Implanted neural recording system vs Electroencephalography

Implanted neural recording system vs Electroencephalography

We are searching data for your request:

Forums and discussions:
Manuals and reference books:
Data from registers:
Wait the end of the search in all databases.
Upon completion, a link will appear to access the found materials.

Can someone help me to understand why in nowadays the scientists develop implanted neural recording system to measure and collect neural signals? Why the use of the electroencephalogram is not sufficient to record neurons? Which is the difference between these systems?

Thanks in advance

EEG records a rough average of activity from a large area of the brain, because it is located far away from the actual neurons. There is no way to separate out the activity of individual or even small groups of cells; instead, you record lower frequency synchronous activity among whole populations.

Implanted systems can record from small groups or sometimes even single neurons, which greatly increases the amount of information available.

Depending on the purposes of the recording or research, EEG might be sufficient; in other cases, implantable electrodes are required. EEG is sufficient to monitor sleep/wake cycles, for example; there is no need to record intracranially to monitor sleep (unless perhaps you are studying mechanisms of sleep itself). Implanted electrodes are often used before surgery to remove parts of the brain that are causing seizures. Using ECoG (basically an "EEG on the dura" rather than on the skull) or depth electrodes allows the surgeon to know more precisely which area is causing seizures, hopefully leading to a more successful surgery and needing to resect less tissue.

Implanted electrodes are also commonly used in animal experiments to learn about how single neurons and small populations encode or process information.

Neural decoding of expressive human movement from scalp electroencephalography (EEG)

  • 1 Laboratory for Noninvasive Brain-Machine Interface Systems, Department of Electrical and Computer Engineering, University of Houston, Houston, TX, USA
  • 2 Center for Robotics and Intelligent Systems, Instituto Tecnológico y de Estudios Superiores de Monterrey, Monterrey, Mexico
  • 3 Department of Biomedical Engineering, University of Houston, Houston, TX, USA
  • 4 Department of Neurobiology, University of Maryland, College Park, MD, USA
  • 5 Department of Dance, University of Maryland, College Park, MD, USA

Although efforts to characterize human movement through electroencephalography (EEG) have revealed neural activities unique to limb control that can be used to infer movement kinematics, it is still unknown the extent to which EEG can be used to discern the expressive qualities that influence such movements. In this study we used EEG and inertial sensors to record brain activity and movement of five skilled and certified Laban Movement Analysis (LMA) dancers. Each dancer performed whole body movements of three Action types: movements devoid of expressive qualities (“Neutral”), non-expressive movements while thinking about specific expressive qualities (“Think”), and enacted expressive movements (𠇍o”). The expressive movement qualities that were used in the “Think” and 𠇍o” actions consisted of a sequence of eight Laban Effort qualities as defined by LMA𠅊 notation system and language for describing, visualizing, interpreting and documenting all varieties of human movement. We used delta band (0.2𠄴 Hz) EEG as input to a machine learning algorithm that computed locality-preserving Fisher's discriminant analysis (LFDA) for dimensionality reduction followed by Gaussian mixture models (GMMs) to decode the type of Action. We also trained our LFDA-GMM models to classify all the possible combinations of Action Type and Laban Effort quality (giving a total of 17 classes). Classification accuracy rates were 59.4 ± 0.6% for Action Type and 88.2 ± 0.7% for Laban Effort quality Type. Ancillary analyses of the potential relations between the EEG and movement kinematics of the dancer's body, indicated that motion-related artifacts did not significantly influence our classification results. In summary, this research demonstrates that EEG has valuable information about the expressive qualities of movement. These results may have applications for advancing the understanding of the neural basis of expressive movements and for the development of neuroprosthetics to restore movements.


  • 1 Department of Biology, Colby College, Waterville, ME, United States
  • 2 Department of Psychology, Colby College, Waterville, ME, United States

Meditation is an umbrella term for a number of mental training practices designed to improve the monitoring and regulation of attention and emotion. Some forms of meditation are now being used for clinical intervention. To accompany the increased clinical interest in meditation, research investigating the neural basis of these practices is needed. A central hypothesis of contemplative neuroscience is that meditative states, which are unique on a phenomenological level, differ on a neurophysiological level. To identify the electrophysiological correlates of meditation practice, the electrical brain activity of highly skilled meditators engaging in one of six meditation styles (shamatha, vipassana, zazen, dzogchen, tonglen, and visualization) was recorded. A mind-wandering task served as a control. Lempel–Ziv complexity showed differences in nonlinear brain dynamics (entropy) during meditation compared with mind wandering, suggesting that meditation, regardless of practice, affects neural complexity. In contrast, there were no differences in power spectra at six different frequency bands, likely due to the fact that participants engaged in different meditation practices. Finally, exploratory analyses suggest neurological differences among meditation practices. These findings highlight the importance of studying the electroencephalography (EEG) correlates of different meditative practices.


Six participants (4 females, mean age 30.3 ± 9.6, see Material and methods and Fig 1B and 1C) performed a delayed saccade task (Fig 1A) while electrophysiological data were recorded from multilead EEG depth electrodes. In each trial, participants were instructed to perform horizontal saccades toward 1 of 2 targets but only after a variable delay period. The information about saccade direction was indicated by a visually presented central cue (Cue 1), followed by a saccade execution Go signal (Cue 2). The task consisted of 3 interleaved experimental conditions (Fig 1A): In the Free condition, a diamond at Cue 1 prompted the participants to freely choose the direction of the forthcoming saccade. In the Instructed condition, an arrow pointing left or right indicated to the participants the direction of the saccade they were to prepare. After a variable delay (3.5–7.75 seconds) during which the participants prepared the adequate saccade while fixating the central fixation point, a Go signal (Cue 2) prompted the participants to immediately execute the saccade. In the Control condition, participants were presented with a square at Cue 1, indicating that they would need to wait for the Go signal (Cue 2) to find out the required saccade direction and execute it immediately. Behavioral saccade onset latency data were collected, and spectral power features were extracted from the iEEG data across multiple time windows and all electrode sites. Power features were computed in 5 standard frequency bands: theta (θ) [4–8 Hz], alpha (α) [8–15 Hz], beta (β) [16–30 Hz], low-gamma (low γ) [30–60 Hz] and high gamma (high γ, HG) [60–140 Hz]. A supervised machine learning framework was implemented to decode (through space, time, and frequency) the experimental conditions (free, instructed, and control) and thereby identify the most discriminant neural patterns that distinguish between free-choice and instructed actions during saccade planning and execution (see Material and methods for details).

A. Experimental design of the delayed motor task. For each trial, participants were instructed to perform horizontal saccades toward one of 2 targets after a delay of 3,750 milliseconds, 5,750 milliseconds or 7,750 milliseconds, depending on a visually presented central cue appearing briefly for 250 milliseconds. B. Top, left, and right views of the number of recording sites that contribute to each vertex (i.e., spatial density) projected on a standard 3D MNI brain. Electrodes contribute to a location when they are within 10 mm of a given site on the brain surface. In all brain images, right side of the image is the right side of the brain. C. Top, left, and right view of the depth-electrode recording sites, projected on a standard 3D MNI brain. Each color represents a participant. Left: Rostral is up Right: Medial views. D. Barplot of mean reaction times for the 3 conditions across all participants (Control, Instructed, Free). Each triangle represents the mean reaction times for 1 participant. The data underlying this panel D can be found in S1 Data. MNI, Montreal Neurological Institute.

Behavioral results

We computed the mean reaction times (RTs, i.e., saccade onset latency, see Material and methods) for each experimental condition across all participants and found that mean RTs were significantly longer for the Control condition (mean RT = 466 ± 66 milliseconds) compared with both Free (mean RT = 334 ± 36 milliseconds t(6) = 2.75, p = 0.0403) and Instructed (mean RT = 321 ± 33 milliseconds t(6) = 2.68, p = 0.0435) conditions (see Fig 1D). No significant differences were found between Free and Instructed conditions (t(6) = 1.31, p = 0.24). These results were also observed at the single participant level in 5 out of 6 participants (see Material and methods section). These results are consistent with the fact that the availability of saccade target information (whether self-generated or instructed) during the delay period allowed the participants to plan the upcoming saccades and hence, execute them faster upon the Go signal compared with the Control condition, in which no directional information was available during the delay period. Mean saccade duration, saccade speed, mean latency, and the number of saccades executed per condition by each participant are reported in S2 Table.

To assess modulations of the neural activity across the 3 delayed saccade trial types (i.e., Free, Instructed, and Control) over space, frequency, and time, we computed time-frequency representations (locked either on stimulus onset, i.e., Cue 1, or saccade execution cue, i.e., Cue 2), as well as single-trial spectral amplitude envelopes in multiple frequency bands (theta (θ) [4–8 Hz], alpha (α) [8–15 Hz], beta (β) [16–30 Hz], low-gamma (low γ) [30–60 Hz] and high gamma (high γ, HG) [60–140 Hz]). Fig 2 illustrates these feature computations and the high quality of the intracranial data by showing time-frequency maps derived from electrodes in FEF and intraparietal sulcus (IPS) in participant 2, as well as single-trial HG activity, aligned to stimulus presentation and to saccade Go signal (ordered by saccade onset latencies). In addition, we used linear discriminant analysis (LDA) to probe the ability of these spectral features to decode experimental conditions from single-trial data. Importantly, we applied this machine learning framework individually to data from each recording site and in a time-resolved manner over the course of the task (see Material and methods section for details).

Time-frequency maps (left) and single-trial HG plots (right) from 2 recording sites in an illustrative participant (P2). Data are shown for the 3 experimental conditions (Control, Instructed, and Free), during planning (Cue 1, stimulus onset), and execution (Cue 2, go signal). Trials in the single-trial gamma plots are sorted according to saccade onset latencies. FEF, frontal eye field HG, high-gamma IPS, intraparietal sulcus Modul., modulations Rel., relative.

Decoding delay-period neural activity in free choice versus instructed saccade trials

To identify the neural patterns related to making autonomous choices, we first compared the delay-period neural responses observed during free-choice saccade trials to those recorded during instructed saccade trials. This was conducted by applying LDA to classify Free versus Instructed saccade trials based on spectral amplitude estimated during the delay interval in all 5 frequency bands (Fig 3, see also S4 Fig). Panels A-C of Fig 3 show that, among all frequency bands, HG activity was the neural feature that provided the highest decoding accuracy (DA) and largest number of significantly decoding sites when classifying Free versus Instructed trials, during the delay period ([0 3,000 milliseconds] after Cue 1). The HG activity led to statistically significant classification in 61 sites (4 out of 6 participants) and yielded a maximum DA of 92.9% and a mean DA of 79% (Fig 3A–3C). Interestingly, out of the 2 participants that did not yield any significantly decoding sites, P5 was the only participant that showed similar RTs across the 3 conditions (see Fig 1D). This might reflect the fact that this participant did not make use of Cue 1, which would explain why HG activity could not significantly decode Free versus Instructed conditions during the delay period in this participant.

A. Summary of all significant electrodes by participant across frequencies showing that the largest clusters were found in the HG frequency band. B. Mean and C. Maximum decoding accuracies across participants and significant electrodes for each frequency band for Free versus Instructed classification (error bars represent SEM). D. Time course of baseline corrected (−500 to−100 milliseconds) HG activity aligned on Cue 1, for all electrodes that significantly classify Free versus Instructed conditions and H. its associated mean decoding accuracy across significant electrodes. E. Maximum decoding accuracies across participants and significant electrodes for each frequency bands for Free versus Instructed multielectrode classification. F. Relative mean HG peak activity (in %) and G. latency (in milliseconds) for electrodes significantly decoding Free versus Instructed conditions during the delay period (from 0 to 3,000 milliseconds after Cue 1). I. Decoding Free versus Instructed conditions with HG activity in 5 successive time windows during the delay period (0 to 500 milliseconds 500 to 1,000 milliseconds 1,000 to 1,500 milliseconds 1,500 to 2,000 milliseconds 2,000 to 3,500 milliseconds after Cue 1, and −2,000 to 0 milliseconds before Cue 2). Only sites with significant decoding accuracies are shown (p < 0.01, with max stats correction across electrodes, time, and frequency bands). J. Percent relative power change ([Free − Instructed]/Instructed) for all significant sites shown in panel I. The data underlying this Figure can be found in S1 Data. DA, decoding accuracy elec., electrodes Freq., frequency HG, high-gamma Inst., instructed Nb, number Rel., relative.

We then used a multifeature classification approach (Free versus Instructed trials) in which observations across all electrode sites were now included simultaneously in the decoding feature space (repeated for each frequency band). We assessed the statistical significance of time-resolved DA using permutation tests, corrected for multiple comparisons across participants (electrodes, frequencies, and time points). As shown in Fig 3E, the multisite DA was highest for HG activity, reaching 86.8%. Given that both single and multisite classification results (Fig 3A–3C and 3E) indicate that HG amplitude is the most prominent predictor of target class (Free versus Instructed), the next sections of the results focus on the characterization of the fine-grained temporal and spatial profiles of HG neural decoding.

Averaging the HG data across all trials from all significantly decoding sites illustrates the temporal dynamics of delay HG activity that distinguish between Free and Instructed saccade trials (Fig 3D). The associated time-resolved mean DA is shown in Fig 3H. Because the analysis is based on averaging across all sites, panels D and H only provide a schematic representation of the temporal dynamics, without statistical assessment. Thus, we conducted standard paired t tests to further quantify the difference in HG peak amplitudes and latencies between Instructed (mean peak HG amplitude = +39% ± 0.74 mean peak latency = 475 milliseconds ± 14) and Free (mean peak HG amplitude = +24% ± 1.18 mean peak latency = 812 milliseconds ± 97) conditions from all 61 significantly decoding sites (in 4 out of 6 participants). The results revealed significant differences (peak amplitude: t(4) = 8.34, p < 0.003, Fig 3F peak latency: t(4) = 21, p < 0.0002, Fig 3G) and were also confirmed in single-trial analyses performed individually in each of the 4 participants (p < 0.05, see Material and methods). All the statistical results within and across participants are listed in S3 Table.

Additionally, Fig 3I and 3J represent HG decoding dynamics resolved across both space and time: In the early part of the delay period, significant decoding electrodes are associated with stronger HG power in the Instructed than in the Free condition. But over time, HG power then becomes higher for the Free saccade planning than for Instructed saccade planning during later stages of the delay period in frontoparietal brain areas. More specifically, we show that all significant electrodes in the (0, 500 milliseconds) time-window after Cue 1 during the delay period are associated with higher HG activity in the Instructed condition (Fig 3I and 3J, [0, 500 milliseconds]). On the other hand, as time goes by, significant electrodes begin to become more associated with higher HG activity in the Free compared with the Instructed condition. From 1,000 to 2,000 milliseconds after Cue 1, we see that all significant electrodes are now associated with higher power in the Free condition (Fig 3I and 3J, [1,000, 1,500 milliseconds], [1,500, 2,000 milliseconds]). Additionally, we find that from 1,500 milliseconds to 2,000 milliseconds after Cue 1, the only electrodes that still significantly decode Free from Instructed conditions are located in frontal regions (Fig 3I and 3J [1,500, 2,000 milliseconds]). Interestingly, we also found sites for which HG activity was still significantly stronger in Free than Instructed, at the end of the delay period, from −2,000 to 0 milliseconds before Cue 2 (Fig 3I and 3J). This suggests that some electrodes may display persistent activity lasting throughout the whole delay period and led to subsequent analyses described in more detail in the following sections.

In order to further characterize the temporal HG dynamics specific to Free and to Instructed saccades (beyond the strict difference between the two) while also taking into account low-level stimulus-related processes, we replicated the same decoding framework but now with the goal of distinguishing each of the main 2 conditions from the Control condition (i.e., Instructed versus Control, and Free versus Control, see Fig 4). First, we found that over the first 3,000 milliseconds after Cue 1, there was an overlap of 56 electrodes between the electrodes that significantly classify Instructed versus Control trials as well as Free versus Control trials. In other words, 82.4% of the sites that significantly discriminate Free versus Control conditions also significantly discriminate Instructed versus Control conditions (Fig 4A). Furthermore, we found that when participants freely chose the saccade direction, the delay HG activity lasted, on average, 618 milliseconds ± 57, whereas the instructed saccade condition displayed mean HG durations of only 368 milliseconds ± 60. The difference was statistically significant (length of time points above the significance threshold in Free versus Control compared with Instructed versus Control classifications, t(4) = 3.52, p < 0.04 across 4 participants and confirmed in intraparticipant analysis in 1 individual with p < 0.05, see Fig 4B). We also found that HG activity during Instructed saccade planning (mean onset = 152 milliseconds ± 39) reached significant classification earlier than HG activity in the free-choice condition (mean onset = 465 milliseconds ± 49) when compared with the control (latency of first significant decoding accuracies in Free versus Control compared with those associated with Instructed versus Control classifications, t(4) = 7.09, p < 0.006 across 4 participants, see Fig 4C). This difference was confirmed with intraparticipant analyses in 3 out of 4 participants with p < 0.05. Lastly, we show that DA peaked significantly earlier in the Instructed condition (mean onset = 527 milliseconds ± 61) than in the Free condition (mean onset = 822 milliseconds ± 70) when compared with the Control condition (Peak DA latencies in Free versus Control compared with Instructed versus Control classifications, t(5) = 3.39, p < 0.03 across 5 participants, confirmed with intraparticipant analyses in 4 out of 5 participants with p < 0.05, see Fig 4D). All the statistical results within and across participants are listed in S3 Table.

A. Location of electrode sites where HG activity discriminates Free versus Control and/or Instructed versus Control mapped on transparent 3D brain images for all participants (p < 0.01, corrected). Left: electrodes colored in green, blue, and yellow, respectively, indicate sites that discriminate Free versus Control trials only, Instructed versus Control only, or both Free versus Control and Instructed versus Control during the delay period (0 to 3,000 milliseconds after Cue 1). Right: colors indicate different participants. B. Duration (length of time points) above the significance threshold C. Decoding onset (i.e., latency of first significant decoding accuracies) D. Latency of the peak decoding accuracies (in milliseconds) for sites significantly decode Free versus Control (in green) and Instructed versus Control (in blue) across participants. E, F. Time course of baseline corrected (−500 to −100 milliseconds) HG activity aligned on Cue 1, for all electrodes that significantly classify Instructed versus Control (E) and Free versus Control (F) conditions, and G, H. Their associated mean decoding accuracy across significant electrodes in time, respectively. I. Temporal generalization of trial-type decoding using HG activity across significant sites derived from the previous analyses (Free versus Control and Instructed versus Control) during the delay period (0 to 3,000 milliseconds after Cue 1) for 4 participants. Generalization matrices show decoding performance plotted as a function of training time (vertical axis) and testing time (horizontal axis). Decoding of Instructed versus Control (left column) trials illustrates the expected profile for transient coding, while decoding of Free versus Control (right column) trials leads to smoother and extended decoding patterns, typical of a single process that is sustained over time. The data underlying this Figure can be found in S1 Data. DA, decoding accuracy HG, high-gamma Inst., Instructed Nb, Number Rel., Relative.

Probing HG delay temporal dynamics via temporal generalization

In order to further characterize the fine temporal organization of information-processing during the delay period in the Instructed and Free-choice conditions, we probed cross-temporal generalization [11,53] of decoding Instructed versus Control and Free versus Control conditions using HG activity (Fig 4I). In brief, temporal generalization consists in training a classifier with data from one time point, t1, and testing it on data from a different time point, t2. In principle, cross-temporal generalization indicates that the neural code identified at t1 also occurs at t2 (see Material and methods). More specifically, we used temporal generalization to better characterize short-lived (transient) and longer-lasting (sustained) HG activity processes underlying Free and Instructed planning. Our findings show that HG activity during instructed saccade planning yields a generalization pattern typical of transient coding (see the first column of Fig 4I), whereas free choice is characterized by a HG decoding process that is more sustained in time (second column of Fig 4I). Taken together, the observed cross-temporal decoding patterns and their accuracies are consistent with the view that decision-related discriminant HG activity during the delay period is more sustained and starts later in Free-choice trials compared with Instructed saccade trials. In contrast, when no choice is involved, task-related information reflected in HG activity is more transient and most relevant shortly after stimulus onset. Importantly, the cross-temporal generalization results also highlight that although HG decoding in the Free choice is more sustained, it does not last systematically the entire duration of the delay period until the GO signal (Cue 2). This is consistent with our earlier observations in Fig 3I and 3J, that over the course of the delay period, fewer sites discriminate between free and instructed trials. This may suggest that although several sites display sustained HG activity for free-choice saccade trials, only a few actually display persistent increases up until saccade execution. This important distinction is probed in the next section.

Spatial distribution of early versus late delay HG activity during free choice

To specifically isolate the brain areas where HG increases index neural processing specific to free saccade decisions, we used a conjunction analysis (Free > Control Ո Free > Instructed) applied to all electrode sites with significant classification of Free versus control and Free versus Instructed trials. Importantly, we conducted this conjunction analysis in 2 distinct time windows during the delay period: An “early” window was defined as the first 2,000 milliseconds after Cue1, and a “late” window from −2,000 to 0 milliseconds before Cue 2. Fig 5A depicts for both time windows the sites with significant decoding accuracies for all participants in which the increase in HG activity was stronger during free choice compared both with control (Free > Control) and instructed (Free > Instructed) trials (see also S5 Fig for more details). Fig 5B combines the results for early and late into a common representation (35 electrodes, 4 participants) indicating, thereby, for each free-choice specific site whether it showed HG increases only in the early (0 to 2,000 milliseconds after Cue 1), only in late (−2,000 to 0 milliseconds before Cue 2) or in both early and late intervals. We found that most free-choice-specific sites exhibited enhanced HG activity only during the early part of the delay period (29 electrodes, 3/4 participants) in a network of regions including superior frontal gyrus (SFG), middle frontal gyrus (MFG), SMA, IPS, and FEF (see Fig 5C and 5D, first 2 rows for illustrative early electrodes in IPS and SFG). However, for 5 sites (in 2/4 participants) located in SFG (2 electrodes), MFG (2 electrodes), and FEF (1 electrode), significant decoding accuracies were found both in the early and late parts of the delay period, indicating HG activity spanning the entire duration of the delay period (see Fig 5C and 5D, last row for an illustrative early + late electrode in SFG). The localization details of all 35 electrodes depicted in Fig 5B are provided in S4 Table. Note that, among all participants, the electrode site with maximum HG DA for both Free versus Control (86.1%) and Free versus Instructed (91.4%) was located in the right IPS (P2, electrode derivation p9-p8, see Fig 5, last row of C, D).

A, Electrode sites with significant decoding accuracies (p < 0.01, corrected) for all participants mapped on transparent 3D brain images when HG activity is significantly stronger in the Free condition than in the Control condition (first row) and when HG activity is significantly stronger in the Free condition than in the Instructed condition (second row) during the delay period, from 0 to 2,000 milliseconds after Cue 1 (first column, early) and from −2,000 milliseconds before Cue 2 (second column, late). B. Electrode sites where HG is higher in Free compared with Instructed and Control, determined by a conjunction analysis (Free > Control U Free > Instructed). Free-choice-specific sites are colored in blue if significant decoding was observed in the early part of the delay in yellow if significant decoding was found in the late part and in green for sites that survived the conjunction analysis both in early and late phases of the delay period. For 3 individual electrodes, we plotted HG activity over time (C, The data underlying this panel can be found in S1 Data), single-trial plots (D, upper row) and time-frequency-maps (D, lower row) for Free, Instructed, and Control conditions. DA, decoding accuracy Freq., frequency HG, high-gamma IPS, intraparietal sulcus MFG, middle frontal gyrus modul., modulations Rel., relative SFG, superior frontal gyrus.

To further appreciate individual participant contributions to the global findings, we also analyzed all electrode sites that survived the conjunction analysis in Fig 5B, grouping the data either by regions of interest (ROIs) or delay-period window (early or late). The results (Fig 6) largely speak to the similarity of temporal HG dynamics across regions and conditions. Data from P2 indicate that the Control condition elicits HG responses in IPS but not in MFG and that the strongest and longer-lasting HG responses in MFG comes from the Free-choice trials (Fig 6A and 6B). This is consistent with an involvement of parietal regions—among other things—in low-level sensory processing and a prominent role of frontal HG activity in deliberation. The distinction between merely “longer-lasting” HG activity (early) and “persistent” HG activity throughout the delay period (early and late) is shown in Fig 6C. We then conducted an analysis to evaluate trial history effects during the free-choice condition. We used an unpaired t test to evaluate statistical differences between n-1 conditional probabilities and random (history-free) probabilities for each participant and found no obvious across-trial dependences in choice behavior during the free condition for 5/6 subjects (see S1 Fig). For one participant (P2), we found a significant trial history effect that was driven by a tendency to alternate behavior (e.g., left, right, left, right…). Interestingly, this participant showed the same delayed and sustained HG free-choice effects observed in the other participants, with the notable difference that the HG response was shorter-lived compared with other participants (Fig 6B). This result seems to be consistent with our interpretation that the sustained HG activity observed during the Free condition is indeed related to deliberation between competing alternatives in frontoparietal brain areas. Specifically, this analysis suggests that the tendency of P2 to alternate between left and right saccade choices was associated with shorter deliberation during the delay (indexed by shorter sustained HG responses). Note that although significant, the observed alternating behavior in P2 was not systematic (see conditional probabilities reported in S1 Fig). Lastly, to verify whether our findings could not be confounded by spatial tuning effects in the delay period, we replicated the classification analyses separately for left and right saccade trials. However, no significantly spatially tuned delay-period decoding was found in the Free, Instructed, or Control conditions (see S6 Fig). More specifically, the 3 illustrative sites shown in Fig 5C and 5D did not show statistically significant differences when trials for left and right saccades were investigated separately (see S6B–S6D Fig). We also ruled out the possibility that our findings were confounded by involuntary saccades made in response to the presentation of Cue 1 by contrasting mean EOG traces for left and right trials and for Free and Instructed conditions (see S3 Fig).

Mean HG activity time courses for free-choice-specific sites grouped here by (A) ROIs, (B) subjects, and (C) early/late. Mean time course of baseline corrected (−500 to −100 milliseconds) HG activity for Free, Instructed, and Control conditions aligned on Cue 1 (first column) and Cue 2 (second columns) in electrodes that have enhanced HG in the free-choice condition compared with both the control and instructed saccade conditions (i.e., determined by a conjunction analysis (see Fig 5B). The data underlying this figure can be found in S1 Data. HG, high-gamma FEF, frontal eye field IPS, intraparietal sulcus MFG, middle frontal gyrus Rel., relative ROI, region of interest SMA, supplementary motor area.

Disentangling the correlates of oculomotor execution and oculomotor planning

Previous reports using intracranial EEG in humans have shown that saccade execution in response to a Go signal is associated with distributed increases in HG power [50]. Yet it has been so far hard to determine whether such gamma activity reflects target selection, motor planning, actual oculomotor commands, or a combination thereof. Analyzing the execution component (cue 2) of the delayed saccade paradigm in the present study provides an opportunity to address some of these questions. Therefore, in the final analysis, we set out to compare HG responses induced by the Go signal in conditions in which target selection already occurred in the delay period (i.e., Free and Instructed) to the Control condition, in which participants were given no information on saccade target before the Go signal. To achieve this, we conducted a supervised classification analysis on Free, Instructed, and Control conditions, but this time using data collected during saccade execution (0 to 2,000 milliseconds after Cue 2). We found that, during saccade execution, the HG power modulations were similar whether the saccade was instructed or self-chosen (i.e., no significant classification between Free and Instructed). However, HG activity associated with saccade execution in the Control condition was significantly different from both Free and Instructed saccades (Fig 7A). These results are consistent with the fact that no significant difference in reaction time (saccade onset latency) were found between Free and Instructed conditions, whereas mean reaction time across participants was significantly longer for the Control condition compared with both Free and Instructed trials (see Fig 1D).

A. Electrodes with significant decoding accuracies (p < 0.01, corrected) for all participants are mapped on transparent 3D brain images when HG activity is significantly stronger in the Control condition than in the Free condition (first row) and when HG activity is significantly stronger in the Control condition than in the Instructed condition (second row) in the interval from 0 to 2,000 milliseconds after Cue 2. B. using a conjunction analysis (Control > Free Ո Control > Instructed), we show sites in which HG is stronger in the Control condition than in the Free and Instructed conditions. Colored sites correspond to 3 individual electrodes, for which we plotted HG activity over time (C, The data underlying this panel can be found in S1 Data), single-trial HG plots (sorted according to RTs) (D, upper row) and time-frequency-maps (D, lower row) for Free, Instructed, and Control conditions. DA, decoding accuracy HG, high-gamma FEF, frontal eye field Freq., frequency IPS, intraparietal sulcus MFG, middle frontal gyrus modul., modulations Rel., relative ROI, region of interest SMA, supplementary motor area.

Next, we examined the trial-by-trial relationship between saccade onset latency and neural activity specifically in areas that exhibit these significant HG differences between the Control condition and the Free and Instructed conditions. To this end, we first used a conjunction analysis to identify the sites of interest defined as sites for which significant classification was mediated by HG power in the Control condition being higher than in the other 2 conditions (Fig 7B). This identified 28 electrodes in parietal and frontal regions (2/6 participants at p < 0.01). Interestingly, in 43% of these significant sites, the reverse pattern was true during the delay period: HG activity was significantly stronger in the Free and the Instructed conditions compared with the Control condition during the delay. Three representative examples of this task-specific HG pattern inversion between delay and execution windows are shown in Fig 7C (first column). This suggests the involvement of these HG responses in action selection processes: When such processes are engaged during the delay period during Free and Instructed conditions, they do not need to be repeated during the execution period. By contrast, in the Control condition, no action selection processes were possible during the delay period (hence, the weaker HG activity), but they were recruited at execution.

By plotting the mean time courses of HG power (Fig 7C), as well as mean time-frequency representations and single-trial HG power plots, sorted according to RT (Fig 7D), it becomes clear that the temporal dynamics of HG power differ quite substantially depending on electrode location and trial type. In order to probe the various relationships between saccade onset and HG activity in these areas in a quantitative manner, we computed Pearson's rank correlation coefficients between saccade onset latency and HG onset latency across trials in each of the 3 experimental conditions (see Material and methods). Significant correlations were observed in a limited number of sites, and the results need to be interpreted with caution. This said, we observed 3 correlation patterns that we considered to be of interest: For some sites, the onset of HG activity after the Go signal did not correlate with saccade onset in any of the 3 conditions (pattern 1, execution independent). Other sites exhibited correlations between HG onset and saccade onsets across all conditions (pattern 2, oculomotor execution). Finally, in the third pattern, correlations between HG onset and saccade onset latencies were only observed in Control trials (pattern 3, oculomotor planning). The recording sites that displayed these patterns came from distinct brain areas we found evidence for pattern 1 (i.e., no correlation with saccade onsets in any condition) for the electrodes located in the IPS (4 sites, 1 participant) (e.g., Fig 7C and 7D, first row and S7 Fig) (e.g., Electrode p9-p8 Control [p = 0.62, r = −0.06], Free [p = 0.69, r = −0.06], and Instructed [p = 0.43, r = −0.12]). Next, pattern 2 (i.e., saccade and HG onset latencies significantly correlated in all 3 conditions) was observed in SMA (1 out of 4 electrodes) (e.g., Fig 7C and 7D second row and S7 Fig Electrode b3-b2 Control [p = 0.001, r = 0.32], Free [p = 0.003, r = 0.29], and Instructed [p = 0.041, r = 0.2]). The fact that these correlations were present in the 3 conditions is consistent with the involvement of SMA in oculomotor execution. Thirdly, in the middle frontal gyri (1 out of 3 electrode), we observed the third pattern 3, namely that HG onsets were correlated with saccade onset latencies only in the Control condition (e.g., Fig 7C and 7D, last row and S7 Fig: Electrode m13-m12 Control [p = 0.04, r = 0.26], Free [p = 0.13, r = 0.26], and Instructed [p = 0.67, r = −0.07]). Despite being of interest, this observation is not surprising given that the relevant sites did not have significant HG response after cue 2 in the Free and Instructed conditions (e.g., third row of Fig 7C and 7D). The detailed list of HG and saccade onset correlations across the 3 conditions are available in S5 Table.

Implantable neurotechnologies: a review of micro- and nanoelectrodes for neural recording

Electrodes serve as the first critical interface to the biological organ system. In neuroprosthetic applications, for example, electrodes interface to the tissue for either signal recording or tissue stimulation. In this review, we consider electrodes for recording neural activity. Recording electrodes serve as wiretaps into the neural tissues, providing readouts of electrical activity. These signals give us valuable insights into the organization and functioning of the nervous system. The recording interfaces have also shown promise in aiding treatment of motor and sensory disabilities caused by neurological disorders. Recent advances in fabrication technology have generated wide interest in creating tiny, high-density electrode interfaces for neural tissues. An ideal electrode should be small enough and be able to achieve reliable and conformal integration with the structures of the nervous system. As a result, the existing electrode designs are being shrunk and packed to form small form factor interfaces to tissue. Here, an overview of the historic and state-of-the-art electrode technologies for recording neural activity is presented first with a focus on their development road map. The fact that the dimensions of recording electrode sites are being scaled down from micron to submicron scale to enable dense interfaces is appreciated. The current trends in recording electrode technologies are then reviewed. Current and future considerations in electrode design, including the use of inorganic nanostructures and biologically inspired or biocomapatible materials are discussed, along with an overview of the applications of flexible materials and transistor transduction schemes. Finally, we detail the major technical challenges facing chronic use of reliable recording electrode technology.

This is a preview of subscription content, access via your institution.


The idea of connecting the human brain to a computer or machine directly is not novel and its potential has been explored in science fiction. With the rapid advances in the areas of information technology, miniaturization and neurosciences there has been a surge of interest in turning fiction into reality. In this paper the authors review the current state-of-the-art of brain–computer and brain–machine interfaces including neuroprostheses. The general principles and requirements to produce a successful connection between human and artificial intelligence are outlined and the authors’ preliminary experience with a prototype brain–computer interface is reported.

For the past 40 years researchers have been trying to transform thought into action. Recent advances in neuroscience and neurotechnology have initiated a renewed interest in the development of brain–machine interfaces (BMIs) or brain–computer interfaces (BCIs) that can restore lost motor or sensory function. The currently available systems depend on neural activity input from cortical surface recording (electroencephalographic [EEG] signals) or extracellular brain recording. So far, several patients have benefited from BCI devices, although the current information transfer rate remains a limiting factor. This technology might be especially useful for patients who are otherwise unable to move or speak. This article provides a review of the current state of technology, with emphasis on neuromotor prosthetics. Furthermore, it summarizes the authors’ experience in the development of a human BMI.

The idea of connecting a computer to a brain is not new. As early as the 1950s it was possible to implant single or multiple electrodes into the cortex of humans and animals for recording or stimulation. 1 The result was sometimes spectacular “control” of an animal’s motor behavior or attempted influence of neurological disorders. 2,3 With the worldwide introduction of computers, and ongoing miniaturization, several research groups have started to look into the potential applicability of such BMIs, BCIs, or neural prostheses for use in patients. These devices, by extracting signals directly from the brain, might help to restore abilities to patients who have lost sensory or motor function because of disease or injury. In essence, the computer is used as a surrogate for the damaged region (eg, the spinal cord in quadriplegic patients) and, in the case of a neuromotor prosthesis, acts to interpret brain signals and drive the appropriate effector (eg, muscles or a robotic arm).

Review of Literature

Probably the most widely accepted neural prosthesis in human use is the cochlear implant, 4,5 which substitutes a small computer chip for the damaged inner ear control organ to enable sound waves to be transformed into electrical signals the brain can interpret. Other research has focused on restoring vision for the blind 6–10 with implantable systems to transmit visual information. One other possible application is the restoration of motor control for patients with movement disorders, a population numbering ≈2 million in the United States alone. 11 In an especially tragic situation, brain stem stroke can leave patients in a locked-in state with minimal eye movements and no speech, but full cognitive functions. Among the other diseases that could be helped by BMIs are degenerative disorders (amyotrophic lateral sclerosis or Lou Gehring disease, multiple sclerosis, muscular dystrophy), brain or spinal cord injury, or cerebral palsy. When a disconnection of the main motor pathway occurs, the information generated in the motor cortical areas cannot travel through the pyramidal tract to reach the executing organ, the muscles. There are several possible approaches in how to overcome this disconnect in the signal pathway: (1) activation of intrinsic alternate pathways (anatomical compensation) (2) repair or regeneration of the damaged pathway (anatomical recovery) and (3) bypassing the damaged area by means of a BMI (functional recovery).

Although BMIs are not capable of activating alternate pathways (anatomical compensation) or truly restoring the structural lesion to its original state (anatomical recovery), they may be helpful in restoring lost function (functional recovery).

Typically, motor BMIs consist of at least 3 distinct modules: (1) the data acquisition module 2) the data interpretation module and (3) the data output module. A functional neuroprosthesis must address each stage efficiently and safely. These 3 stages of the process are discussed here.

Data Acquisition Module: EEG signals and Microelectrodes

The purpose of this module is to extract electrical signals from the brain with sufficient bandwidth and at a favorable signal-to-noise ratio. Although the EEG signals can be obtained in a noninvasive manner through single or multiple electrodes that are mounted on the head or assembled in headgear, 12–18 the signal represents a field potential rather than specific cellular activity. Patients and volunteers have been shown to be able to learn to self-regulate slow cortical potentials 19 or to voluntarily control a specific frequency component of the EEG. 20–22 Yet because the EEG signal, especially when recorded with noninvasive methods, is a gross potential generated by the synchronous activity of large numbers of neurons, its resolution is limited temporally and spatially. This means that the rate at which information, and therefore patient intent, is extractable from the EEG signal is limited. This limited bandwidth may be sufficient for many applications, such as interaction with a computer cursor, but may be insufficient for more complex signals, such as when detailed movement trajectories must be specified. In contrast, other groups have focused on obtaining signals from the cortex through invasive methods. Kennedy et al describe their method of implanting glass cone electrodes that are filled with a substrate containing nerve growth-stimulating substances. 23–25 This extracellular recording electrode is bioactive and requires nerve fibers to sprout into the glass cone, a process which occurs over a period of several weeks. With this device, the researchers were able to produce long-term recording from the’ cerebral motor cortex of patients. 25–27 More commonly used are bio-inactive arrays that provide extracellular recording through multiple microwires 28,29 or multichannel electrode arrays in the more superficial motor areas 30 or deep brain structures. 31 It is of note that any of these electrodes are capable of recording from multiple neurons at the same time, a requirement for extraction of signals with the necessary bandwidth for use in prosthetics.

Recent research has focused on exploring various cortical and subcortical areas for optimum electrode array placement. Although traditionally the primary motor cortex (M1) was assumed to be the optimum location for extracting neural signals for use with BMIs designed to substitute for movement, the consensus is growing that alternative or multiple locations may provide the best signals. Among the alternative brain regions investigated are parietal cortex, 32–34 the premotor cortex, 35,36 or simultaneously across M1 and premotor cortices, 37 frontoparietal cortices, 29,38 or subcortically from the basal ganglia. 31 Because various regions’ neural activities occur at varying times in the movement planning and execution process, signals from some areas may be more suited than others depending on the type of information extracted and computational algorithm used.

Data Interpretation Module

Once the signal has been obtained from the recording site, it is fed into a signal processing unit through telemetric communication 27 or direct wire contact. 30 Most researchers recommend early pre-amplification and digitization of the analog recording data to minimize the signal deterioration. The main goal of the interpretation module is to transform the digitized brain signal into a code that best-represents the desired action. For a motor prosthetic, this may be movement of a cursor, clicking of a button, or specification of a complex time-varying movement trajectory, such as reaching for a glass. Continuous movement such as in the hand tracking a moving target can be represented by a Cartesian X-Y-Z coordinate system in extrapersonal space. This approach has been used by several research groups and has resulted in very good approximation of 2-dimensional and 3-dimensional arm or hand movement. 15,27,28,32,39–42 Furthermore, distinct states of limb movement, ie, reaching versus grasping, can now be identified by interpretation modules. 38 The ability to distinguish between these different modes of desired action expands the usefulness of the system and may aid in increasing accuracy and decreasing the effects of errors in decoding. For example, if decoding of continuous motion results in some error, or jitter, around the desired position, holding a full glass of water may be problematic. However, if the system can determine that the user intends to hold the glass still, the position of the artificial effector can be held fixed by turning off the position decoding.

It will be important for a BMI to be able to decode discrete movement classes, such as initiation and termination, or selection between several choices, such as in typing, along with continuous decoding for optimum functionality. 37 The mathematical models used for interpretation and decoding of intent include linear regression algorithms, best fit models, and neural networks, 21,29,43–46 and the best BMI may use >1 of these simultaneously. Although initially decoding was performed off-line, it is now possible with advanced algorithms and recording systems to accurately predict movement in up to 90% of trials either in real time or with only milliseconds delay. 30,43,47 Moreover, accuracy of decoding may be augmented through feedback to the user. It is known that subjects have the ability to consciously alter their neural activity in certain brain areas with sufficient training. 48 Patients, then, should be able to improve decoding of their intent through practice. This reduces the burden on the algorithm and ameliorates potential concerns about drift in the population of neurons that is being observed.

Data Output Module

After the data set is translated into appropriate coordinates or output classes, it can be used to drive a variety of output devices that become the “effector organ” in lieu of muscle-activated limbs. For one, the information has been successfully used to control a computer cursor. 14,15,17,26,27,35,41,42 This opens ample opportunities ranging from simple move-and-click functions to Internet and e-mail use or command or a virtual keyboard. 18 The same signal can also be used to control a robotic device for instance as a substitute for a moving human limb. 49–52 Possibly the most attractive yet difficult to achieve is provided when the computer generated signal can be fed back into the patient’s own limb to activate muscles, for example, through a functional electrical stimulation system. One group has been able to train normal subjects and a neuroprosthesis user to control EEG rhythms to stimulate a motionless hand to open and close or to move a cursor to targets on a computer screen. 15 Although provocative, training took months, however, and ultimate functional movement was less than practical because of low resolution of the EEG signal.

Brown University Experience

Recently, research at Brown University has begun to focus on human control of neural activity, both on-line and off-line. We have been able to take advantage of the implantation of deep brain stimulators for improvement of motor disorders to explore the ability of patients to control their neural signals. Parkinson patients and essential tremor and dystonia patients who are unable to benefit from medication may opt to have a DBS implanted in one of several basal ganglia nuclei to improve tremor, rigidity, bradykinesia, or dyskinesias. 53 As part of this neurosurgical procedure, extracellular recording from brain regions “en route” to the basal ganglia target is routinely performed to give added localization information to the neurosurgeon. With institutional review board and patient permission, we have recorded from 4 to 6 neurons simultaneously in the premotor and prefrontal regions during visuomotor arm movement tasks. With a few minutes of practice, patients have been able to control their neural activity to bring a cursor to a target. 35 Off-line decoding using a maximum likelihood estimator revealed that small, pseudo-randomly selected neuronal ensembles in the human cortex contain information about movement direction and intent (to move or not to move) 36 (Ojakangas et al, data unpublished, 2004). These results are promising in that they demonstrate the possibility of using neuronal activity in nonprimary motor areas for neural prosthetic development, which may be required or even advantageous with certain types of patients.


With the current state of technology, all essential steps for development of a human motor neural prosthesis are in place. Research at multiple institutions continues to refine surgical implantation techniques and analysis algorithms, as well as computer software, that can take more efficient advantage of signals directly derived from human brains. Especially because Food and Drug Administration approval was recently granted for a pilot study using the Cyberkinetics, Inc Braingate electrode array system, the dream of turning “thought into action” may soon become reality for patients with severe motor disabilities.

We acknowledge the generous support by Cyberkinetics, Inc, Foxborough, Mass. G. Friehs, V. Zerris, and M. Fellows are paid consultants, and J. Donoghue is chief scientific officer for Cyberkinetics, Inc, a company involved in developing a human neuromotor prosthetic device.


An ongoing challenge confronting basic scientists, as well as those at the translational interface, is the ability to access a rapid and cost-effective tool to uncover mechanistic details of neural function and dysfunction. For example, identifying the presence of stroke, establishing altered neural dynamics in traumatic brain damage, and monitoring changes in neural profile in athletes on the sidelines all pose major hurdles. In this paper, using scalp electroencephalography (EEG) signals with relatively little data, we provide theoretical and empirical support for a method for the noninvasive detection of neural silences. We adopt the term silences or regions of silence to refer to the areas of brain tissue with little or no neural activity. These regions reflect ischemic, necrotic, or lesional tissue, resected tissue (e.g., after epilepsy surgery), or tumors 1,2 . Dynamic regions of silence also arise in cortical spreading depolarizations (CSDs), which are slowly spreading waves of silences in the cerebral cortex 3,4,5 .

There has been growing utilization of EEG for diagnosis and monitoring of neurological disorders such as stroke 6 , and concussion 7 . Common imaging methods for detecting brain damage, e.g., magnetic resonance imaging (MRI) 8,9 , or computed tomography 10 , are not portable, are not designed for continuous (or frequent) monitoring, are difficult to use in many emergency situations, and may not even be available at medical facilities in many countries. However, many medical scenarios can benefit from portable, frequent/continuous monitoring of neural silences, e.g., detecting changes in tumor or lesion size/location and CSD propagation. Noninvasive scalp EEG is, however, widely accessible in emergency situations and can even be deployed in the field with only a few limitations. It is easy and fast to setup, portable, and of lower cost compared with other imaging modalities. Additionally, unlike MRI, EEG can be recorded from patients with implanted metallic objects in their body, e.g., pacemaker 11 .

Source vs. silence localization

An ongoing challenges of EEG is source localization, the process by which the location of the underlying neural activity is determined from the scalp EEG recordings. The challenge arises primarily from three issues: (i) the underdetermined nature of the problem (few sensors, many possible locations of sources) (ii) the spatial low-pass filtering effect of the distance and the layers separating the brain and the scalp and (iii) noise, including exogenous noise, background brain activity, as well as artifacts, e.g., heart beats, eye movements, and jaw clenching 12,13 . In source localization paradigms applied to neuroscience data 14,15,16 , e.g., in event-related potential paradigms 17,18 , scalp EEG signals are aggregated over event-related trials to average out background brain activity and noise, permitting the extraction of the signal activity that is consistent across trials. The localization of a region of silence poses additional challenges, of which the most important is how the background brain activity is treated: while it is usually grouped with noise in source localization (e.g., authors in 16 state: “EEG data are always contaminated by noise, e.g., exogenous noise and background brain activity”), estimating where background activity is present is of direct interest in silence localization where the goal is to separate normal brain activity (including background activity) from abnormal silences. Because source localization ignores this distinction, as we demonstrate in our experimental results below, classical source localization techniques, e.g., multiple signal classification (MUSIC) 19,20 , minimum norm estimation (MNE) 15,21,22,23 , and standardized low-resolution brain electromagnetic tomography (sLORETA) 24 , even after appropriate modifications, fail to localize silences in the brain (“Methods” details our modifications on these algorithms).

To avoid averaging out the background activity, we estimate the contribution of each source to the recorded EEG across all electrodes. This contribution is measured in an average power sense, instead of the mean, thereby retaining the contributions of the background brain activity. Our silence localization algorithm, referred to as SilenceMap, estimates these contributions, and then uses tools that quantify our assumptions on the region of silence (contiguity, small size of the region of silence, and being located in only one hemisphere) to localize it. Because of this, another difference arises: silence localization can use a larger number of time points (than typical source localization). For example, 160 s of data with the sampling frequency of 512 Hz provides SilenceMap with around 81,920 data points to be used, boosting the signal-to-noise ratio (SNR) over source localization techniques, which typically rely on only a few tens of event-related trials to average over and extract the source activity that is consistent across trials.

Further, we confront two additional difficulties: lack of statistical models of background brain activity, and the choice of the reference electrode. The first is dealt with either by including baseline recordings (in absence of silence which we did not have for our experimental results) or utilizing a hemispheric baseline, i.e., an approximate equality in power measured at electrodes placed symmetrically with respect to the longitudinal fissure (see Fig. 1b). While the hemispheric baseline used here provides fairly accurate reconstructions, we note that this baseline is only an approximation, and an actual baseline is expected to further improve the accuracy. The second difficulty is related: to retain this approximate hemispheric symmetry in power, it is best to utilize the reference electrode on top of the longitudinal fissure (see Fig. 1a). Using these advances, we propose an iterative algorithm to localize the region of silence in the brain using a relatively small amount of data. In simulations and real data analysis, SilenceMap outperformed existing algorithms in localization accuracy for localizing silences in three participants with surgical resections using only 160 s of EEG signals across 128 electrodes (see “Results” for more details on finding the minimum amount of EEG data for localizing silences using SilenceMap).

a The EEG recording protocol and the locations of scalp electrodes. One of 10 reference electrodes (shown in red) is chosen along the longitudinal fissure for rereferencing against. b Average power of scalp potentials for different choices of reference electrodes. c Symmetric brain model of a patient (UD) with a right occipitotemporal lobectomy. d Steps of SilenceMap in a low-resolution source grid. A measure of the contribution of brain sources in the recorded scalp signals ( ( ilde<eta >) ) is calculated relative to a hemispheric baseline. In the brain colormap, yellow indicates no contribution. A contiguous region of silence is localized based on a convex spectral clustering (CSpeC) framework in the low-resolution grid. e Steps of SilenceMap in a high-resolution source grid. The source covariance matrix (Cs) is estimated through an iterative method, and the region of silence is localized using the CSpeC framework. f Choosing the best reference electrode to reference against (Cz in this example), which results in minimum scalp power mismatch (ΔPow). The localized region of silence for this patient (UD) has 13 mm COM distance (ΔCOM) from the original region, with more than 38% overlap (JI = 0.384), and it is 32% smaller (Δk = 0.32).


The brain and the spinal cord together make up the central nervous system (CNS). The functions of the human brain have been the focus of neuroscience research for a long time. However, the spinal cord is largely ignored, and the functional interaction of these two parts of the CNS is only partly understood. This study developed a novel method to simultaneously record spinal cord electrophysiology (SCE) and electroencephalography (EEG) signals and validated its performance using a classical resting-state study design with two experimental conditions: eyes-closed (EC) and eyes-open (EO). We recruited nine postherpetic neuralgia patients implanted with a spinal cord stimulator, which was modified to record SCE signals simultaneously with EEG signals. For both EEG and SCE, similar differences were found in delta- and alpha-band oscillations between the EC and EO conditions, and the spectral power of these frequency bands was able to predict EC/EO behaviors. Moreover, causal connectivity analysis suggested a top-down regulation in delta-band oscillations from the brain to the spinal cord. Altogether, this study demonstrates the validity of simultaneous SCE-EEG recording and shows that the novel method is a valuable tool to investigate the brain-spinal interaction. With this method, we can better unite knowledge about the brain and the spinal cord for a deeper understanding of the functions of the whole CNS.

Susceptibility of brainstem to kindling and transfer to the forebrain

Purpose: The kindling of seizures with stimulation of brainstem sites has been reported inconsistently in the literature. The characteristics of the kindling observed, involving high intensities of stimulation and immediate onset of generalized tonic-clonic convulsions, raise questions regarding the nature of kindling from these sites.

Methods: We implanted chronic electrodes in either the nucleus reticularis pontis oralis (RPO), mesencephalic reticular formation (MRF), dorsal periaqueductal gray (dPAG), or ventrolateral periaqueductal gray (vlPAG) in male Long-Evans rats, with a recording electrode in the amygdala. Rats received conventional high-frequency kindling stimulation once daily for 30 days. To test for transfer, we kindled the amygdala beginning 7 weeks after the last brainstem kindling trial.

Results: Tonic-clonic seizures were evoked by stimulation from all brainstem sites. Seizures were brief and were associated with characteristic low-amplitude high-frequency afterdischarge (AD). Kindling of the dPAG resulted in the development of classic AD and increased AD duration. Prior kindling of the dPAG facilitated subsequent kindling of the amygdala however, no transfer was observed with prekindling of other brainstem sites.

Discussion: The variability in the response to kindling stimulation suggests that certain brainstem sites are resistant to kindling, whereas other sites are more susceptible to kindling but are still relatively resistant in comparison to sites in the forebrain. The development of classic AD in later trials of dPAG stimulation suggests that epileptogenesis can occur even in the initial absence of classic AD when low-amplitude high-frequency AD is present.

Author information

These authors contributed equally: Christopher Heelan, Jihun Lee.


School of Engineering, Brown University, Providence, RI, USA

Christopher Heelan, Jihun Lee, Ronan O’Shea, Laurie Lynch & Arto V. Nurmikko

Connexon Systems, Providence, RI, USA

Department of Surgery (Neurosurgery), Dalhousie University, Halifax, Nova Scotia, Canada

Department of Neuroscience, Brown University, Providence, RI, USA

Carney Institute for Brain Science, Brown University, Providence, RI, USA

Wilson Truccolo & Arto V. Nurmikko

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar

You can also search for this author in PubMed Google Scholar


A.N. and J.L. conceived the project. J.L. and A.N. designed the neural experimental concept. J.L. designed the NHP experiments. J.L. and L.L. executed the NHP experiments. C.H. designed and executed the neural decoding machine learning experiments. C.H. and R.O. developed the neural processing toolkit. C.H. performed the results analysis. W.T. provided the neurocomputational expertise and D.B. the surgical leadership. C.H., A.N, and J.L. wrote the paper. All authors commented on the paper. C.H. composed the Supplementary movies.

Corresponding authors