pt. 03, N=1, under general anesthesia, lateral temporal lobe) or during the implantation of DBS leads to treat movement disorders (Pt. 01 and Pt. 02; N=2, one awake (Pt. 02) and one under general anesthesia (Pt. 01), dorsolateral prefrontal cortex) using Neuropixels probes. In our paper, we also reported unsuccessful recordings – and lessons learned -- from six cases performed while developing these approaches. Unsuccessful recordings were either due to electrode fracture (N=2, with the devices and pieces fully recovered) or excessive noise during the recordings (N=4). Two different types of arrays were tested. In the first two attempts, we used the original Neuropixels 1.0 probe but found it to be too fragile.  Instead, we developed a variant  featuring a thicker shank (Neuropixels 1.0-ST: thickness: 100µm, width: 70 µm, length: 10 mm). This version enabled considerably easier insertions and robust use during neurosurgical cases. This probe version, combined with an improved grounding and reference electrode configuration, enabled us to observe spiking activity from populations of isolatable single neurons in three participants (N=3, Pt. 01-03).

As we also observed movement artifact, we include both the raw data sampled during the recording in the operating room as well as the interpolated data set (saved as a binary file) per participant for spike sorting. Further information is included in the ReadMe file: README_PaulkEtAlNeuropixels.txt

Data included from:
Participant Pt. 01, Neuropixels recording during a procedure for deep brain stimulation (DBS) surgery for treatment of a movement disorder, participant under generalized anesthesia; recording from right dorsolateral prefrontal cortex. 
Participant Pt. 02, Neuropixels recording during a procedure for deep brain stimulation (DBS) surgery for treatment of a movement disorder, participant awake and with monitored anesthesia care (MAC); recording from left dorsolateral prefrontal cortex.  
Participant Pt. 03, Neuropixels recording during a left anterior temporal lobectomy, under generalized anesthesia; recording from the left anterior temporal lobe. 

Further files are as such:

A. raw recording Data, action potential (AP) band, binary file: Pt0X.imec0.ap.bin    

B. meta data for the raw recording Data: Pt0X.imec0.ap.meta    

C. raw recording Data, local field (LF) band, , binary file: Pt0X.imec0.lf.bin        

D. meta data for the raw recording Data: Pt0X.imec0.lf.meta     

F. recording Data realigned to adjust for movement using MANUAL approaches, action potential (AP) band, binary file: Pt0X_aligned_0.bin   

G. recording Data realigned to adjust for movement using AUTOMATIC DREDge approaches, action potential (AP) band, .dat binary file: Pt0X_DREDgealigned_0.dat with tracked motion and a channel map

NOTE: For the Pt. 01 data set, the beginning included adjusting the ground and reference and inserting the electrode. For this reason, the aligned data for sorting (Pt01_aligned_0.bin) starts 330 sec into the recording. 

NOTE: For the Pt. 02 data set, the beginning included adjusting the ground and reference and inserting the electrode. For this reason, the aligned data for sorting (Pt02_aligned_0.bin) starts 220 sec into the recording. 



National Cancer Institute, Award: K24-NS088568

Tiny Blue Dot Foundation

National Institute of Neurological Disorders and Stroke, Award: R01NS11662301

National Institute on Deafness and Other Communication Disorders, Award: R01DC01403406

Simons Foundation, Award: 543045

Howard Hughes Medical Institute, Award: Stanford University

A.P. Giannini Foundation

Wu Tsai Neurosciences Institute, Stanford University, Award: Interdisciplinary Scholars Fellowship

Burroughs Wellcome Fund, Award: Career Award at the Scientific Interface

Brain & Behavior Research Foundation

Grossman Institute