Auditory cortex iEEG responses to pure tone stimuli in normal-hearing and deafened cochlear implant rats.

A, Schematic of 60-channel cortical surface electrode arrays, covering 3.25 x 3.25 mm on the surface of auditory cortex to record tone-evoked responses in normal-hearing (NH) trained rats. B, Some animals were bilaterally deafened and fitted with a unilateral cochlear implant (CI). C, Behavioral training of subset of animals on auditory go/no-go frequency recognition task. Animals were first trained when normal-hearing (NH, N=4, d’: 2.9±0.1) and then re-trained to respond to cochlear implant stimulation (CI, N=3, d’: 0.9±0.2). D, Examples of trial-averaged single-site iEEG responses (top, ERPs; middle, ERP spectrograms; bottom, HG). Clear transients were evoked by tones in normal-hearing animals (NH, left column) and in cochlear implant rats (CI, right column). E, iEEG response magnitude was similar between normal-hearing (NH, ERP amplitude: 1.9±0.1 µV; HG amplitude: 0.6±0.1 a.u.) and cochlear-implant rats (CI, ERP amplitude: 2.0±0.2 µV, p=0.53 compared to normal-hearing ERPs, linear mixed effects; HG amplitude: 0.6±0.1 a.u., p=0.39 compared to normal-hearing HG).

Tone-evoked and cochlear implant-evoked iEEG measurements are spatially organized.

A, F, Trial-averaged tone-evoked event-related potentials (ERPs, panel A) and high gamma (HG, panel F) were spatially restricted to regions along the iEEG recording array; these active regions shifted as a function of stimulus. All panels represent data from the same animal (N=1), unless otherwise noted. B, G, Spatial correlations of evoked response areas decreased monotonically as a function of increasing stimulus separation, truncated at large separations due to low sample sizes. Evoked iEEG responses as assessed by ERPs and HG were nonrandom for normal-hearing and implanted rats (p<10-8 compared to all shuffled responses), but were locally tonotopic in normal-hearing rats (p<10-4) but not implanted rats (ERPs: p=0.5, HG: p=0.7). C, H, Determining the preferred stimulus of spatially-evoked activity revealed gradients shifting from high-to-low to high tone frequencies or implant channels (ERPs, panel C; HG, panel H). D, I, Local tonotopic gradients for each recording site plotted on a unit circle (blue) as a function of magnitude (strength of gradient) and angle (direction of gradient). The vectors across each recording site were averaged to produce an overall tonotopic vector (black) whose magnitude represents a metric of overall topography (ERPs, panel D; HG, panel I). The mean vector was plotted against the mean vector computed from n=1,000 shuffled maps (ERPs, panel D: normal-hearing z-score: 6.6, p<10-10; implanted z-score: 7.9, p<10-14; HG, panel I: normal-hearing z-score: 2.1, p=0.02; implanted z-score: 1.5, p=0.07). E, J, Z-scored magnitudes of non-random spatial structure across animals (ERPs, panel E: normal-hearing mean z-score: 4.7±1.0, implanted mean z-score: 2.6±1.0, p=0.05, linear mixed effects; HG, panel J: normal-hearing mean z-score: 3.0±0.9, implanted mean z-score: 2.1±0.9, p=0.48).

Trial-by-trial implant-evoked spatial response patterns are more variable than tone-evoked patterns.

A, C, Trial-by-trial evoked iEEG measurements in both normal-hearing and implanted rats revealed peaks across time (left) and space (right) (ERP, panel A; HG, panel C). All panels represent data from the same animal (N=1), unless otherwise noted. B, D, Variability of iEEG measurements across trials (root mean square, rms) was consistently higher for spatial distributions of cochlear implant-evoked compared to tone-evoked activity, for ERPs in panel B (bottom, ERP spatial NH mean rms: 0.32±0.01, CI mean rms: 0.33±0.02, p=0.011, linear mixed effects) but not for temporal distributions (top, ERP temporal NH mean rms: 1.6±0.3 over all animals, CI mean rms: 1.9±0.5, for animals monitored before and after deafening, p=0.054) and HG in panel D (top, HG temporal NH mean rms: 2.4±0.3 over all animals, CI mean rms: 3.3±0.9, p=0.17; bottom, HG spatial NH mean rms: 1.2±0.2, CI mean rms: 1.3±0.3, p=0.35).

Trial-by-trial iEEG measurements encode stimulus identity.

A, Trial-by-trial ERPs evoked by 1.4 kHz (left), 8 kHz (middle), and 32 kHz tones (right) are plotted for three recording sites (blue) and reveal discernible evoked transients and spatially restricted patterns of evoked magnitudes (stimulus onsets depicted by a dotted line). All panels represent data from the same animal (N=1), unless otherwise noted. B, Trial-by-trial iEEG measurements concatenated such that columns represent recording sites (spatial) and post-stimulus sampling (temporal), rows by stimulus trials. Data then reorganized by PCA and reduced to the 15 components according to magnitude of explained variance. C, Classification matrices are plotted as means across bootstrapped repeats (N=1,000) versions of LDA classifiers trained using 13 randomly selected trials for each stimulus, and classification predictions of single and remaining trials reveal significant prediction of stimulus identity (dashed line: chance-level error distance; solid lines: actual mean error distances). D, Stimulus prediction probabilities plotted across animals (black: mean, grey: s.e.m.) and as function of either octaves (normal-hearing) or channels (cochlear implant) from actual stimulus, reflecting a gradient in which adjacent stimuli share more encoded features than other stimuli. E, Decoder performance (mean error distance) across individual animals was somewhat worse for cochlear implant ERPs compared to tone-evoked ERPs (ERP mean error rate relative to chance for normal-hearing: 0.74±0.05, cochlear implant: 0.78±0.06; for animals assessed both first when normal-hearing and then after cochlear implantation p<10-4, linear mixed effects), but not for HG (HG mean error rate relative to chance for normal-hearing: 0.90±0.03, cochlear implant: 0.91±0.04; for animals assessed both first when normal-hearing and then after cochlear implantation p=0.15).

Trial-by-trial decoding from either spatial or temporal aspects of iEEG signals.

A, Raw trial-by-trial ERPs for a single animal are plotted for 3 channels. All panels represent data from the same animal (N=1), unless otherwise noted. B-C, Trial-by-trial iEEG measurements are reduced into either spatial only (B) or temporal only (C) by extracting either the magnitude of evoked activity or averaging across all recording sites. D, Re-training PCA-LDA decoders on spatial-only (middle) or temporal-only (right) does not abolish the encoding of stimulus identity. E, Decoder errors (mean error distance re. chance) across animals are plotted for decoders trained on spatial+temporal, spatial-only, and temporal-only iEEG measurements. Mean error distance for spatial+temporal ERPs, normal-hearing: 0.74±0.04, implant: 0.86±0.06; for spatial+temporal HG, normal-hearing: 0.90±0.03, implant: 0.97±0.04. Mean error distance for spatial-only ERPs, normal-hearing: 0.77±0.04, implant: 0.77±0.08; for HG, normal-hearing: 0.89±0.04, implant: 0.87±0.05. Mean error distance for temporal-only ERPs, normal-hearing: 0.82±0.05, implant: 0.81±0.07; for HG, normal-hearing: 0.87±0.04, implant: 0.93±0.05.

TCA-based decoding of single-trial iEEG measurements.

A, Schematic of TCA with 3-dimensional tensors with orthogonal dimensions of spatial, temporal and trial factors. All panels represent data from the same animal (N=1), unless otherwise noted. B, TCA-reduced iEEG measurements of stimulus-evoked ERPs in normal-hearing rats are plotted across 15 components of spatial (top row), temporal (middle row) and trial factors (bottom row). C, Classification matrices showing mean decoder predictions from bootstrapped repeats (N=1,000) versions LDA classifiers trained using 13 randomly selected trials for each stimulus indicated that iEEG measurements encode stimulus identity from single stimulus presentations (dashed line: chance-level error distance; solid lines: actual mean error distances). D, Stimulus prediction probabilities are plotted across animals (black: mean, grey: s.e.m.) as a function of either octaves or channels away from actual stimulus, for normal-hearing and implanted rats. Correct prediction probabilities are high across all conditions and prediction errors cluster toward the actual stimuli (orange: mean and s.e.m. of decoder performance on shuffled trials; grey: mean error distance, dashed line: chance-level error distance). E, Decoder errors (mean error distance) across individual animals were slightly smaller for evoked iEEG measurements in normal-hearing compared to implanted rats (ERP mean error rate re. chance normal-hearing: 0.77±0.04, implanted: 0.80±0.06; HG mean error rate re. chance normal-hearing: 0.90±0.03, implanted: 0.88±0.05). Comparisons between normal-hearing and implanted iEEG measurements showed significant decrease in estimated error distances for ERP (top, linear mixed-effects model: p<10-5) and HG (bottom, p=0.03).

TCA data reductions reveal latent spatial factors with coarse, non-random spatial organization.

A, TCA-based analysis of evoked neural activity, divided into 15 unique components where each component is separated into orthogonal dimensions of spatial factors (top row), temporal factors (middle row) and trial factors (bottom row). All panels represent data from the same animal (N=1), unless otherwise noted. B, Spatial maps for tone-evoked or electrode-evoked responses recreated from TCA reduced data. C, Reducing re-organized spatial factors revealed a TCA map (left) with gradients in the same location and direction as the maps reduced from raw measurements (right). D, Spatial correlations of evoked response areas decreased monotonically as a function of increasing stimulus separation. TCA-reduced models of evoked iEEG responses as assessed by ERPs and HG exhibited non-random spatial organization in normal-hearing and implanted rats (p<10-8 compared to all shuffled responses). TCA-reduced models were locally tonotopic only for HG in normal-hearing rats (normal-hearing ERPs: p=0.18, normal-hearing HG: p<10-4, cochlear implant ERPs: p=0.43, cochlear implant HG: p=0.40). E, Z-scored magnitudes of mean tonotopic vectors across animals were similar (ERPs, normal-hearing mean z-score: 6.2±3.6, cochlear implant mean z-score: 3.9±2.1, p=0.16, linear mixed effects; HG, normal-hearing mean z-score: 2.9±1.1, implanted mean z-score: 2.5±0.7, p=0.75).

Lack of information transfer between acoustic and electrical stimulation representations in the same animals.

A, Evoked iEEG measurements from a subset of data from normal-hearing rat reduced using TCA. All panels represent data from the same animal (N=1), unless otherwise noted. B, Evoked iEEG measurements from previously withheld trials from normal-hearing rat reduced using TCA, constraining the model with the spatial and temporal factors from the original TCA model, leaving only the trials as the optimizable variables. C, Linear discriminant analysis classifiers are trained on the trial factors extracted from the original TCA-reduced models of evoked data and then used to predict stimulus identity from the trial factors from the re-optimized TCA-reduced models of withheld normal-hearing measurements. D, Classification matrices are plotted as the mean decoder predictions across bootstrapped repeats (N=1,000) of linear-discriminant analysis (LDA) classifiers. Mutual information (IT) is indicated above these mean classification matrices. E, Stimulus prediction probabilities plotted across animals (black: mean, grey: s.e.m.) and as function of either octaves (normal-hearing) or channels (cochlear implant) from actual stimulus, reflecting a gradient in which adjacent stimuli share more encoded features than other stimuli (for ERP, but not HG). F, Evoked iEEG measurements in a cochlear implant rat were reduced using TCA, constraining the model with the spatial and temporal factors from the original TCA model optimized on normal-hearing data and leaving only the trials as the optimizable variables. Linear discriminant analysis classifiers were trained on the trial factors extracted from the original TCA-reduced models of evoked data in normal-hearing conditions and then used to predict stimulus identity from the trial factors from the TCA-reduced models of implant-evoked measurements. G, Classification matrices are plotted as the mean decoder predictions across bootstrapped repeats (N=1,000) of linear-discriminant analysis (LDA) classifiers for the normal-hearing-trained, cochlear-implant-tested condition. H, Mutual information (IT) is computed from each confusion matrix and divided by the maximum possible number of bits given the respective matrix dimensions. Normal-hearing-trained, implant-tested decoders reveal little-to-no information transfer (IT mean across animals, ERP: 1.2+/-0.6%; HG: 0.3+/-0.1%) compared to both the normal-hearing-only condition (IT mean across animals, ERP: 2.8+/-1.2%, p=0.04; HG: 0.4+/-0.4%, p=0.19, one-tailed t-test) and the shuffled implant-tested condition, in which rows were shuffled within columns to estimate a chance-level baseline (IT mean across animals, ERP: 2.4+/-1.0%, p=0.04; HG: 0.3+/-0.2%, p=0.21, one-tailed t-test).

ECAP measurements of spatio-temporal cochlear implant tuning.

A, Forward masking was performed to assess the spatial tuning of cochlear stimuli in the cochlea via ECAPs measured by adjacent cochlear implant electrodes. These measurements were made in a separate group of animals (N=3) than those used for iEEG recordings. B, Forward masking was performed to assess temporal tuning of cochlear implant stimuli in the cochlea via ECAPs measured by adjacent cochlear implant electrodes. The masker stimulus was delivered by the same electrode as the probe. C, Spatial tuning functions were averaged across all probe electrodes in N=3 animals (black, mean; gray: s.e.m.; orange, average of individual subjects). D, Temporal tuning functions were averaged across all probe electrodes and N=3 animals (black, mean; gray, s.e.m.; orange, average of individual subjects).

Experimental timeline and animal tracker.

A, Rats underwent a combination of behavioral training (when typically-hearing), iEEG recordings, and then behavioral testing (after deafening) in a specific sequence, but not all animals included here participated in every phase. B, iEEG measurements were obtained in a total of N=10 animals. In a subset of these animals (N=4), we also obtained iEEG measurements in response to both acoustic tones before deafening and cochlear implant stimulation after deafening. In addition to iEEG recordings, we also behaviorally trained either prior (tones) or after iEEG recordings with cochlear implant stimuli.

Maps of stimulus preferences for tone- or cochlear implant-evoked iEEG measurements.

A, Maps of preferred stimuli are derived from event-related potentials (ERP) for 10 animals, including tone evoked iEEG measurements (left column) and cochlear implant-evoked iEEG\ measurements (right column). For the cochlear implant, single channel stimuli were either delivered via tones presented to the cochlear implant processor (top 3 maps) and or single channel stimuli were triggered by a direct connection between computer and the cochlear implant processor (bottom 3 maps). In some cochlear-implanted animals, apical electrodes were non-functional. The numbers of these non-functional electrodes are denoted with a superimposed red circle and backslash.

Numbers of components for PCA and TCA models

A. The means (+/- s.e.m.) of explained variance per component for PCA models across all ERP data (top row) decrease rapidly from 30 percent to near zero within 30 components. The means (+/- s.e.m.) of explained variance per component for PCA models for HG data (bottom row) and decrease from 12 percent rapidly to a steady decline within 30 components. Orange dashed line: linear regression of components 16-30 are plotted. B. The means (+/- s.e.m.) of model error for TCA models of ERP (top row) and HG (bottom row) decrease as a function of number of components.