Moving beyond the motor cortex: a brain-wide evaluation of target locations for intracranial speech neuroprostheses

Abstract

Speech is the fastest and most natural form of communication, which can be impaired in certain disorders. Speech brain-computer interfaces (BCIs) offer a solution by decoding brain activity into speech. Current neuroprosthetic devices focus on the motor cortex, which might not be usable in all patient populations. Fortunately, many other brain regions have been associated with the speech production process. Here, we investigate which regions are potential (alternative) targets for a speech BCI across a brain-wide distribution within a single study. The distribution includes sulci and subcortical areas, sampled with both a high temporal and a high spatial resolution. Thirty participants were recorded with intracranial electroencephalography during speech production, resulting in 3249 recorded contacts across the brain. We trained machine learning models to continuously predict speech from a brain-wide global to a single-channel local scale. Within each scale we examined a variation of selected electrode contacts based on anatomical features within participants. We found significant speech detection in both gray and white matter tissue, no significant difference between gyri and sulci at any of the analysis scales and limited contribution from subcortical areas. The best potential targets in terms of decoding accuracy and consistency are located within the depth of and surrounding the lateral fissure bilaterally, such as the (sub)central sulcus, transverse temporal gyrus (Heschls' gyrus), the supramarginal cortex and parts of the insula. These results highlight the potential benefits of extending beyond the motor cortex and reaching the sulcal depth for speech neuroprostheses.

Full text

Moving beyond the motor cortex: a brain-wide
evaluation of target locations for intracranial speech
neuroprostheses
Maxime Verwoert1,* , Maarten C. Ottenhoff1, Simon Tousseyn1,2, Johannes P. van Dijk2,3,4,
Pieter L. Kubben1,5, and Christian Herff1,*
1Department of Neurosurgery, Mental Health and Neuroscience Research Institute, Maastricht, The Netherlands
2
Academic Center for Epileptology, Kempenhaeghe/Maastricht University Medical Center, Heeze, The Netherlands
3Department of Orthodontics, Ulm University, Ulm, Germany
4Department of Electrical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands
5Academic Center for Epileptology, Kempenhaeghe/Maastricht University Medical Center, Maastricht, The
Netherlands
*corresponding author(s): Christian Herff (c.herff@maastrichuniversity.nl), Maxime Verwoert
(m.verwoert@maastrichtuniversity.nl)
ABSTRACT
Speech is the fastest and most natural form of communication, which can be impaired in certain disorders. Speech brain-
computer interfaces (BCIs) offer a solution by decoding brain activity into speech. Current neuroprosthetic devices focus on the
motor cortex, which might not be usable in all patient populations. Fortunately, many other brain regions have been associated
with the speech production process. Here, we investigate which regions are potential (alternative) targets for a speech BCI
across a brain-wide distribution within a single study. The distribution includes sulci and subcortical areas, sampled with both a
high temporal and a high spatial resolution. Thirty participants were recorded with intracranial electroencephalography during
speech production, resulting in 3249 recorded contacts across the brain. We trained machine learning models to continuously
predict speech from a brain-wide global to a single-channel local scale. Within each scale we examined a variation of selected
electrode contacts based on anatomical features within participants. We found significant speech detection in both gray and
white matter tissue, no significant difference between gyri and sulci at any of the analysis scales and limited contribution from
subcortical areas. The best potential targets in terms of decoding accuracy and consistency are located within the depth of
and surrounding the lateral fissure bilaterally, such as the (sub)central sulcus, transverse temporal gyrus (Heschls’ gyrus), the
supramarginal cortex and parts of the insula. These results highlight the potential benefits of extending beyond the motor
cortex and reaching the sulcal depth for speech neuroprostheses.
Introduction
Decoding neural signals to restore natural and efficient communication for individuals with speech impairments is a key
objective of speech brain-computer interface (BCI) research
1
. Speech production is a highly complex process involving an
extensive and interconnected network of brain regions
2,3
. Yet, the current state-of-the-art speech BCIs are primarily focused on
decoding signals from the sensorimotor cortex
4–9
. While this region has shown promise in enabling communication for patients
with dysarthria or anarthria due to amyotrophic lateral sclerosis (ALS) or a brain-stem stroke, it may not cover the full intent of
the user
10
nor translate to individuals with different diseases or disease progressions. This is especially the case for individuals
with damage to the motor cortex, which may happen due to a stroke or degeneration of cortical motor neurons in ALS11.
Given the extensive network of brain areas contributing to produce speech, it is essential to consider alternative regions across
the entire brain. Existing non-invasive techniques that can capture signals across the brain, such as functional magnetic resonance
imaging (fMRI), scalp electro-encephalography (EEG), and magneto-enecephalography (MEG), lack the simultaneous temporal
and spatial resolution necessary to capture the rapid and complex dynamics of natural speech production for a continuous BCI.
The intracranial technologies electro-corticography (ECoG) and micro-electrode arrays (MEAs) provide high temporal-spatial
fidelity and have demonstrated success in speech BCIs
4–6,9
. However, both cover only the brain’s outer surface in typically a
few specific cortical regions (i.e., the sensorimotor cortex). While research with these technologies provide crucial insights,
these tools cannot access deeper areas of the brain—such as the sulci, inter-hemispheric regions, and subcortical nuclei.
Stereo-electroencephalography (sEEG), in contrast, offers an opportunity to probe these deeper areas of the brain. Research
utilizing sEEG has revealed involvement of subcortical structures in speech production
12,13
, as well as cortical areas nestled
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 3, 2024. ; https://doi.org/10.1101/2024.11.29.626019doi: bioRxiv preprint

within the sulci that are beyond the reach of surface-based recording technologies
14,15
. An added advantage of sEEG is the
typical broad placement of electrodes across multiple regions, often bilaterally, within a single individual
16
, enabling a more
diverse survey of brain regions implicated in speech. Although sEEG sampling can be spatially sparse, by aggregating data
across multiple participants with varying electrode placements, we can achieve a brain-wide investigation, as we demonstrate in
the current study.
Different representations of speech (e.g., articulatory, acoustic or semantic components) can be decoded from varying
locations
17,18
, which may be beneficial for individuals with diverse brain injuries affecting speech. For instance, in motor-based
impairments like apraxia of speech, where motor planning is disrupted
19
, BCIs may benefit from targeting areas associated
with higher-level phonological or acoustic representations, such as the posterior superior temporal cortex
20–22
. In this study, we
apply a continuous speech/silence decoder to 3249 widely distributed recording sites, covering gyri, sulci and deeper structures
to investigate potential targets for speech neuroprostheses.
Methods
Participants
We recorded data from 30 individuals with medication-resistant epilepsy in the current study (Table 1). The data from the first
10 participants (P01 to P10) have previously been published
23
. All participants joined the study voluntarily and signed written
informed consent. The study has been approved by the Institutional Review Boards of both Maastricht Unviversity (METC
2018-0451) and Epilepsy Center Kempenhaeghe. Participants were native Dutch speakers and were implanted with sEEG
electrodes (Fig. 1A) to localize the epileptic onset zone. Electrode placement was solely determined based on clinical needs.
Table 1. Demographic information, the performed task(s), the total number of channels and shafts that were recorded and the total number
of channels available in the electrode selection categories for each participant. M = male; F = female; UW = unique words, RW = repeated
words, US = unique sentences, GM = gray matter, WM = white matter, SC = subcortical, GS = gyrus-sulcus.
Participant Age Sex Task Channels Shafts GM WM Cortical SC Gyrus Sulcus GS
P01 20 F UW 127 12 67 53 48 1 16 24 8
P02 43 M UW 127 19 24 101 17 13 6 11 -
P03 24 M UW 127 15 74 50 53 10 30 17 6
P04 46 F UW 115 12 72 33 56 1 25 25 6
P05 50 F UW 60 5 34 25 17 16 8 9 -
P06 16 M UW 127 14 90 26 82 - 39 34 9
P07 47 M UW 127 11 62 62 49 10 28 16 5
P08 22 F UW 54 5 13 41 10 - 3 7 -
P09 20 F UW 117 11 72 41 57 2 29 21 7
P10 36 M UW 122 13 60 46 54 6 28 22 4
P11 36 M RW 127 15 79 44 62 9 31 22 9
P12 21 M RW 57 6 21 34 19 - 8 8 3
P13 36 F RW 94 10 46 42 32 10 18 14 -
P14 31 F RW 127 13 72 54 51 9 23 15 13
P15 34 M RW 115 11 70 43 65 - 20 26 19
P16 36 F RW 90 9 26 63 20 7 13 7 -
P17 60 M RW 65 5 35 30 15 14 3 10 2
P18 18 F RW 127 10 66 60 53 - 17 19 17
P19 60 F RW 121 9 60 58 39 10 22 11 6
P20 25 F RW 127 14 82 45 71 - 21 28 22
P21 20 F US 107 11 68 38 44 5 24 20 -
P22 59 M US 111 9 44 63 34 9 17 16 1
P23 40 M US 127 12 63 61 47 5 13 33 1
P24 37 F US 127 14 64 62 54 3 23 26 5
P25 60 M US 115 9 58 56 56 2 21 15 20
P26 61 F US 87 7 38 49 30 7 7 17 6
P27 25 F US 127 9 87 40 80 1 37 15 28
P28 28 M US 87 8 50 36 38 - 17 9 12
P29 24 M US 81 8 34 46 25 6 8 17 -
P30 33 M US 127 12 66 61 46 7 20 18 8
2/13
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 3, 2024. ; https://doi.org/10.1101/2024.11.29.626019doi: bioRxiv preprint

Tasks
Participants performed one of three speech production tasks in which they spoke 100 Dutch words or sentences out loud. In the
SWPD dataset
24
(P01 to P10), unique words (UW) were presented for two seconds. The next set of participants (P11 to P20)
spoke 20 words, in random order, that were chosen to be relevant for patients in locked-in state
25
. The block of 20 words was
repeated 5 times (RW). The last set of participants (P21 to P30) were presented with unique sentences (US). Since the reading
speed varies between people, the first five participants of this set (P21 to P26) chose beforehand how much time they would like
for reading each sentence. For the last five participants we modified the task design to allow the experimenter to press a key to
end the speech trial once the participant finished reading the sentence. Across all tasks, each speech trial was followed by a
1-second rest period.
Data acquisition
The neural data was recorded with one or two Micromed SD LTM amplifiers (Micromed S.p.A., Treviso, Italy) of 64 channels
each, with a sampling rate of either 1024 Hz or 2048 Hz. The 2048 Hz signal was subsequently downsampled to 1024 Hz. The
sEEG electrode shafts (Microdeep intracerebral electrodes; Dixi Medical, Beçanson, France) had a diameter of 0.8 mm, contact
length of 2 mm and an inter-contact distance of 1.5 mm. The number of contacts on an electrode shaft could vary between 5
and 18. The number of implanted shafts varied between 5 and 19. The audio data was recorded with the microphone of the
recording laptop (HP Probook) with a sampling rate of 48 kHz. LabStreamingLayer
26
and our in-house software T-Rex
27
was
used to synchronize neural, audio and stimulus data.
Electrode localization
The electrodes were localized for each participant with the img_pipe
28
Python package and Freesurfer. The pipeline included
the co-registration between a pre-implantation T1-weighted anatomical magnetic resonance imaging (MRI) scan and a post-
implantation computerized tomography (CT) scan, manual identification of two contacts from each electrode, a linear inter-
and/or extrapolation of the contacts accounting for the inter-electrode distance, a manual inspection of each contacts location
and a non-linear warping to an average brain (MNI152) for visualisations only. The native MRI scan was segmented using the
Fischl atlas
29
to extract subcortical labels and parcellated using the Destrieux atlas
30
to extract cortical labels for each contact.
There were 3807 implanted electrode contacts of which 3249 were recorded. From here on, we only focus on the recorded
contacts. There were 140 unique labels, 1419 had the ’Cerebral White Matter’ label (43.67%) and 318 the ’Unknown’ label
(9.79%). The ’Unknown’ label was given to contacts outside of brain tissue. The remaining 46.54% of contacts were spread
across a variety of locations, including cortical and subcortical regions.
Re-referencing
The data is recorded with a white matter reference (WMR), a contact located in white matter that is hand-picked by the treating
neurologist as a suitable reference contact. It is usually a channel with a low amplitude signal that does not show epileptic
activity. However, the use of one single contact for a reference can bias the signal in other contacts, solely due to spatial
differences. Moreover, the reference contact would need to be located on the implantation shaft itself for a simulation of the
performance of a single shaft. Thus, we applied an electrode shaft re-reference (ESR) by subtracting the average signal from all
contacts within the same shaft from each channel. This is similar to the commonly used common average reference, except that
it is restricted to the same electrode shaft rather than all implanted contacts across the brain. It is noteworthy that, in most cases,
there were more electrodes implanted than recorded due to a limited number of channels in the amplifiers. Thus, the neurologist
had to select which contacts to record data from. Sometimes this lead to missing contacts within a shaft of electrodes.
Signal processing
The neural signals were filtered to the broadband high-frequency activity (70-170 Hz) with an IIR bandpass filter (filter order 4)
and the Hilbert envelope was extracted. Two IIR bandstop filters (filter order 4) were applied to attenuate the first two harmonics
of the 50 Hz line noise. The filters were employed forward and backward to eliminate a potential phase-shift. The envelope
was averaged in 50 ms windows with a 10 ms frameshift as the neural features. The audio signal was downsampled to 16
kHz and the Short-Term-Fourier-Transform was applied to extract audio features, also in 50 windows with a 10 ms frameshift.
This procedure ensured an alignment between the labels and the neural features. An energy threshold was calculated with the
maximum plus minimum value of the average energy across spectral bins, multiplied by a static value of 0.45. This threshold
was applied to the audio features to extract ’speech’ and ’silence’ labels.
Analysis scales
The data was analyzed at three different scales. In the ’global’ scale, all recorded electrode contacts within an individual were
included as features in the model. In the ’shaft’ scale, the analysis was restricted to all contacts within a shaft. Finally, in the
’local’ scale, each contact was analyzed individually (Fig. 1B).
3/13
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 3, 2024. ; https://doi.org/10.1101/2024.11.29.626019doi: bioRxiv preprint

Electrode selection categories
For each of the analysis scales, we additionally selectively included contacts located in specific tissue types or structures in
separate analyses to determine the contribution of each. Due to a margin of error in electrode labeling, we utilized the proximal
tissue density (PTD) for each contact to determine if it was located in ’white matter’ or ’gray matter’. The PTD score indicates
the relative proportion of gray versus white matter in the 26 MRI voxels surrounding the center of the contact
31
, using the
following formula:
PT D =nGM −nWM
nGM +nWM
where
nGM
and
nWM
represent the number of gray matter and white matter voxels, respectively. We included subcortical
areas and labeled these as gray matter. For the white matter analysis, we only included contacts with
PT D <0
. For the gray
matter analysis, we only included contacts with
PT D >0
. The PTD could not be calculated for some contacts as they were
surrounded entirely by non-brain tissue, these were excluded in both analyses.
Next, we used the labels from the atlases to further sort the contacts into the categories ’cortical’, ’subcortical’, ’gyrus’,
’sulcus’ and ’gyrus-sulcus’. The ’cortical’ category refers to those contacts labeled by the atlas as gray matter. However, these
do not include the subcortical structures as they do in the general ’gray matter’ analysis. In this case, the subcortical structures
(e.g., the hippocampus) are labeled as ’subcortical’. The ’gyrus’ category refers to contacts located near the cortical surface
(e.g., the precentral gyrus), whereas the ’sulcus’ refers to contacts located deeper within a cortical fold (e.g., the central sulcus)
which could not be reached with a standard electrode grid. The ’gyrus-sulcus’ category refers to contacts located on the border
between a gyrus and sulcus, which could not be clearly separated and therefore have this combined label in the atlas (e.g., the
subcentral gyrus and sulcus). Note that some regions within the lateral fissure are gyral, even when they cannot be reached with
electrode grids (e.g., the short insular gyrus).
Speech detection
Neural features were enhanced with causal temporal information from non-overlapping windows up to 200 ms into the past.
This means that for each channel, there were 4 additional neural features with a shifted time alignment to the labels. Only
temporal windows from the past were used, as future windows could not be used in a real-time BCI. The data was split into a
10-fold cross-validation. Neural features were normalized to zero mean and unit variance using the mean and standard deviation
of the training data. The same normalization was then applied to the testing data.
We employed a continuous binary classification approach to distinguish between speech and silence. This simple framework
facilitates analysis across tasks and serves as a foundation for a basic BCI. The goal was not to optimize classification
performance but rather to identify potential brain areas involved in speech production. For this, we utilized a Linear Discriminant
Analysis (LDA) classifier. Given the class imbalances, with generally more silence than speech, we used balanced accuracy as
the primary outcome metric to account for this disparity.
This analysis was computed for each of the electrode selection categories described in the previous section. Additionally,
the single channel analysis was repeated for each individual temporal window from -500 ms up until +500 ms to investigate
temporal dynamics.
Significance threshold
We computed a distribution of chance level accuracy scores by randomly shuffling the speech/silence labels, calculating the
balanced accuracy between the original and shuffled labels in 10 folds similar to the speech detection pipeline, and repeating this
procedure 1,000 times for each participant. Next, we calculated the 99th percentile (
α
= 0.01) within each fold per participant
and set the individual threshold at the maximum across folds. As the final significance threshold, we took the maximum score
across participants. This procedure led to a strict threshold of 52.44%, well above the theoretical chance level of 50%.
Regions-of-interest
Due to variations in sampling within and between regions as they are labeled by the atlases, we grouped results together in 9
larger regions-of-interest (ROIs) across the brain. A detailed list of the anatomical regions from the atlases included in each ROI
is provided in the supplementary material (Table S1). Note that only the subcortical regions that were sampled were included in
the list.
Results
In this work, we recorded speech production sEEG data from 30 participants with electrode contacts covering a wide range of
locations (Fig. 1A) and structures/tissue types (Fig. 1B). We used a binary classification approach to decode the recorded neural
activity into speech or silence labels (Fig. 1C). Electrodes were analyzed at three levels (global, shaft and local), with additional
analyses discriminating tissue types and brain structures (Fig. 1B).
4/13
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 3, 2024. ; https://doi.org/10.1101/2024.11.29.626019doi: bioRxiv preprint

Figure 1. Graphical overview of the methods. A) Locations of recorded electrode contacts in an averaged brain from a top view. Colors
differentiate the 30 participants. B) Graphical illustration of the analysis scales and electrode selection categories within a coronal brain slice.
The ’all’ category is not displayed as it includes all contacts and no specific selection. C) Example of how the spectrogram of the recording
speech audio was used to extract the ’speech’ and ’silence’ labels used for the decoders.
Global to local scale
Speech can be detected well above chance level for all 30 individuals (ranging between 60.13% and 93.79%) at the global scale
using all recorded contacts (Fig. 2A). This holds true when only including contacts in gray matter as well. The accuracy of the
results dropped slightly for certain individuals in the other categories and the performance goes down to below significance
threshold for most individuals using only subcortical contacts. It is important to note that the lowest performing categories,
subcortical and gyrus-sulcus, also had a much lower number of channels within individuals compared to the other categories
(Table 1). Independent samples t-test were performed to compare certain electrode selection categories with one another. Gray
matter only was not significantly different from white matter only (t(58) = 1.47, p= 0.15), nor was the sulcus significantly
lower than the gyrus category (t(58) = 0.67, p= 0.51). However, the cortical category was higher than the subcortical one (t(51)
= 9.20, p< 0.001) and both the gyrus (t(51) = 3.25, p= 0.002) and the sulcus (t(51) = 2.49, p= 0.016) scored higher than
the gyrus-sulcus category. The variation in performance between individuals could be related to the sheer amount of features
included, as there was a significant correlation (Pearson’s r(58) = 0.43, p= 0.018) between the amount of features and accuracy
(Fig. 2B), and/or the specific anatomical regions that were sampled.
At the shaft scale, there are large variations between individual shaft results with the median accuracy being below the
significance threshold in all of the categories (Fig. 2C). However, the maximum scores (89, 89, 87, 88, 59, 88, 87 and 86%) are
not much lower than in the global scale (94, 94, 89, 93, 61, 89, 88 and 88%). At this level, between shafts overall, there was a
significant difference between the gray and white matter categories (t(602) = 2.25, p= 0.025) and between the cortical and
subcortical ones (t(352) = 3.00, p= 0.003). There was no significant difference between the gyrus and sulcus (t(464) = 0.23, p
= 0.82), nor either with the gyrus-sulcus category (gyrus: t(329) = 0.29, p= 0.77; sulcus: t(295) = 0.12, p= 0.91). At the shaft
level, there was not a significant correlation (Pearson’s r(634) = 0.10, p= 0.07) between the size of the shaft and accuracy
(Fig. 2D), suggesting that the anatomical location of the shaft may be a more important factor than the sheer amount of features
included.
In the local scale, the results are similar those at the shaft scale, with large variations between channels within categories
and most channels below the significance threshold (Fig. 2E). However, the maximum scores of single channels (88, 88, 87.
86, 59, 86, 86 and 79%) are still similar to those in the shaft and global scales. In these large groups, there are no statistically
significant differences between gray matter and white matter (t(3170) = 1.95, p= 0.05), between the gyrus and sulcus (t(1105)
= 0.28, p= 0.78), the gyrus versus gyrus-sulcus (t(790) = 1.22, p= 0.22), nor the sulcus versus gyrus-sulcus (t(747) = 1.00, p=
0.32). The subcortical category is significantly lower than the cortical category (t(1485) = 3.06, p= 0.002). There was a strong
correlation (Pearson’s r(634) = 0.95, p< 0.001) between the maximum accuracy of a single channel within a shaft and the
accuracy of the complete shaft (Fig. 2F), suggesting that one or a few channels within a shaft are driving most of the results.
Anatomical contributions
We zoom into the local scale as particular anatomical regions may be more important than the generic categories presented
above. Since the global and shaft scales cover multiple regions at once, we can only compare regions at the local channel level.
In Figure 3A we see the classification accuracy distribution across all significant channels. This figure indicates that the best
5/13
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 3, 2024. ; https://doi.org/10.1101/2024.11.29.626019doi: bioRxiv preprint

Figure 2. Distribution of balanced accuracy results within electrode selection categories across the analysis scales and correlations. A)
Global scale balanced accuracy results. B) Correlation between global accuracy and number of features. C) Shaft scale balanced accuracy
results. D) Correlation between shaft accuracy and shaft size. E) Local scale balanced accuracy results. F) Correlation between maximum
channel accuracy within a shaft and accuracy of the complete shaft. In A-C-E the individual dots represent the averaged balanced accuracy
per individual, shaft or channel, respectively. The box plots show the median and inter-quartile range, whiskers show the maximum range
excluding outliers. The dashed gray line represents the significance threshold in A-C-E and the best-fit linear trend, calculated using a
least-squares linear regression, in B-D-F. *p< 0.05, **p< 0.01, ***p< 0.001.
decoding results are clustered within and around the lateral fissure, with a more focal clustering in the right hemisphere.
Out of the 137 sampled anatomical regions (excluding white matter and unknown labels), we further focus on the 23 regions
that were sampled by at least 5 different participants and had at least 3 significant channels. The right anterior transverse
temporal gyrus (Heschl’s gyrus) scored the highest in terms of accuracy (Fig. 3B) and second in terms of consistency (the
percentage of significant channels within that region, Fig. 3C), the left Heschl’s gyrus was not sampled enough. The right
inferior circular insula scored the second highest in accuracy and third in consistency, while only at the 9th and 14th place in
the left hemisphere. The left central sulcus and subcentral gyrus-sulcus (GS) both scored high in accuracy and consistency,
with over 60% significant channels in the subcentral GS. These regions, along with the precentral gyrus, were not sampled by
enough participants in the right hemisphere. However, the left precentral gyrus was, and only had 2/12 significant contacts.
When we look at the larger regions-of-interests (ROIs; Fig. 3D), including regions not sampled enough individually, we see that
generally the motor cortex was equally involved between the hemispheres.
We further see bilateral regions surrounding the ones described above, such as the superior and middle temporal cortex, the
supramarginal cortex, the inferior frontal gyrus and other parts of the insula. The most consistent results overall come from the
auditory, motor and supramarginal regions (Fig. 3D). Note that parts of the superior temporal cortex are incorporated in the
auditory cortex ROI, the remaining parts of the relative large superior temporal cortex ROI are therefore not very consistently
involved. The insular ROI follows, as it was particularly the inferior part that scored high in accuracy and consistency. It is
worth noting that, while we look at the relative number of significant channels, the channels were not sampled exactly uniformly
within regions and between hemispheres, which limits the interpretability of comparisons.
Temporal dynamics
We explored how the spatial distribution and number of significant channels varied between temporally misaligned features
and tasks (Fig. 4). The results were grouped in five larger time segments (Fig. 4A-E), representing time windows well before
alignment (-500 until -300 ms), just prior to alignment (-300 until -100 ms), surrounding alignment (-100 until +100 ms), just
after alignment (+100 until +300 ms) and well after alignment (+300 until +500 ms). Neural features just prior to alignment
resulted in the largest number of channels significantly detecting speech (Fig. 4F), with a steeper drop-off after than before
alignment. The variation between tasks reflects the variation in the length of the speech trials (short words versus long
sentences). All three tasks engaged peri-sylvian, temporal and sensorimotor areas, while only the sentences (US) task engaged
frontal areas.
6/13
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 3, 2024. ; https://doi.org/10.1101/2024.11.29.626019doi: bioRxiv preprint

Figure 3. Anatomical dissemination at the local scale. A) Spatial distribution of significant balanced accuracy scores across channels in an
averaged brain. The colorbar ranges from the significance threshold (52%) to
≥80%
. Smaller gray dots represent insignificant channels. B)
Distribution of balanced accuracy scores within anatomical regions, sorted by average accuracies. The color of the bar indicates the
hemisphere. C) Percentage of significant channels within anatomical regions, sorted by percentage. The text above the bar describes the
amount of significant and total amount of channels within that region. The color of the bar indicates the hemisphere. D) Percentage of
significant channels within larger regions-of-interest within each hemisphere, sorted by the average percentage between the two hemispheres.
G = gyrus, S = sulcus, GS = gyrus-sulcus.
7/13
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 3, 2024. ; https://doi.org/10.1101/2024.11.29.626019doi: bioRxiv preprint

Figure 4. Temporal dynamics. A-E) Significant channel locations with colors representing the three tasks, divided into five time segments.
Smaller gray dots represent non-significant channels. F) Number of significant channels across time-frames from all (brain-wide) contacts.
UW = unique words task, RW = repeated words task, US = unique sentences task.
Discussion
Speech detection was performed at three scales of analysis and eight electrode selection categories. Speech was reliably detected
for all thirty individuals using all available data within individuals (global scale), but the performance varied greatly between
electrode shafts (shaft scale) and individual channels (local scale). While at a global scale the performance was correlated with
the amount of included channels, this was not the case at the shaft scale, but instead correlated strongly with the accuracy of the
best individual channel along the shaft. These results combined suggest that the location of individuals channels is important,
while a combination of multiple distributed regions may be beneficial as well. Previous research has indeed showed that overtly
and covertly produced speech can be reconstructed well from distributed regions in offline32,33 and in real-time33 decoding.
Overall, the gray matter category scored slightly higher than the white matter category at the shaft scale. At the global
scale the difference was not significant likely due to difference in the number of data-points. Whilst lower than gray matter,
the nonetheless high accuracy results from contacts located in white matter indicates that we may be able to decode valuable
information directly from white matter tracts. This activity may reflect volume conduction from a combination of nearby and
distant gray matter regions31. For example, direct electrical stimulation of the anterior arcuate fasciculus tract, superior to the
insula, has been found to induce thoughts or alter the conscious awareness of thoughts, similar to stimulation of the posterior
parietal cortex itself
34
. Our results are in line with previous decoding studies with sEEG that have also found contributions
from white matter contacts
15,35,36
. However, in our work, these findings could, at least in part, have been due to spreading of
information from cortical contacts within the same electrode shaft due to re-referencing. It is also a challenge to locate the
exact white matter tracts, requiring additional anatomical analyses and/or diffusion-tensor imaging, and therefore beyond the
scope of our current work.
There were clear differences between the cortical and subcortical categories in all analysis scales. While there were a few
significant channels in subcortical areas, the accuracy did not reach nearly as high as in cortical and even white matter regions.
Previous studies have found language-related functions in the hippocampus
12,13,37,38
and basal ganglia
39
and other subcortical
areas through a large meta-analysis
40
. Studies reveal specialized functions of these areas, such as semantic integration
37
and
syntactic processing
39
, and therefore may be too subtle for general speech detection as in the current study. Alternatively, in
electrophysiology, we may need to investigate these areas with lower frequencies37 or even single cell recordings13,41 .
Within the cortical channels, there was no statistically significant difference between the gyrus and sulcus categories at any
of the analysis scales across the brain, despite general differences in cytoarchitecture and function
42
. Even within the local
scale, only looking at significant contacts, we see both gyral and sulcus regions detecting speech. However, the gyri are more
8/13
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 3, 2024. ; https://doi.org/10.1101/2024.11.29.626019doi: bioRxiv preprint

often mentioned in the speech decoding literature as sulci cannot be sampled with the commonly used electrocorticography
grids. Note that, in this work, gyri do not necessarily mean they can be sampled with electrode grids on the surface of the brain,
as they can be located within the lateral fissure. A previous sEEG study did analyze differences in speech detection accuracy
across the level of depth along shafts and found generally higher accuracies at the surface level with overt and mouthed speech,
but not with imagined speech
15
. These results suggest that the deeper regions may be just as important for an imaginary-based
speech neuroprosthesis.
Next, we dived further into significant anatomical regions at the local scale. The overall spatial distribution across the
brain corresponds with the core language network
2,3
and its bilaterality is consistent with previous sEEG speech production
studies
15,17,18,22
. At the top positions in terms of both accuracy and consistency are Heschl’s gyrus (the auditory cortex), the
inferior insular sulcus and the (sub)central sulcus. While named gyrus, Heschl’s gyrus is also nestled deeper within the lateral
fissure. Interestingly, the precentral gyrus (primary motor cortex) was sampled plenty in the left hemisphere, but only showed
2/12 significant contacts. This is in stark contrast with the sulcal parts of the motor cortex, suggesting that reaching the sulcal
depth may be beneficial for motor-based BCIs43.
We further looked at larger regions-of-interest to generate a better overview across regions, including the individual areas
that were left out of Figure 3B and C due to not being sampled by plenty of different participants or not containing plenty
of significant contacts. Whilst sampling of the motor cortex was not ideal, the auditory areas of the superior temporal lobe
showed the most consistent results, despite known suppression of regions within the auditory cortex during self-generated
speech
44
. The supramarginal cortex may also be a good alternative for a speech neuroprosthesis, as was previously shown with
a micro-electrode array even for imagined speech
45
. There was limited contribution from the inferior frontal cortex, which
may be due to the reading task not requiring spontaneous speech planning
46
. The reading could also explain the relatively
stronger occipital than frontal involvement overall. However, when we split the results between the three tasks for temporal
dynamics, we do note that only the most complex (sentences) task reached significance in the frontal areas, likely engaging
more higher-order executive control47 than single word reading.
As most of the significant regions are also known to be activated during speech or sound perception
3,48
, the question is
whether they will remain activated without or with a limited amount of overt speech. Previous research has indeed found
similar, albeit more attenuated, activation patterns with purely imaged speech, compared to overt or mouthed speech
15
. This
even applies to auditory regions
15,49,50
as they may be involved in the internal representation of speech
49
. However, responses
to perceived speech would be another problem for a practical speech BCI. This is even the case for a speech neuroprosthesis on
the motor cortex, with the potential of leading to false positives through speech perception
51,52
or even through unintentional
inner speech
52
. Ensuring executive control is an important ethical aspect in the development of speech neuroprostheses
53
and
efforts must therefore be made to mitigate false positives.
In the sensorimotor cortex, differential spatial patterns between produced speech, perceived speech and rest can be used as
an ’intended’ speech detector
51
. A similar spatial decoder may also be possible in the superior temporal cortex, considering
differential patterns found between produced and perceived speech here as well
22,44
, with a gradient of more anterior regions
preferring self-generated and posterior regions preferring perceived speech
22
. The hippocampus may also be interesting as it is
found to have a stronger coupling with auditory cortex during self-generated speech versus hearing one’s own pre-recorded
speech and is thought to have a predictive role in speech production
38
. However, we did not see hippocampal involvement
in our current work. The posterior insula, similar to the area we see in this work, is also found to be more activated during
self-generated than perceived speech
54
. However, the insula may be more related to the coordination of speech with autonomic
functions and may therefore not be present during purely imagined speech
55
. In our earlier work, we have also shown that
speech is better reconstructed from this region using neural features during speech articulation rather than before, unlike the
other regions
17
. In our current work, while the largest amount of significant channels overall were based on neural features just
prior to the speech features, we could not distinguish pre- and post-articulation signals due to the continuous nature of our
analysis and different temporal dynamics between our speech tasks. Nonetheless, a speech detector from within or between
such regions that can additionally differentiate between self-generated and perceived speech may be used to mitigate the false
positive effect from speech perception, as well as as on/off switch that a participant can control to mitigate the potential effect
from unintended inner thoughts52,53.
Conclusion
This study provides a comprehensive overview of speech decoding throughout the brain, and underscores the potential of
regions beyond the cortical surface for speech neuroprostheses. We showed that white matter contacts provided strong decoding
accuracy. Sulci regions, which cannot be measured with ECoG or MEA, yielded decoding results on-par with gyral locations.
Additional to the usually recorded left hemisphere, we found that locations on both hemispheres provided very high accuracies,
with the most consistent results derived from auditory regions, followed by the motor and supramarginal cortices. Frontal areas
provided significant decoding results only in the sentence task, indicating the importance of different types of speech in the
9/13
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 3, 2024. ; https://doi.org/10.1101/2024.11.29.626019doi: bioRxiv preprint

training data. Overall, these findings indicate that there is large potential for speech neuroprostheses even when the motor
cortex cannot be targeted.
Data availability
The raw data is publicly available at
https://doi.org/10.17605/OSF.IO/AK3DP
and code can be found on
https:
//github.com/neuralinterfacinglab/SpeechTargets.
References
1.
Silva, A. B., Littlejohn, K. T., Liu, J. R., Moses, D. A. & Chang, E. F. The speech neuroprosthesis. Nat. Rev. Neurosci.
1–20 (2024).
2.
Indefrey, P. The spatial and temporal signatures of word production components: a critical update. Front. psychology 2,
255 (2011).
3.
Hertrich, I., Dietrich, S. & Ackermann, H. The margins of the language network in the brain. Front. Commun. 5, 519955
(2020).
4.
Moses, D. A. et al. Neuroprosthesis for decoding speech in a paralyzed person with anarthria. New Engl. J. Medicine 385,
217–227 (2021).
5.
Metzger, S. L. et al. A high-performance neuroprosthesis for speech decoding and avatar control. Nature 620, 1037–1046
(2023).
6. Willett, F. R. et al. A high-performance speech neuroprosthesis. Nature 620, 1031–1036 (2023).
7.
Luo, S. et al. Stable decoding from a speech bci enables control for an individual with als without recalibration for 3
months. Adv. Sci. 10, 2304853 (2023).
8.
Angrick, M. et al. Online speech synthesis using a chronically implanted brain–computer interface in an individual with
als. Sci. reports 14, 9617 (2024).
9.
Card, N. S. et al. An accurate and rapidly calibrating speech neuroprosthesis. New Engl. J. Medicine 391, 609–618 (2024).
10.
Gallego, J. A., Makin, T. R. & McDougle, S. D. Going beyond primary motor cortex to improve brain–computer interfaces.
Trends neurosciences 45, 176–183 (2022).
11.
Ragagnin, A. M., Shadfar, S., Vidal, M., Jamali, M. S. & Atkin, J. D. Motor neuron susceptibility in als/ftd. Front.
neuroscience 13, 532 (2019).
12.
Hamamé, C. M., Alario, F.-X., Llorens, A., Liégeois-Chauvel, C. & Trébuchon-Da Fonseca, A. High frequency gamma
activity in the left hippocampus predicts visual object naming performance. Brain language 135, 104–114 (2014).
13.
Tankus, A. et al. A speech neuroprosthesis in the frontal lobe and hippocampus: Decoding high-frequency activity into
phonemes. Neurosurgery DOI: 10.1227/neu.0000000000003068 (2024).
14.
Afif, A., Minotti, L., Kahane, P. & Hoffmann, D. Middle short gyrus of the insula implicated in speech production:
intracerebral electric stimulation of patients with epilepsy. Epilepsia 51, 206–213 (2010).
15.
Soroush, P. Z. et al. The nested hierarchy of overt, mouthed, and imagined speech activity evident in intracranial recordings.
NeuroImage 269, 119913 (2023).
16.
Herff, C., Krusienski, D. J. & Kubben, P. The potential of stereotactic-eeg for brain-computer interfaces: current progress
and future directions. Front. neuroscience 14, 123 (2020).
17.
Verwoert, M. et al. Whole-brain dynamics of articulatory, acoustic and semantic speech representations. bioRxiv 2024–08
(2024).
18.
Thomas, T. M. et al. Decoding articulatory and phonetic components of naturalistic continuous speech from the distributed
language network. J. Neural Eng. 20, 046030 (2023).
19.
Ogar, J., Slama, H., Dronkers, N., Amici, S. & Luisa Gorno-Tempini, M. Apraxia of speech: an overview. Neurocase 11,
427–432 (2005).
20.
Mesgarani, N., Cheung, C., Johnson, K. & Chang, E. F. Phonetic feature encoding in human superior temporal gyrus.
Science 343, 1006–1010 (2014).
10/13
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 3, 2024. ; https://doi.org/10.1101/2024.11.29.626019doi: bioRxiv preprint

21.
Meng, K., Grayden, D. B., Cook, M. J., Vogrin, S. & Goodarzy, F. Identification of discriminative features for decoding
overt and imagined speech using stereotactic electroencephalography. In 2021 9th International Winter Conference on
Brain-Computer Interface (BCI), 1–6 (IEEE, 2021).
22.
Nourski, K. V. et al. Electrophysiology of the human superior temporal sulcus during speech processing. Cereb. Cortex 31,
1131–1148 (2021).
23.
Herff, C. & Verwoert, M. Dataset of speech production in intracranial electroencephalography. Open Science Framework
https://doi.org/10.17605/OSF.IO/NRGX6 (2022).
24.
Verwoert, M. et al. Dataset of speech production in intracranial electroencephalography. Sci. Data 9, 434, DOI:
10.1038/s41597-022-01542-9 (2022).
25. Vansteensel, M. J. et al. Fully implanted brain–computer interface in a locked-in patient with als. New Engl. J. Medicine
375, 2060–2066 (2016).
26. Kothe, C. Lab streaming layer (lsl). https://github.com/sccn/labstreaminglayer 26, 2015 (2014).
27.
Amigó-Vega, J. et al. The easy and versatile neural recording platform (t-rex): Design and development study. JMIR
Neurotechnology 2, e47881 (2023).
28.
Hamilton, L. S., Chang, D. L., Lee, M. B. & Chang, E. F. Semi-automated anatomical labeling and inter-subject warping
of high-density intracranial recording electrodes in electrocorticography. Front. Neuroinformatics 11, 62 (2017).
29.
Fischl, B. et al. Whole brain segmentation: Automated labeling of neuroanatomical structures in the human brain. Neuron
33, 341–355, DOI: https://doi.org/10.1016/S0896-6273(02)00569-X (2002).
30.
Destrieux, C., Fischl, B., Dale, A. & Halgren, E. Automatic parcellation of human cortical gyri and sulci using standard
anatomical nomenclature. NeuroImage 53, 1–15, DOI: https://doi.org/10.1016/j.neuroimage.2010.06.010 (2010).
31.
Mercier, M. R. et al. Evaluation of cortical local field potential diffusion in stereotactic electro-encephalography recordings:
A glimpse on white matter signal. NeuroImage 147, 219–232, DOI: https://doi.org/10.1016/j.neuroimage.2016.08.037
(2017).
32.
Martin, S. et al. Decoding spectrotemporal features of overt and covert speech from the human cortex. Front. neuroengi-
neering 7, 14 (2014).
33.
Angrick, M. et al. Real-time synthesis of imagined speech processes from minimally invasive recordings of neural activity.
Commun. biology 4, 1055 (2021).
34.
Koubeissi, M. Z., Fernandez-Baca Vaca, G., Maciunas, R. & Stephani, C. A white matter tract mediating awareness of
speech. Neurology 86, 177–179 (2016).
35.
Li, G. et al. Detection of human white matter activation and evaluation of its function in movement decoding using
stereo-electroencephalography (seeg). J. neural engineering 18, 0460c6 (2021).
36.
Bouton, C. et al. Decoding neural activity in sulcal and white matter areas of the brain to accurately predict individual
finger movement and tactile stimuli of the human hand. Front. Neurosci. 15, 699631 (2021).
37.
Piai, V. et al. Direct brain recordings reveal hippocampal rhythm underpinnings of language processing. Proc. Natl. Acad.
Sci. 113, 11366–11371 (2016).
38.
van de Ven, V., Waldorp, L. & Christoffels, I. Hippocampus plays a role in speech feedback processing. NeuroImage 223,
117319 (2020).
39.
Thibault, S. et al. Tool use and language share syntactic processes and neural patterns in the basal ganglia. Science 374,
eabe0874 (2021).
40.
Turker, S., Kuhnke, P., Eickhoff, S. B., Caspers, S. & Hartwigsen, G. Cortical, subcortical, and cerebellar contributions to
language processing: A meta-analytic review of 403 neuroimaging experiments. Psychol. Bull. (2023).
41.
Dijksterhuis, D. E. et al. Pronouns reactivate conceptual representations in human hippocampal neurons. Science 385,
1478–1484 (2024).
42.
Jiang, X., Zhang, T., Zhang, S., Kendrick, K. M. & Liu, T. Fundamental functional differences between gyri and sulci:
implications for brain function, cognition, and behavior. Psychoradiology 1, 23–41 (2021).
43.
Jensen, M. A. et al. A motor association area in the depths of the central sulcus. Nat. neuroscience 26, 1165–1169 (2023).
44.
Flinker, A. et al. Single-trial speech suppression of auditory cortex activity in humans. J. Neurosci. 30, 16643–16650
(2010).
11/13
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 3, 2024. ; https://doi.org/10.1101/2024.11.29.626019doi: bioRxiv preprint

45.
Wandelt, S. K. et al. Representation of internal speech by single neurons in human supramarginal gyrus. Nat. human
behaviour 1–14 (2024).
46.
Castellucci, G. A., Kovach, C. K., Howard III, M. A., Greenlee, J. D. & Long, M. A. A speech planning network for
interactive language use. Nature 602, 117–122 (2022).
47.
Bourguignon, N. J. A rostro-caudal axis for language in the frontal lobe: the role of executive control in speech production.
Neurosci. & Biobehav. Rev. 47, 431–444 (2014).
48.
Berezutskaya, J., Freudenburg, Z. V., Güçlü, U., van Gerven, M. A. & Ramsey, N. F. Neural tuning to low-level features of
speech throughout the perisylvian cortex. J. Neurosci. 37, 7906–7920 (2017).
49.
Yamamoto, A. K. et al. A special role for the right posterior superior temporal sulcus during speech production. Neuroimage
203, 116184 (2019).
50.
Zhang, W., Liu, Y., Wang, X. & Tian, X. The dynamic and task-dependent representational transformation between the
motor and sensory systems during speech production. Cogn. Neurosci. 11, 194–204 (2020).
51.
Schippers, A., Vansteensel, M. J., Freudenburg, Z. V. & Ramsey, N. F. Don’t put words in my mouth: Speech perception
can generate false positive activation of a speech bci. medRxiv (2024).
52.
Kunz, E. M. et al. Representation of verbal thought in motor cortex and implications for speech neuroprostheses. bioRxiv
2024–10 (2024).
53.
Van Stuijvenberg, O. C., Samlal, D. P., Vansteensel, M. J., Broekman, M. & Jongsma, K. R. The ethical significance of
user-control in ai-driven speech-bcis: a narrative review. Front. Hum. Neurosci. 18, 1420334 (2024).
54.
Woolnough, O., Forseth, K. J., Rollo, P. S. & Tandon, N. Uncovering the functional anatomy of the human insula during
speech. elife 8, e53086 (2019).
55.
Ackermann, H. & Riecker, A. The contribution (s) of the insula to speech production: a review of the clinical and functional
imaging literature. Brain Struct. Funct. 214, 419–433 (2010).
Acknowledgements
This publication is part of the project INTENSE (with project number 17619 of the research programme NWO Crossover
Programme) which is (partly) financed by the Dutch Research Council (NWO). C.H. acknowledges funding by the Kavli
Foundation.
Author contributions statement
C.H. and P.K. and M.V. designed the experiments, C.H, M.C.O. and M.V. collected the data, M.V. ran the analyses, M.V. wrote
the manuscript. All authors reviewed the manuscript and declare no competing interests.
12/13
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 3, 2024. ; https://doi.org/10.1101/2024.11.29.626019doi: bioRxiv preprint

Supplementary Material
Table S1. Included Destrieux/Fischl atlas labels per region-of-interest. The same labels were applied to both hemispheres.
Region-of-interest Labels
Auditory cortex ’S_temporal_transverse’, ’G_temp_sup-G_T_transv’, ’G_temp_sup-Plan_tempo’
Inferior frontal cortex ’G_front_inf-Opercular’, ’G_front_inf-Orbital’, ’G_front_inf-Triangul’, ’G_orbital’,
’S_orbital-H_Shaped’, ’S_suborbital’, ’S_front_inf’, ’S_orbital_lateral’, ’Lat_Fis-ant-’
Vertical’, ’Lat_Fis-ant-Horizont’
Insular cortex ’S_circular_insula_ant’, ’S_circular_insula_inf’, ’S_circular_insula_sup’, ’G_insular_short’
’G_Ins_lg_and_S_cent_ins’
Middle temporal cortex ’G_temporal_middle’
Motor cortex ’S_central’, ’G_and_S_subcentral’, ’G_precentral’, ’S_precentral-inf-part’
Occipital cortex ’G_occipital_middle’, ’G_occipital_sup’, ’G_oc-temp_lat-fusifor’, ’G_oc-temp_med-
Lingual’, ’G_and_S_occipital_inf’, ’Pole_occipital’, ’S_oc_middle_and_Lunatus’, ’S_oc_
sup_and_transversal’, ’S_occipital_ant’, ’S_oc-temp_lat’, ’S_oc-temp_med_and_Lingual’
Subcortical structures ’Amygdala’, ’Hippocampus’, ’Putamen’
Superior temporal cortex ’G_temp_sup-Lateral’, ’S_temporal_sup’, ’G_temp_sup-Plan_polar’
Supramarginal cortex ’G_pariet_inf-Supramar’, ’Lat_Fis-post’
13/13
.CC-BY-ND 4.0 International licenseavailable under a
(which was not certified by peer review) is the author/funder, who has granted bioRxiv a license to display the preprint in perpetuity. It is made
The copyright holder for this preprintthis version posted December 3, 2024. ; https://doi.org/10.1101/2024.11.29.626019doi: bioRxiv preprint