Long-term performance of intracortical microelectrode arrays in 14 BrainGate clinical trial participants

Abstract

Brain–computer interfaces have enabled people with paralysis to control computer cursors, operate prosthetic limbs, and communicate through handwriting, speech, and typing. Most high-performance demonstrations have used silicon microelectrode “Utah” arrays to record brain activity at single neuron resolution. However, reports so far have typically been limited to one or two individuals, with no systematic assessment of the longevity, decoding accuracy, and day-to-day stability properties of chronically implanted Utah arrays. Here, we present a comprehensive evaluation of 20 years of neural data from the BrainGate and BrainGate2 pilot clinical trials. This dataset spans 2,319 recording sessions and 20 arrays from the first 14 participants in these trials. On average, arrays successfully recorded neural spiking waveforms on 35.6% of electrodes, with only a 7% decline over the study enrollment period (up to 7.6 years, with a mean of 2.8 years). We assessed movement intention decoding performance using a “decoding signal-to-noise ratio” (dSNR) metric, and found that 11 of 14 arrays provided meaningful movement decoding throughout study enrollment (dSNR > 1). Three arrays reached a peak dSNR greater than 4.5, approaching that achieved during able-bodied computer mouse control (6.29). We also found that dSNR increases logarithmically with the number of electrodes, providing a pathway for scaling performance. Longevity and reliability of Utah array recordings in this study were better than in prior nonhuman primate studies. However, achieving peak performance consistently will require addressing unknown sources of variability.

Introduction

Brain-computer interfaces have emerged as a promising approach for restoring movement and communication to people with paralysis. Successful demonstrations have included point-and-click cursor control (Jarosiewicz et al. 2015; Pandarinath et al. 2017), robot control (Hochberg et al. 2012; Collinger et al. 2013; Benabid et al. 2019), functional electrical stimulation (Ajiboye et al. 2017), typing (Willett et al. 2021; Vansteensel et al. 2016, 2024; Shah et al. 2024; Jude et al. 2025), and speech (Moses et al. 2021; Willett et al. 2023; Metzger et al. 2023; Card et al. 2024; Wairagkar et al. 2025). Many of the highest performing demonstrations to date have used silicon microelectrode arrays (the “Utah” array) to record brain activity at single neuron resolution (Collinger et al. 2013; Pandarinath et al. 2017; Willett et al. 2021; Card et al. 2024; Wairagkar et al. 2025), providing rich information about the user’s motor intent. However, nearly all reports in humans have been limited to case studies of one or two participants, including those that have focused on summarizing array recording performance over time (Simeral et al. 2011; Downey et al. 2018; Sponheim et al. 2021; Colachis et al. 2021; Hughes et al. 2021). Key questions remain about how decoding accuracy varies across individuals, how long neural recordings last over years, and how stable the recordings are across days. Answers to these questions are needed to place isolated demonstrations into context and understand whether they represent typical signal quality. A more comprehensive evaluation of the Utah array can also set a benchmark against which improvements can be measured, informing the design of future recording devices.

Here, we present a post-hoc analysis of neural recording and motor decoding performance of the Utah array in the first 14 participants enrolled in the BrainGate and BrainGate2 pilot clinical trials between 2004 and 2019 (constituting all available participants at the time when this analysis was started, but not including more recent participants enrolled between 2021 and 2025). Encouraging safety results from these same 14 participants were recently reported in an interim study (Rubin et al. 2023). Like prior assessments in nonhuman primates (Suner et al. 2005; Chestek et al. 2011; Barrese et al. 2013; Sponheim et al. 2021), here we evaluated the percentage of electrodes which successfully recorded neural spiking waveforms over time (array “yield”). Additionally, since most arrays were placed in the same brain region (hand knob area of motor cortex) and all participants regularly engaged in BCI-enabled cursor control tasks, we were able to compare neural decoding quality across the 9 participants for which high-bandwidth (30 kHz) neural data was available. We assessed the stability of neural tuning over time, how cursor decoding accuracy compares to mouse movement accuracy in able-bodied volunteers, and how decoding accuracy scales with the number of electrodes (which is important for understanding how higher channel count devices might improve decoding). These functional measurements are more directly relevant for BCI applications than array yield, particularly because low-amplitude background activity can be highly useful for neural decoding even if it fails to generate clear spiking waveforms (Nason et al. 2020).

Results

Between 2004 and 2019, 14 participants were enrolled in the BrainGate and BrainGate2 pilot clinical trials; each participant had one or two 96-channel platinum-metalized Utah microelectrode arrays placed in cortex. Nineteen of the total twenty arrays were placed in the hand area of motor cortex (Figure 1A, B), with one array placed in the middle frontal gyrus. The BrainGate and BrainGate2 clinical trials were designed to evaluate the safety of implanted microelectrode arrays and to discover fundamental insights into the effective design of intracortical BCIs. Testing new neural decoding architectures and task designs was common throughout the trials, and as such it is not feasible to summarize across all types of data collected.

Fig. 1 |. Datasets curated for assessing recording and decoding performance.

(A) One or two 96 channel intracortical microelectrode arrays were placed in the cortex of 14 participants as part of the BrainGate and BrainGate2 clinical trials. 19 of 20 arrays were placed in the hand knob area of the dominant precentral gyrus (B), and one was placed in the middle frontal gyrus of Participant T10. (C) Example spike waveforms. For this study, 30 Khz recordings were decimated to 15 Khz, band pass filtered with a pass band of 250–5000 Hz, re-referenced with linear regression referencing (LRR) and thresholded at −4.5 times the robust standard deviation of the voltage signal for each channel. (D) To assess decoding performance, we retrospectively analyzed sessions in which participants were asked to attempt to move a computer cursor towards a defined target. Imagery varied between sessions and participants. Tasks included Fitts, Radial-8, grid tasks and their variations. To assess array recording yields, data from all types of sessions were used (including sessions where no cursor control occurred). (E) Example cursor trajectories for a Radial-8 task. (F) Overview of recording duration and sessions analyzed. Total array implant days ranged from 296 to 2780 with a mean of 1063. After excluding sessions with excessive electrical noise, 2,319 recording sessions and 634 decoding sessions were analyzed. T10’s middle frontal gyrus array was only used for recording evaluation, as it was in a different brain area with neural tuning properties not comparable to the hand knob area. High bandwidth neural data (30 kilo-samples per second) required for retrospectively evaluating decoding performance in a uniform manner were not consistently available for participants S1, S2, S3, A1, and T1, and so these participants were not included in decoding analyses.

To summarize findings spanning 14 participants over 15 years, here we curated two types of datasets: (1) a “yield dataset” containing a single neural recording of at least 5 minutes duration from every research session, which was used to estimate array yield (Figure 1D), and (2) a “decoding dataset” consisting of all neural recordings during which a BCI was used to move a computer cursor towards well-defined targets (Figure 1E), which was used to estimate neural decoding accuracy, neural tuning stability and electrode count scaling properties. This dataset consists of a more limited number of days and participants, because not every experimental session included BCI cursor control, and because for the first 5 participants high bandwidth neural data (30 kHz) was either not recorded during cursor control (S1, S2, A1), not recorded for the first year post-implant (S3), or could not be consistently aligned to cursor task data (T1). Note that prior publications have shown successful demonstrations of neural cursor control in S1 and S2 (Hochberg et al. 2006), S3 (Kim et al. 2008; Simeral et al. 2011; Jarosiewicz et al. 2013; Bacher et al. 2015; Jarosiewicz et al. 2015), A1 (Kim et al. 2008), and T1 (Jarosiewicz et al. 2013).

Neural spiking activity persists for multiple years with little average decline

We first investigated how the number of electrodes able to detect spiking activity (“yield”) changed over time, looking across a total of 2,319 recording sessions. Spiking activity was defined as the detection of at least 2 spiking events per second when using a threshold of −4.5 times the robust standard deviation of the voltage signal (see Methods). We observed substantial variability across 20 arrays in 14 different participants (Figure 2A), with some arrays sharply declining in yield within the first year (S2, S3 and T2) or having consistently low yield (T3), some arrays increasing over time (T7, T8), and many arrays staying relatively constant or gradually declining over time (e.g., T5, T6, T10, T11). Yield trajectories in the first year were not always predictive of yields over longer periods (e.g., T7’s medial array increased in yield after the first year). Yield curves were estimated using locally weighted scatterplot smoothing (LOWESS)(Cleveland 1979). Although the cause of S2 and T2’s sudden declines in yield are not known, S3’s decline coincided with and may have been triggered by mechanical trauma to the pedestal that occurred outside of research study sessions. T3’s array, which had a consistently very low yield, was believed to have had its wire bundle unintentionally placed under tension during surgery, which could have resulted in the recording tips being largely epicortical rather than intracortical.

Fig. 2 |. Long-term trends in neural spiking activity recording performance.

(A) Array yield over the recording duration of twenty 96-channel microelectrode arrays. Each point represents the number of electrodes detecting spiking activity in a single neural recording of 5 minutes duration or longer (one recording per session). Spiking electrodes are defined as electrodes detecting a threshold crossing rate greater than 2 Hz when using a −4.5 × robust standard deviation threshold (see Methods). LOWESS curves are shown in black. For participants with two arrays, data points in the lighter shade represent the more medial array. Implant date is indicated above each participant panel. Participant S2’s array experienced a steep increase in yield after repair of the percutaneous pedestal (day 190) which placed conductive epoxy on each bond site to restore connections. Months later, array yield declined sharply again (day 323), and it is unknown whether this was due to failure of the previous repair or an unrelated event. Participant S3 experienced an accidental forceful contact of the pedestal against her headboard on the evening of day 181, with a decrease in yield that slowly, partially recovered. (B) Overall deltas in yield. Deltas were calculated by taking the mean number of spiking electrodes in sessions recorded during the last 3 months post implant and subtracting the mean number of spiking electrodes recorded during the first 3 months of recording. (C) 6 month deltas in yield, up to six and a half years post-implant. Each point represents the mean spiking electrode count over a 6 month period minus the mean spiking electrode count over the previous six month period. The red line plots the mean of deltas. Outliers are labeled, with participant T2, S2, and S3 all experiencing a steep decline. (D) Array proportions exceeding a certain yield over the first 3 years of recording. If data was not recorded from an array during a certain time period, that array was not included in the calculation for that time period.

Averaged over all days, participants, and arrays, mean array yield (percentage of electrodes recording spiking activity) was 35.6%. Arrays experienced only a modest 7% average decline in yield between the first 3 months of recording and the last three months of recording (Figure 2B; 41% yield in the first three months to 34% yield in the last three months). However, there was large variability in yield across arrays (standard deviation: 20.5%). A minority of arrays in a subset of participants contributed substantially to this variability, including the three arrays that showed a sharp decrease in yield over the first year (S2, S3, and T2; Figure 2C). Overall, results were favorable compared to prior work in nonhuman primates, with over 50% of arrays maintaining at least 20% yield over the first three years (Figure 2D; compare to Sponheim et al. 2021, where less than 15% of arrays maintained a 20% yield). This demonstrates the potential of long-term, reliable detection of neural spike waveforms with planar microelectrode arrays in humans.

We also examined electrode impedance and found that it increased sharply post-implant, followed by a consistent and gradual decline (Supplementary Figure 1) consistent with prior reports (Barrese et al. 2013; Barrese, Aceros, and Donoghue 2016). Spike waveform amplitude and background noise amplitude also declined over time (Supplementary Figure 2A–B), and smaller waveform amplitudes were associated with lower impedance values (Supplementary Figure 2C). Since both the spike waveform and background noise amplitudes declined at similar rates, the waveform signal-to-noise ratio remained relatively constant over time (Supplementary Figure 2D). This overall decline in signal amplitude and impedance may be due to insulation degradation (Barrese et al. 2013; Barrese, Aceros, and Donoghue 2016), particularly around the edges of the planar array where low impedances were sometimes observed progressively over time (e.g., T2, T3, T5, T6; Supplementary Figure 3). Nevertheless, these changes did not appear to affect detection of spiking signals or decoding performance (see next section).

Movement intention can be decoded reliably for years in most arrays

Next, we examined the extent to which cursor movement intention could reliably be decoded from intracortical recordings, how decoding varied across arrays, and how it compared to the accuracy of able-bodied cursor control using a computer mouse. A total of 634 sessions across 14 arrays and 9 participants were analyzed to evaluate movement intention decoding during closed loop cursor control tasks (Figure 1F).

To evaluate the amount of intended movement information in the neural signals recorded by an array, we quantified how well an offline linear decoder could transform neural activity into a “movement intention” unit vector pointing from the cursor to the target, using neural features from a short time window at the beginning of each movement. The decoder output from each trial can be visualized as a dot in 2D space and colored according to the ground truth direction; the more accurate the decoder is, the more the dots should form a tight ring whose colors form an orderly gradient (Figure 3A). Accurate decoder output represents movement intention vectors that point more accurately at the target with a consistent magnitude, and would thus enable higher quality cursor control, given that cursor control algorithms often use linear mappings between neural features and cursor velocity (Kim et al. 2008; Gilja et al. 2012; Willett et al. 2017, Pandarinath, et al. 2017).

Fig. 3 |. Retrospective assessment of movement decoding signal-to-noise ratio across time.

(A) Decoding signal-to-noise ratio (dSNR) and example dSNR plots. dSNR is a vector-based, movement intention decoding metric that quantifies how well movement intention can be decoded from all electrodes of a given array during a target-directed cursor control task. To estimate dSNR, retrospectively computed linear decoder output within a 300 ms window is decomposed into a signal component (pointing toward the target) and a noise component to calculate a decoding SNR. Each point in the dSNR plots denotes a trial, color coded by target location. Spike band power neural features were retrospectively computed in a uniform manner for all participants from whom high bandwidth neural recordings were available for analysis. A linear decoder was then trained in a uniform manner to decode these neural features, ensuring that results are independent of whatever algorithms were used at the time of data collection. (B) Example SNR plots for able-bodied cursor control using a computer mouse. Cursor trajectories from an initial push towards the target were used to calculate dSNR (C) Decoding SNR over time. Each point represents a dSNR value for one session. LOWESS curves are shown in black for each array. (D) Overall deltas in dSNR. Deltas were calculated by taking the mean dSNR in sessions recorded during the last 3 months post implant and subtracting the mean of dSNR recorded during the first 3 months. (E) Decoding SNR distributions. Arrays are displayed in implant order. The dotted line indicates mean decoding SNR across 9 able-bodied participants using a computer mouse. (F) Decoding SNR and array yield are correlated (r = 0.52, p = 2e-68). When excluding the T6 array, which had low array yield but high information content in spike band power features, this relationship is more pronounced (r = 0.70, p = 9e-133)

To quantify the quality of the offline decoder output, the decoding signal-to-noise ratio (dSNR) metric (Willett, Murphy, et al. 2017) decomposes each decoder output vector into a signal and a noise vector (Figure 3A) and takes the ratio of their magnitudes (see Methods). dSNR is independent of the particular task parameters, decoding architecture, and signal processing used in real-time during data collection, and thus enables meaningful post-hoc comparisons across a diversity of datasets that would otherwise not be possible via task performance metrics alone. It can also be used to quantify target-directed able-bodied mouse movements (Figure 3B), contextualizing the neural dSNR values. To compute dSNR for able-bodied mouse movements, we used the endpoint of the computer cursor after an initial “ballistic” center-out cursor movement to define the “decoder output” vector.

There was wide variability in dSNR across participants and arrays (Figure 3C), although on average dSNR increased by 0.34 across the duration of recording (ranging from 1.2 to 7.6 years, Figure 3D). Most arrays maintained dSNR values greater than 1, indicating a persistent ability to decode meaningful movement intention signals. A minority of arrays had low dSNR at the end of first year post-implant and did not recover dSNR over the course of study participation (T2, T3, T11 Lateral). High-performing arrays approached able-bodied dSNR values (Figure 3E; T5 Lateral, T5 Medial, T11 Medial), but typical arrays yielded about 1/3 the dSNR of able-bodied computer mouse movement (average computer mouse dSNR=6.29).

dSNR was correlated with array yield (pearson r: 0.52, p: 2.10e-68; Figure 3F), but array yield did not fully account for the large variability of decoding performance, in part due to information contained in “spike band power” features (e.g. participant T6). The results presented above used spike band power features to decode movement intention, since spike band power (power contained in the 250–5000 Hz band) can capture low-amplitude spiking signals that might otherwise be missed by threshold crossing features. We evaluated how dSNR varied as a function of neural features used for decoding, and found that spike band power led to higher performance on average as compared to threshold crossing rates, even at lower thresholds (Supplementary Figure 5). Overall, these outcomes show that long-term persistence of meaningful decoding performance is typical and can occur even at low array yields, although absolute dSNR levels can vary widely across participants and arrays.

Neural representations of movement intention remain stable on short time scales

Next, we investigated the day-to-day changes in how movement intention is represented in the recorded neural activity. Neural tuning stability was analyzed by using a linear encoding model to describe how spike band power features were tuned to a movement intention vector that points from the cursor to the target. After fitting, model coefficient similarity was assessed by correlating the coefficients of models fit on different days (see Methods). Only arrays from the decoding dataset where movement intention could be decoded reliably (dSNR >1) were analyzed (3 arrays excluded), yielding a total of 593 recording sessions and 11 arrays from 7 participants.

Day-to-day tuning correlations exhibited a variety of patterns (Figure 4A), due in part to potential changes in motor imagery across days and/or tasks, varying task instructions, and variable intervals between consecutive cursor control sessions. Nevertheless, many arrays showed clear regimes of long-lasting tuning stability. Average stability as a function of time elapsed (Figure 4B) was generally high within the time span of 1 month (8 of 11 arrays, r>0.6). This is encouraging, as short-range stability enables decoders calibrated on prior days to be used on future days without requiring user interruption for re-calibration (Card et al. 2024; Wilson et al. 2023; Fan et al. 2023; Hosman et al. 2023).

Fig. 4 |. Stability of neural tuning to movement intention across days.

(A) Tuning stability heatmaps. A linear encoding model was used to describe neural activity as a function of a point-at-target vector that describes the user’s movement intention during a cursor control task. The Pearson correlation between model coefficients for each session pair is shown in each (i,j) entry. Blocks of blue color indicate time periods of high stability. (B) Tuning stability as a function of day separation. Tuning remains stable on the time scale of a month (8 of 11 arrays, r>0.6), which could enable decoders calibrated on prior days to be used on future days without requiring re-calibration. Line colors correspond to the heatmap outlines in (a).

Electrode count scaling provides a pathway to higher performance

Finally, we examined how decoding performance scales with the number of electrodes. We computed average dSNR across all cursor control sessions, for each array and participant, as a function of electrode count by randomly subsampling from all 96 electrodes (n=500 resamplings per electrode count). We found that dSNR scaled as an approximately logarithmic function of the number of electrodes, up to 192 electrodes for participants with dual arrays and 96 electrodes for participants with a single array (Figure 5). Only arrays providing meaningful movement intention decoding (dSNR >1) were evaluated. By linearly extrapolating to 1024 electrodes, a subset of arrays (T5 Lateral, T5 Medial, T6, T11 Medial) demonstrated a projected dSNR that was 70% or greater than the average computer mouse dSNR observed across 9 able-bodied volunteers. These results are supportive of the many ongoing efforts towards higher channel count devices (Musk and Neuralink 2019; Sahasrabuddhe et al. 2021; Hettick et al. 2022), although it appears that for arrays with lower dSNR values (e.g., T8), many orders of magnitude more electrodes would be required to reach able-bodied computer mouse performance.

Fig. 5 |. Decoding SNR scales logarithmically with electrode count.

For each array, decoding SNR (dSNR) was assessed using a random subset of electrodes to investigate how dSNR varies as a function of electrode count (for each electrode count, dSNR was averaged over 500 random subsets). For each single array, electrode counts ranged from 6 to 96. For participants with dual arrays, combined electrode counts ranged from 6 to 192. Extrapolated data is represented with a dotted line and combined arrays are represented in black. dSNR scales logarithmically with the number of electrodes. Here, for arrays exhibiting higher performance per electrode (e.g., t5, t11), performance approaches able-bodied levels as electrode count is extrapolated up toward 1024 electrodes.

Discussion

We analyzed the performance of Utah microelectrode arrays placed in the first fourteen participants of the longest running clinical trial for implanted brain-computer interfaces. Overall, our findings show that Utah arrays typically continue to record neural spiking waveforms for years with little average decline, contrary to some prior reports in nonhuman primates (Barrese et al. 2013; Sponheim et al. 2021). Even when array yields are low, decoding SNR can still be high due to the information contained in spike band power features (e.g., participant T6) (Nason et al. 2020), and incorporating spike band power features is now standard practice (Willett et al. 2023; Card et al. 2024; Gusman et al. 2025). All arrays except two yielded functional movement intention signals, and nearly all (11 of 14) did so throughout the duration of study participation. Array recordings generally showed higher decoding SNR over the most recent 10 years compared to previous arrays (but see T11 Lateral; Figure 3E), possibly reflecting process improvements over time. Neural tuning stability was also reasonably high in many arrays (correlation > 0.6 for day pairs within a 1-month time span). Good tuning stability allows decoders calibrated on a previous day to be used on a new day, reducing or eliminating the need for interrupting the user to recalibrate (Jarosiewicz et al. 2015; Sussillo et al. 2016; Wilson et al. 2023; Fan et al. 2023; Hosman et al. 2023; Card et al. 2024)

Despite favorable results overall, some arrays did fail to consistently produce useful decoding signals over time (3 of 14 arrays evaluated for decoding; T2, T3, T11 Lateral; Figure 3A), experienced large declines in yield within the first year (2 of 19; S2, S3; Figure 1A) or had consistently low yield (4 of 19; S1, T3, T6, T11 Lateral; Figure 1A). Additionally, there was substantial variability in decoding quality that was not fully accounted for by yield (Figure 3F). While some top performing arrays provided movement intention signals that enabled directional decoding comparable to able-bodied computer mouse movements, there was a wide range of decoding SNRs, and adding more electrodes to low performing arrays was not projected to enable top performance within the range of ~1000 electrodes. To ensure consistently high decoding SNR, it may be necessary to discover the underlying factors driving variability in signal quality, which could include disease state, underlying injury, implant location, implant technique, variability in device manufacture, and foreign body response.

It has been suggested that very late-stage ALS may lead to motor cortical signal loss (Lawyer and Netsky 1953; Eisen and Weber 2001; Vansteensel et al. 2024). This study included six participants living at the time with ALS (A1, T1, T3, T6, T7, T9), although none reached a completely locked-in state. For these participants, array yield was relatively high in some cases (A1, T1, T7, T9) and low in others (T3, T6), with no clear relationship to their remaining ability to move, speak, or use their eyes for communication. Decoding SNR could not be assessed in A1 and T1 (lack of stored high-bandwidth recordings); dSNR was poor in T3 (with low yield) but did not appear to decline over time in participants T6, T7 and T9. Although these dSNR and array yield measurements are somewhat variable, they appear to be consistent with the broader study group.

Implant location may also affect decoding SNR if the hand area of motor cortex contains tuning “hot spots” for directional cursor movement. If so, more precise presurgical neuroimaging or broader coverage of relevant areas of motor cortex could improve decoding SNR. Additionally, variable foreign body response could be a factor, since silicon microelectrode arrays are known to cause tissue disruption in animal models (Nolta et al. 2015) and neural connectivity can decline even while neurons are spared (Gregory et al. 2023; Chen et al. 2024). Surgical technique may also affect array performance, potentially via minimization of tissue disruption and foreign body response.

Finally, it is important to note that arrays included in this study spanned 15 years of varying manufacturing processes, which would have contributed to some of the variability we observed. Also, arrays implanted in early trial participants used arrays with electrodes 1.0 mm in length (S1, S2, A1, T1, T3, T6) whereas later electrodes were 1.5 mm in length (S3, T2, T5, T7 through T11); combined with individual differences in cortical thickness, this would be expected to introduce variability regarding which cortical layers were recorded.

Data collected over decades of human intracortical BCI research has demonstrated the growing potential of intracortical BCIs as a viable assistive technology, and many commercial efforts are currently underway to develop novel recording devices with improved durability and greater number of electrodes (Musk and Neuralink 2019; Sahasrabuddhe et al. 2021; Hettick et al. 2022). As these new recording technologies become available, progress toward more performant BCI systems will require standardized benchmarks with which to compare devices. For cursor control applications, a variety of performance metrics have been applied, including Fitts bit rate, achieved bit rate, typing rate, target acquisition time and path length (Hochberg et al. 2006; Simeral et al. 2011; Gilja et al. 2015; Jarosiewicz et al. 2015; Nuyujukian et al. 2015; Pandarinath et al. 2017; Brandman et al. 2018). However, these metrics are sensitive to task parameters (e.g., target hold time or click decoding latency) and decoding algorithms, which vary widely across studies and have rarely been tuned for optimal comparative performance. Here, we used a decoding SNR metric that quantifies the amount of linear movement direction information present in a neural population, independent of decoding algorithm and task parameters, enabling a meaningful comparison across 10 years of ever-evolving research session designs. This metric could be used to compare devices aiming to enable BCI cursor control, and variants of this metric may enable comparison across other domains (e.g., speech), serving as an informative complement to clinical outcomes assessments that measure the combined efficacy of a particular device and neural decoding approach in restoring useful function.

Funding Statement

This work was supported by Office of Research and Development, Department of Veterans Affairs (N2864C, N9228C, A4820R, A2295R, B6453R, B6459L, A6779L, P1155R, A2827R, A3803R), NIH NIDCD (R01DC009899), NIH NIDCD (U01DC017844), NIH NIDCD (R01DC014034), NIH NIDCD (U01DC019430), NIH NICHD-NCMRR (N01HD53403), NIH NICHD-NCMRR (N01HD10018), NIH NICHD-NCMRR (R01HD077220), NIH NINDS, (U01NS098968), NIH NINDS (U01NS062092), NIH NINDS (UH2NS095548), NIH NINDS (U01NS123101), NIH NICHD (RC1HD063931), Simons Foundation (543,045), Howard Hughes Medical Institute, Doris Duke Charitable Foundation, Conquer Paralysis Now (004698), AHA (19CSLOI34780000), ALS Association (20-MALS-553), Larry and Pamela Garlick, Samuel and Betsy Reeves, MGH-Deane Institute, The Executive Committee on Research (ECOR) of Massachusetts General Hospital, Wu Tsai Neurosciences Institute, Bio-X Institute at Stanford, Robert J. and Nancy D. Carney Institute for Brain Science, Brown University School of Engineering, Brown University Office of the Vice President for Research

Data and materials availability

All data required to reproduce these results will be publicly released upon acceptance in a peer-reviewed journal.