Stability of ECoG high gamma signals during speech and implications for a speech BCI system in an individual with ALS: a year-long longitudinal study

Abstract

Objective.

Speech brain-computer interfaces (BCIs) have the potential to augment communication in individuals with impaired speech due to muscle weakness, for example in ALS and other neurological disorders. However, to achieve long-term, reliable use of a speech BCI, it is essential for speech-related neural signal changes to be stable over long periods of time. Here we study, for the first time, the stability of speech-related electrocorticographic (ECoG) signals recorded from a chronically implanted ECoG BCI over a 12 month period.

Approach.

ECoG signals were recorded by an ECoG array implanted over the ventral sensorimotor cortex (vSMC) in a clinical trial participant with ALS. Because ECoG-based speech decoding has most often relied on broadband high gamma signal changes relative to baseline (non-speech) conditions, we studied longitudinal changes of high gamma band (HG) power at baseline and during speech, and we compared these with residual high frequency (HF) noise levels at baseline. Stability was further assessed by longitudinal measurements of signal-to-noise ratio (SNR), activation ratio (ActR), and peak speech-related HG response magnitude (HG response peaks). Lastly, we analyzed the stability of the event-related HG power changes (HG responses) for individual syllables at each electrode.

Main Results.

We found that speech-related ECoG signal responses were stable over a range of syllables activating different articulators for the first year after implantation.

Significance.

Together, our results indicate that ECoG can be a stable recording modality for long-term speech BCI systems for those living with severe paralysis.

Clinical Trial Information.

ClinicalTrials.gov, registration number NCT03567213.

1. Introduction

Speech impairments due to muscle weakness, without language or cognitive impairments, may occur with a variety of neurological disorders, including amyotrophic lateral sclerosis (ALS) (1). In the worst-case scenario, known as Locked In Syndrome (LIS) (2), patients may be unable to initiate conversations and may only be able to respond to closed questions with eye movements or blinks. Brain computer interfaces (BCIs) have the potential to augment or restore communication for these patients and allow them to control assistive technology using only their neural signals. In a survey of 28 Dutch people with LIS, 13 of whom had progressive neuromuscular diseases, the most preferred BCI application was direct personal communication through attempted speech (3).

Decoding speech from ECoG signals has most commonly relied on speech-related changes in broadband high gamma (HG) activity, with the strongest responses typically ranging between 60 and 200 Hz (4–9). HG activity is tightly correlated with changes in neural population firing rates, co-localizes with BOLD responses, and has been used effectively to measure cortical activation with high spatial and temporal precision (10, 11). Metzger et al. were able to decode speech from ECoG signals at a remarkable rate from a subject with vocal tract paralysis by combining HG activity with information from lower frequency signals (8, 9). While BCI devices that record neural signals through microelectrode arrays (MEAs) have achieved comparable speech decoding performance to ECoG BCIs, and are currently the most widely employed BCI recording mechanism (12, 13), ECoG has been investigated as a recording modality for BCI because it is associated with less tissue damage (14) and greater signal stability over time when compared to MEAs (15). In addition, ECoG provides robust physical artifact resistance, far greater spatial resolution, and increased signal amplitude and bandwidth when compared to electroencephalography (EEG) (16).

Reliable long-term decoding results from ECoG signals are a strong indicator of the stability of the underlying recording system (17, 18), but only a few studies have specifically focused on individuals with ALS (19–21). ECoG recordings from the upper limb area of the sensorimotor cortex (SMC) in an individual with late-stage ALS were stable for 36 months and allowed for several years of BCI home use (19). The work of Silversmith, et al. in plug-and-play cursor control, and Benabid, et al. in long-term exoskeleton control demonstrate the ability of upper limb ECoG BCIs to function over many days without the need for retraining (18, 22, 23). Luo et al. achieved stable speech BCI decoding for three months without recalibration using an implanted ECoG BCI (21). Mean bandpower remained relatively stable across both high and low frequencies in a study by Moly et al., where bimanual motor-based avatar and exoskeleton tasks were reliably carried out for six and four months, respectively, without recalibration (17). It has yet to be seen if these results will also hold true for recordings at speech-related SMC areas during rest and during speech tasks.

Other studies of long-term ECoG stability in humans have corroborated the results of BCI studies, in their analyses of SMC recordings at rest (19, 24, 25) and during upper limb motor tasks (19). In a 15-patient study by Nurse et al., individual electrodes tended to show more variability in responses over time, but when taken as a whole, spectral qualities and daily signal bandpowers were largely unchanged for up to two years, only suffering small increases in high frequency noise levels (25). However, another study by Sun, et al., of 121 epilepsy patients who were implanted with ECoG strip electrodes for two years found that data collected within the first 5 months of implantation was not fully representative of subsequent electrophysiology. This study observed significant changes in total bandpower and normalized power within frequency bands during the first 5 months, with the most prominent changes occurring in the HG frequency band, which then stabilized for the remainder of the study period (26).

In this study, we analyzed the stability of ECoG signals during periods of rest and attempted speech over a 12-month period, recorded from ECoG electrodes implanted over the ventral sensorimotor cortex (vSMC) in an individual with ALS. The speech task was run once at the beginning of each testing session. Average HG power was extracted for baseline (rest) and speech epochs, and average high frequency noise (HF) bandpower was extracted during the baseline periods in each session. From these frequency bands, other metrics of signal stability such as signal-to-noise ratios (SNR), HG activation ratios (ActR), and peak HG response magnitudes were calculated for each session over the course of the study. Lastly, HG response similarity was assessed by calculating the root mean squared error (RMSE) between subsets of trial-averaged event-related HG power changes. This allowed for analysis of spectral response stability. The stability of HG responses was also evaluated with offline decoding.

2. Methods

2.1. Participant and Ethics Statement

This study was performed in accordance with the Declaration of Helsinki. This study includes data from one clinical trial participant (ClinicalTrials.gov Identifier: NCT03567213) who gave written informed consent for the study. The neural implants used in this study were approved under an Investigational Device Exemption approved by the US Food and Drug Administration. The study protocol was approved by the Johns Hopkins Medicine Institutional Review Board (IRB00167247). The participant is a right-handed man who was diagnosed with ALS 8 years prior to his participation. He was 61 years old at the time of implant and had moderate to severe weakness of bulbar and upper extremity muscles. Residual phonation and articulation could still support overt speech, albeit with reduced speed and limited intelligibility. Evaluations by a speech-language pathologist at multiple points during and after the 12-month study concluded that the participant had mixed spastic-flaccid dysarthria, the nature and severity of which remained stable across timepoints. Mandibular strength and range of movement were also found to be preserved. A detailed report of the evaluations can be found in Supplementary Note S1.

2.2. CortiCom Investigational Device

The investigational CortiCom device used in this study was implanted in July 2022 without complications. Two 64-channel ECoG grids (PMT Corporation, Chanhassen, MN) were implanted on the pial surface of the brain (Figure 1c.–d.). Exact placement was informed by anatomical landmarks, preoperative fMRI, and somatosensory evoked potentials. The grids were positioned to cover areas of SMC important for speech (27–30) and upper limb control (31) according to previous ECoG mapping studies. Each ECoG grid had a surface area of 12.11 cm² (36.66 mm × 33.1 mm) with an 8 × 8 electrode configuration (center-to-center spacing of 4 mm) embedded in soft silastic sheets. Each platinum-iridium disc electrode had a thickness of 0.76 mm and an exposed surface diameter of 2 mm. We used one of two wires implanted on the top surface of the grids as references for ECoG signal amplification. One percutaneous pedestal connector connected to the ECoG grids was surgically anchored to the skull.

Figure 1:

Hardware setup and task overview. a.-b. Acquisition of neural signals during the syllable repetition task. The subject was presented with an auditory cue that they were instructed to immediately repeat to the best of their ability. c. Electrode coverage of the precentral gyrus (PreCG), postcentral gyrus (PostCG), and dorsal laryngeal speech areas. d. Neural signals were recorded by two 64-electrode ECoG grids placed over the SMC and passed through a NeuroPlexE headstage via a 128-channel percutaneous connector. Electrodes not analyzed in this study are grayed out. All other electrode colors are scaled by the trial-averaged HG response magnitude over the course of the study.

2.3. Task and data acquisition

Signals were recorded during a syllable repetition task in which a recording of one of 12 spoken syllables was played on a speaker and the participant repeated aloud each syllable (Figure 1a–b). Each syllable was repeated 5 times at each session, for a total of 60 trials per session, in a pseudorandom order that was repeated over the entire study period. The inter-stimulus interval was 4.5 seconds. The participant was tested up to three times per week, at the outset of each testing session. The consonant-vowel syllables used in this task were selected to cumulatively activate electrodes across the lower and upper grid that could potentially be used for speech decoding (Figure 1c) (28). Only the electrodes that covered areas that were reliably activated during speech were included in our analyses (Figure 1c–d).

2.4. Impedance testing

Impedances of all electrodes were measured once per week, on average, by the Impedance Tester in the Blackrock Central Software Suite. This test returns a specific impedance value only if an electrode has an impedance greater than 15 kOhms, so it was not possible to track lower impedances in detail over time.

2.5. Neural signal processing

Initial filtering (0.3 Hz – 7500 Hz), amplification, and digitization were done by a NeuroPlexE headstage (Blackrock Microsystems, Salt Lake City, UT) using a referential derivation in which all electrodes were referenced to one of the two reference wires mentioned in section 2.2. The same reference wire was used for all sessions. The headstage was attached to the implanted investigational CortiCom device via a 128-channel percutaneous connector. An HDMI cable connected the headstage to a Digital Hub (Blackrock Microsystems, Salt Lake City, UT), which sent signals to a Digital NeuroPort Biopotential Signal Processing System (Blackrock Microsystems, Salt Lake City, UT) via a fiberoptic cable, where they were downsampled to 1 kHz. Electrodes 19, 38, and 48 were excluded from analysis due to consistently high impedance (>15 kOhm) and electrode 52 was excluded after visual inspection of the raw signals, which revealed that this channel’s signals did not exhibit nominal patterns of ECoG signals on multiple occasions. In total, we investigated ECoG signal stability in 68 electrodes throughout the study period, 60 from the ventral array and 8 from the dorsal array. Five sessions were excluded from analysis due to technical problems with the recordings.

HG and HF frequency band limits were chosen according to trial-averaged spectrograms containing data from every 10th session (Supplementary Figure 1). The HF band was further confirmed as noise by the procedure described in Larzabal, et al (24). The average PSD slope for the 300–499 Hz frequency band at 61, 229, and 386 days after implantation was 0.0091%, 0.0063%, and 0.01%, respectively. All of these values are well below the reference’s HF PSD slopes, which range between 0.21% - 0.3%. HG (70–170 Hz) and HF (300–499 Hz) data from all electrodes used for analyses in this study was extracted by passing raw signals through an 8th order Butterworth IIR filter. Logarithmic power in the extracted frequency bands was computed in 50ms windows with 10ms overlap.

2.6. Neural signal analysis

For each session, trial-averaged HG (baseline and active) and HF (baseline) bandpower was calculated in dB for each electrode. Baseline was defined as one second immediately preceding stimulus onset, when the subject had completed the previous trial and was in a state of rest. The activation period was defined as the 3.5 seconds after stimulus onset, the period encompassing speech planning and active speech (32). Baseline and activation time periods were confirmed by the average HG response for each electrode over all trials (Supplementary Figure 2).

SNR and ActR were calculated from the bandpower data for each session according to the equations below, where P_HG and P_HF are the average values of the trial-averaged bandpower traces.

$$\mathit{\text{SNR}}=10{\text{log}}_{10}\frac{P_{HG\mathit{\text{baseline}}}}{P_{HF\mathit{\text{baseline}}}}$$ (eq. 1)

$$\mathit{\text{ActR}}=10{\text{log}}_{10}\frac{P_{HG\mathit{\text{active}}}}{P_{HG\mathit{\text{baseline}}}}$$ (eq. 2)

HG response peaks were calculated by finding the maximum value of the trial-averaged HG values for each session after being z-scored to a portion of the baseline signal (eq. 3).

$$H{G}_{\mathit{\text{peak}}}=\text{max}\left(\frac{{\text{log}}_{10}(P_{HG})-\overline{b}}{\mathit{\text{std}}(b)}\right),\text{b}=\text{baseline}\text{HG}\text{from}-0.8-0\text{s}$$ (eq. 3)

To assess HG response stability, HG power for each syllable was extracted as described in section 2.5 across both baseline and active periods, from −1 to 3.5 seconds with respect to onset of the auditory stimulus, and z-scored to the 0.8 seconds of the baseline signal before cue onset. Sessions were divided into subsets of 14 consecutive sessions, which allowed us to average 70 trials for each syllable. Each subset spanned an average of 6.2 weeks, and the average HG response waveform was calculated across trials from each subset. Each average response was compared to its subsequent group’s average response by calculating the root mean squared error (RMSE) for each channel during the active period.

2.7. Offline decoding

For full syllable decoding, two sets of training data were used, each containing 40 sessions worth of data from either the first 5 months of implantation or within 6–10 months post implantation, for a total of 200 repetitions per syllable. All 68 channels analyzed for stability in this study were used in decoding analyses. Linear discriminant analysis (LDA) and neighborhood components analysis (NCA) were applied to the event-related HG responses from each training set (0–3.5s after cue onset), extracted as described in section 2.5. All 50ms windows were input as individual features.

40-fold cross validation was performed on each training set and training method combination. HG responses from the last 20 sessions (341 to 393 days post implantation) were used as test data against each training set using k nearest neighbors (kNN) classification with k = 30, determined by calculating the decoding accuracy of the first session in the test set (341 days post implantation) for a range of k values between 1–60.

This process was repeated using only three of the twelve classes (“YAH”, “MOO”, and “GEE”), which were chosen to represent distinct consonant-vowel combinations. Separability of these three classes was initially determined visually, as the average HG event related responses for these syllables were distinct (Supplementary figures S9–S11), and this was confirmed by t-distributed stochastic neighbor embedding (tSNE) analyses after fitting the testing data to the LDA and NCA models that were trained on the 6–10 month training data (Supplementary figure S13).

Lastly, analyses of individual vowel and consonant sounds were conducted using the 6–10 month training set. The decoding process remained the same as above, with the stimuli now grouped either by consonant sounds (“B”, “G”, “H”, “L”, “M”, “TH”, “T”, “V”, “Y”, “J”, or “Z”) or vowel sounds (”OO”, “AH”, or “EE”).

2.8. Statistical testing

Linear regression was used to identify statistically significant trends in all metrics for whole grid averages and individual electrodes over time intervals spanning some or all of the 12-month period of this study (see below). Regression slope significance was calculated within the python linregress function by a Wald test, with the null hypothesis being that the slope of the regression line was equal to zero. The dependent variable was the specified metric, and the independent variable was days since implantation. In HG response stability analysis, RMSE values of 2 standardized units (z-scores) or higher were considered significant.

When we computed regression slopes over the course of 12 months, we observed an overall trend of increasing signal strength across most of our metrics (Table 1). However, the greatest increase in signal strength appeared to occur within the first 5 months after implantation (Figure 2). Prior long-term studies of signal strength from chronically implanted ECoG electrodes had also observed this early trend (26, 33), attributing this to short-term implant-related changes in the electrode-tissue interface. This raised concern that any long-term decrease in signal strength could potentially be obscured by an early increase. To avoid biasing our analysis of long-term signal stability, we computed regression lines for each signal strength metric not only across 12 months, but also across the first 5 months and across the remainder of the 12-month period.

Table 1.

Whole Grid Regression Slopes

Metric	0–12 Months Rate of Change (per day)	0–5 Months Rate of Change (per day)	6–12 Months Rate of Change (per day)
HG Active (dB)	0.00567	0.02503	0.00363
HG Baseline (dB)	0.00527	0.01840	0.00344
HF (dB)	0.00190	−0.00437	0.00402
SNR (dB)	0.00336	0.02277	−0.00059
ActR (dB)	0.00040	0.00664	0.00019
HG Response Peaks (z-score)	0.00023	0.00455	0.00010

Table 1 summarizes the whole grid (across all 68 electrodes used in this study) regression slopes for each metric, with bolded values representing slopes that are statistically significant from zero (least squares linear regression, Wald test, p < 0.05) as determined by a Wald test within the python linregress function. All metrics showed very small rates of change over the entire 12 months, with a decrease in the magnitude of their rate of change after the first 5 months.

Figure 2:

Session averages for all metrics. a.-b. Average active HG, baseline HG, and baseline HF bandpower, SNR, and ActR across all electrodes. Error bars are ± average standard deviation (SD). c. Peaks of average HG response waveforms.

2.9. Tuning fork experiments

To ensure no acoustic artifacts were present during attempted overt speech (34, 35), a tuning fork experiment was conducted as described by Wilson et al. (36). The fundamental frequency (F₀) of the participant’s voice was approximately 130Hz. A 128-Hz tuning fork was held next to the participant or gently pressed against the participant’s skull. In these tuning fork experiments, we observed no increase in energy near the 128-Hz tuning fork frequency in the ECoG channels used in this study, as detailed in a previous publication (21).

3. Results

3.1. Whole Grid Average Trends

We calculated the mean bandpower by session for each electrode (Supplementary Figures S3–S5) and across all electrodes (Figure 2a) and conducted linear regressions on the averages of all electrodes (Table 1). We then derived several metrics from the extracted bandpower information. Session means (Supplementary Figures S6–S8) and linear regressions (Table 1) for these metrics were calculated following the same procedure as HG and HF bandpower. We calculated SNR as a ratio of HG power to HF power (eq. 1). Similarly, we compared average HG power during speech with its average power at baseline (ActR) (eq.2) to estimate how discriminable active and resting states were. The whole grid session means for SNR and ActR are shown in Figure 2b.

The last metric we chose was the HG response peaks (Figure 2c), defined as the maximum value of the trial-averaged event-related HG trace for each session (eq. 3). Event-related HG traces were calculated by z-scoring the HG signal for each trial to the average of a portion of the baseline (0.8s preceding speech). This allowed us to determine whether the strength of speech-related HG responses changed over time relative to baseline. The whole grid session averages of all metrics appeared to show little change over time, with the average bandpower, SNR, and ActR SD within 2–3 dB for most sessions.

3.2. Individual Electrode Trends

The regression lines for each electrode for the 0–5 month and the 6–12 month periods can be seen in figure 3a. Figure 3b shows that there were fewer electrodes with regression slopes that were significantly different from zero in the 6–12 month period when compared to the 0–12 month period for all metrics and the 0–5 month period for all metrics except for HF noise, in which the number of significant channels increased by 18%. The number of significant channels decreased by 24%, 14%, 56%, 85%, and 80% for Active HG, Baseline HG, SNR, ActR, and HG response peaks, respectively.

Figure 3:

a. Regression lines for all metrics by electrode. Solid lines indicate significant trends (least squares linear regression, Wald test, p < 0.05) b. Number of channels with significant positive or negative regression slopes for each regression period.

3.3. Stability of HG Responses

Finally, we analyzed the stability of speech-related HG responses for different syllables. Trial-averaged HG responses were obtained for each group of sessions. Figure 4 shows examples of these waveforms along with the results of the RMSE analyses of the active period for all syllables. All channels returned RMSE values of less than 1.05 standard deviations throughout the course of the study.

Figure 4:

Similarity of average HG responses over time. a. Subset average HG response waveforms for the syllable ‘BAH’ for all electrodes in each anatomical area. b. Root mean squared error (RMSE) between subset average HG responses for each channel. Solid line represents the mean of all RMSE values.

3.3.1. Offline Decoding Stability Using HG Responses

Although this task was designed for signal stability monitoring rather than decoding, to confirm HG response stability, we performed offline decoding using various combinations of training data and training methods. Each combination of training data and training method was cross-validated with 40-fold cross-validation. For full-syllable decoding with the 0–5 month training set, the cross-validation accuracies were 20%, 32.5%, 75%, and 77.5% for the 12-class LDA, 12-class NCA, 3-class LDA, and 3-class NCA, respectively. Cross-validation of the 6–10 month training set returned accuracies of 27.5%, 40%, 97.5%, and 90% for the same models. Consonant sound cross-validation returned accuracies of 40% for LDA and 47.5% for NCA. Vowel sound cross-validation returned accuracies of 47.5% and 52.5% for the LDA and NCA, respectively.

The data from the last 20 sessions of the study were used as test sets. Both 12 class and 3 class decoding of full syllables returned classification accuracies well above their respective chance levels (Figure 5a). On average, changing the training set from 0–5 months to 6–10 months increased the 12-class classification accuracy by 4.5% and 0.1% when using LDA and NCA, respectively. The average increases for 3 class accuracy were 6.9% for LDA and 9.6% for NCA. We also analyzed the full syllable classification accuracy over time for both training sets (Figure 5c–d). No significant trends were found for 12 or 3 class decoding using either training method. Analyses of consonant and vowel sounds returned similar results, with decoding accuracies markedly above chance (Figure 5b) and no significant trends in accuracy over time (Figure 5e).

Figure 5:

Offline Decoding Results. a. Full syllable classification accuracy for all test files for each training method and training dataset combination. b. Consonant and vowel classification accuracy for all test files for the 6–10 month training set. c.-d. Full syllable classification accuracy over time for the 0–5 month (c.) and the 6–10 month (d.) training set. e. Consonant and vowel classification accuracy over time for the 6–10 month training set. Shaded regions in c.-e. represent 95% confidence intervals for the regression lines. Regression slopes were not significantly different from zero for these data (least squares linear regression, Wald test).

4. Discussion

We studied the stability of ECoG signals with particular emphasis on HG signal components most likely to be used for long-term speech BCIs. Several other studies of ECoG stability have shown stability in baseline (i.e., at rest) recordings during upper limb motor tasks. Here, we report evidence for the longitudinal stability of ECoG signals during speech tasks, including ECoG HG responses during periods of speech-related cortical activation (32).

When analyzing metrics across the entire set of electrodes, all metrics returned very gradual increases in response quality. There was an apparent decrease in regression slope magnitudes between the 0–5 month period and the 6–12 month period (Table 1). This was consistent with the observations of Sun et al., where increases in average total spectral power were observed during the first four months of implantation (26). In the current study, it was unclear how much the increasing trends during the first 5 months were influenced by hardware adjustments within the first 100 days. In the whole grid averages, there appeared to be a marked increase in HG signal power around 64 days after implantation (Figure 2), which was when an HDMI cable was changed, but upon further inspection, this jump in power was not observed in individual channels. Many channels showed consistently increasing HG power until approximately 150 days (5 months) after implantation (Supplementary Figures S3–S4).

The greater stability of the metrics that consisted of a ratio (SNR, ActR), or that were standardized to baseline (HG response peaks) suggest that underlying neural information can still be accurately recorded without distortion in cases where both metrics undergo a common change. Indeed, all of the bandpower metrics (HG active and baseline bandpower, HF baseline bandpower) increased at approximately the same rate from 6 months onwards, and this increase was mitigated when SNR, ActR, and HG response peaks were calculated. Offline decoding accuracy remained stable and above chance regardless of whether the training set was taken from the 0–5 month period or the 6–10 month period, so there is reason to believe that the small increases in HG and HF bandpower observed here were not detrimental to decoding neural responses. It is important to note that while our decoding accuracies might have been improved by the use of neural networks or additional preprocessing steps, such as normalization across the HG frequency range and more sophisticated feature selection, the main focus of this study was to monitor signal stability, so we have chosen to focus on decoding stability rather than achieving the highest possible accuracy.

Individual electrode trends gave additional insight into the whole grid trends, as all metrics except for HF noise bandpower saw a decrease in the number of electrodes with significant trends after the first 5 months (Figure 3b). We observed a small increase in HF noise over time in the electrodes that showed significant trends during the 6–12 month period. Since the increase in HF bandpower of these electrodes, which we have interpreted here as the noise floor, seems to have not affected the stability of the SNR during the first year of implantation, this is not immediately concerning, but should be monitored carefully over a longer period.

We did not observe significant changes, defined as an RMSE greater than 2 standard deviations, over time in the HG responses that were standardized to their respective baseline periods and averaged separately for each syllable (Figure 4b). This suggests that ECoG signal stability is robust for a wide range of articulators (27). Combined with the stability of the offline decoding accuracy observed here, this also demonstrates stability of the temporal envelopes of HG responses during speech, which is likely to have an important bearing on online speech decoding.

Electrodes in the ventral grid displayed more prominent speech related HG responses than those in the dorsal grid (Figure 2c, Supplementary Figures S9–S11) and appeared to follow a gradient that increased towards the posterior inferior corner of the ventral grid. Initially, this was thought to be an issue of electrode proximity to the reference, but the effect remained after applying common average referencing (Supplementary Figure S12). The ventral grid responses were expected to have higher amplitudes based on their anatomical location, as these electrodes covered areas commonly associated with articulation and phonation along ventral pre and post central gyri (37). The dorsal grid electrodes that were included in these analyses targeted the dorsal laryngeal area (38), which we identified by the physiological responses observed within the dorsal vocalic anatomical area previously defined by Chartier et al (29). The HG responses recorded by these electrodes were lower in amplitude.

It should be noted that the data presented here are only from one subject, and speech was limited to single syllables. Further studies are needed across a wider range of subjects before definitive conclusions can be made. Additionally, it would be beneficial to track these metrics for longer than 12 months to see whether the trends for speech tasks mirror those for motor tasks beyond the first year of implantation. However, the stability observed in this study, especially that of HG responses, shows that ECoG is a promising recording method for long-term speech BCI systems.

5. Conclusions

These results support ECoG as a promising option for long term, stable use of BCIs for patients with speech impairments due to ALS. Here we found that after an initial increase in signal strength in the first 5 months, HG responses were stable until at least 12 months post-implantation. Importantly, we did not observe any indicators of a significant decrease in signal strength or stability.

Data Availability Statement

The data that support the findings of this study are available from the corresponding author upon reasonable request. The data are not publicly available due to privacy or ethical restrictions.