Skip to main content
eLife logoLink to eLife
. 2022 Aug 12;11:e78677. doi: 10.7554/eLife.78677

Information flows from hippocampus to auditory cortex during replay of verbal working memory items

Vasileios Dimakopoulos 1, Pierre Mégevand 2,3, Lennart H Stieglitz 1, Lukas Imbach 4,5, Johannes Sarnthein 1,5,
Editors: Timothy D Griffiths6, Laura L Colgin7
PMCID: PMC9374435  PMID: 35960169

Abstract

The maintenance of items in working memory (WM) relies on a widespread network of cortical areas and hippocampus where synchronization between electrophysiological recordings reflects functional coupling. We investigated the direction of information flow between auditory cortex and hippocampus while participants heard and then mentally replayed strings of letters in WM by activating their phonological loop. We recorded local field potentials from the hippocampus, reconstructed beamforming sources of scalp EEG, and – additionally in four participants – recorded from subdural cortical electrodes. When analyzing Granger causality, the information flow was from auditory cortex to hippocampus with a peak in the [4 8] Hz range while participants heard the letters. This flow was subsequently reversed during maintenance while participants maintained the letters in memory. The functional interaction between hippocampus and the cortex and the reversal of information flow provide a physiological basis for the encoding of memory items and their active replay during maintenance.

Research organism: Human

eLife digest

Every day, the brain’s ability to temporarily store and recall information – called working memory – enables us to reason, solve complex problems or to speak. Holding pieces of information in working memory for short periods of times is a skill that relies on communication between neural circuits that span several areas of the brain. The hippocampus, a seahorse-shaped area at the centre of the brain, is well-known for its role in learning and memory. Less clear, however, is how brain regions that process sensory inputs, including visual stimuli and sounds, contribute to working memory.

To investigate, Dimakopoulos et al. studied the flow of information between the hippocampus and the auditory cortex, which processes sound. To do so, various types of electrodes were placed on the scalp or surgically implanted in the brains of people with drug-resistant epilepsy. These electrodes measured the brain activity of participants as they read, heard and then mentally replayed strings of up to 8 letters. The electrical signals analysed reflected the flow of information between brain areas.

When participants read and heard the sequence of letters, brain signals flowed from the auditory cortex to the hippocampus. The flow of electrical activity was reversed while participants recalled the letters. This pattern was found only in the left side of the brain, as expected for a language related task, and only if participants recalled the letters correctly.

This work by Dimakopoulos et al. provides the first evidence of bidirectional communication between brain areas that are active when people memorise and recall information from their working memory. In doing so, it provides a physiological basis for how the brain encodes and replays information stored in working memory, which evidently relies on the interplay between the hippocampus and sensory cortex.

Introduction

Working memory (WM) describes our capacity to represent sensory input for prospective use (Baddeley, 2003; Christophel et al., 2017). Maintaining content in WM requires communication within a widespread network of brain regions. The anatomical basis of WM was shown noninvasively with EEG/MEG (Michels et al., 2008; Sarnthein et al., 1998; Tuladhar et al., 2007; Polanía et al., 2012; Näpflin et al., 2008; Bidelman et al., 2021; Pavlov and Kotchoubey, 2022; Hsieh and Ranganath, 2014) and invasively with intracranial local field potentials (LFP; Cogan et al., 2017; Raghavachari et al., 2001; Rizzuto et al., 2003; Maris et al., 2011; van Vugt et al., 2010; Leszczyński et al., 2015; Johnson et al., 2018a; Johnson et al., 2018b; Boran et al., 2019; Li et al., 2022; Schwiedrzik et al., 2018) and single-unit recordings (Boran et al., 2019; Schwiedrzik et al., 2018; Kamiński et al., 2017; Kornblith et al., 2017; Rutishauser et al., 2021).

In cortical brain regions, WM maintenance correlates with sustained neuronal oscillations, most frequently reported in the theta-alpha range ([4 12] Hz; Michels et al., 2008; Sarnthein et al., 1998; Tuladhar et al., 2007; Polanía et al., 2012; Näpflin et al., 2008; Pavlov and Kotchoubey, 2022; Hsieh and Ranganath, 2014; Cogan et al., 2017; Raghavachari et al., 2001; Rizzuto et al., 2003; Maris et al., 2011; van Vugt et al., 2010; Leszczyński et al., 2015; Johnson et al., 2018a; Johnson et al., 2018b; Boran et al., 2019; Li et al., 2022) or at even lower frequencies (Kumar et al., 2021; Rezayat et al., 2021). Also in the hippocampus, WM maintenance was associated with sustained theta-alpha oscillations (van Vugt et al., 2010; Boran et al., 2019). As a hallmark for WM maintenance, persistent neuronal firing was reported during the absence of sensory input, indicating the involvement of the medial temporal lobe in WM (Boran et al., 2019; Kamiński et al., 2017; Kornblith et al., 2017; Boran et al., 2022).

At the network level, synchronized oscillations have been proposed as a mechanism for functional interactions between brain regions (Fries, 2015; Pesaran et al., 2018). It is thought that these oscillations show temporal coupling of the low-frequency phase for long-range communication between cortical areas (Sarnthein et al., 1998; Polanía et al., 2012; Maris et al., 2011; Johnson et al., 2018a; Johnson et al., 2018b; Boran et al., 2019; Solomon et al., 2017). This synchronization suggests an active maintenance process through reverberating signals between brain regions.

We here extend previous studies with the same task (Michels et al., 2008; Boran et al., 2019) by recording from four participants with hippocampal LFP and direct cortical recordings (ECoG) from electrodes over primary auditory, parietal, and occipital cortical areas. Given the low incidence of the epileptogenic zone in parietal cortex, parietal ECoG recordings are rare. To benefit from the wide spatial coverage of scalp EEG, we analyzed the directed functional coupling between hippocampal LFP and the beamforming sources of scalp EEG in all 15 participants. We found that the information flow was from auditory cortex to hippocampus during the encoding of WM items, and the flow was from hippocampus to auditory cortex for the replay of the items during the maintenance period.

Results

Task and behavior

Fifteen participants (median age 29 years, range [18–56], 7 male, Table 1) performed a modified Sternberg WM task (71 sessions in total, 50 trials each). In the task, items were presented all at once rather than sequentially, thus separating the encoding period from the maintenance period. In each trial, the participant was instructed to memorize a set of 4, 6, or 8 letters presented for 2 s (encoding). The number of letters was thus specific for the memory workload. The participants read the letters themselves and heard them spoken at the same time. Since participants had difficulties reading eight letters within the 2 s encoding period, also hearing the letters assured their good performance. After a delay (maintenance) period of 3 s, a probe letter prompted the participant to retrieve their memory (retrieval) and to indicate by button press (‘IN’ or ‘OUT’) whether or not the probe letter was a member of the letter set held in memory (Figure 1a). During the maintenance period, participants rehearsed the verbal representation of the letter strings subvocally, i.e., mentally replayed the memory items. Participants had been instructed to employ this strategy, and they confirmed after the sessions that they had indeed employed this strategy. This activation of the phonological loop (Baddeley, 2003) is a component of verbal WM as it serves to produce an appropriate behavioral response (Christophel et al., 2017).

Table 1. Participant characteristics and results of Granger causality analysis.

For each participant, we report the atlas parcels that contained EEG sources with the maximal t-value and the t-value of sources in auditory cortex (Heschl gyrus) during encoding and maintenance (non-parametric cluster-based permutation test p<0.05). In each participant, the vast majority of the significant LCMV sources were in the left hemisphere, both during encoding (≥87%) and during maintenance (≥81%). We also report the net information flow (ΔGranger) for correct and incorrect trials in the direction auditory cortex → hippocampus during encoding and in the direction hippocampus → auditory cortex during maintenance.

Participant Pathology Encoding Maintenance
Maximal LCMV source Significant LCMV sources in the left hemisphere (%) max. t-value Heschl t-value Heschl ΔGranger correct trials Heschl ΔGranger incorrect trials Maximal LCMV source Significant LCMV sources in the left hemisphere (%) max. t-value Heschl t-value Heschl ΔGranger correct trials Heschl ΔGranger incorrect trials
1 hippocampal sclerosis Heschl / Temporal Inferior L 100 17.8 17.8 –0.036 0.087 Frontal Mid Orb / Heschl L 100 10.1 10.1 0.025 –0.037
2 non-lesional Heschl L 100 19.8 19.8 –0.017 0.016 Temporal Inferior L 96 11.4 10 0.093 0.002
3 focal cortical dysplasia Temporal Superior L 91 24.2 16.3 –0.060 –0.013 Heschl L 91 14.6 14.6 0.065 0.005
4 unclear etiology Frontal Inferior L 100 18.3 16.6 –0.006 0.003 Heschl L 100 13.4 13.4 0.035 –0.002
5 brain contusion Temporal Superior L 100 6.9 5.4 –0.003 –0.002 Heschl L 96 7.6 7.6 0.021 0.048
6 hippocampal sclerosis Supramarginal L 98 11.1 9.7 –0.049 0.025 Temporal Pole Superior L 93 19.8 17.9 0.039 –0.024
7 xanthoastrozytoma Lingual R 87 12.9 11.5 –0.059 0.042 Caudate L 85 9.3 8.2 0.042 –0.036
8 focal cortical dysplasia Caudate L 100 18.8 16.2 –0.040 0.013 Parietal Superior L 100 9.9 8.3 0.017 –0.031
9 gliosis Cingulum Anterior L 100 12.3 11.2 –0.051 0.012 Parietal Inferior L 100 12.6 11.2 0.050 –0.015
10 hippocampal sclerosis Heschl L 100 7.9 7.9 –0.070 –0.016 Cingulum Mid L 100 2.7 2.3 0.018 –0.081
11 hippocampal sclerosis Cingulum Anterior L 100 10.8 6.3 –0.052 –0.011 Cuneus L 86 4.8 4.3 0.020 –0.062
12 hippocampal sclerosis Temporal Superior L 95 8.1 6.7 –0.022 0.000 Temporal Pole Mid L 100 7.3 4.9 0.012 0.003
13 hippocampal sclerosis Temporal Superior R 93 10.8 6.9 –0.018 0.019 Parietal Superior R 81 5.6 4.9 0.020 –0.022
14 hippocampal sclerosis Temporal Superior L 100 16.8 13.3 –0.053 0.022 Heschl L 100 11.8 11.8 0.085 –0.089
15 hippocampal sclerosis Heschl L 98 9.1 9.1 –0.061 –0.005 Heschl L 100 10.4 10.4 0.069 –0.004

LCMV, linearly constrained minimum variance; ΔGranger, difference of GC spectra.

Figure 1. Task and recording sites.

(a) In the task, sets of consonants are presented and have to be memorized. The set size (4, 6, or 8 letters) determines working memory workload. In each trial, presentation of a letter string (encoding period, 2 s) is followed by a delay (maintenance period, 3 s). After the delay, a probe letter is presented. Participants indicate whether the probe was in the letter string or not. (b) Response accuracy decreases with set size (71 sessions). (c) Reaction time increases with set size (53 ms/item). (d) The tip locations of the hippocampal local field potentials electrodes for all participants (N=15) are projected in a hippocampal surface.

Figure 1.

Figure 1—figure supplement 1. Hippocampal contact locations.

Figure 1—figure supplement 1.

The recording location of the hippocampal local field potentials electrode of each participant is projected in a left hippocampal surface.

The mean correct response rate was 91% (both for IN and OUT trials). The rate of correct responses decreased with set size from a set size of 4 (97% correct responses) to set sizes of 6 (89%) and 8 (83%) (Figure 1b). Across the participants, the memory capacity averaged 6.1 (Cowan’s K, [correct IN rate +correct OUT rate –1]×set size), which indicates that the participants were able to maintain at least six letters in memory. The mean response time (RT) for correct trials (3045 trials) was 1.1±0.5 s and increased with workload from set size 4 (1.1±0.5 s) to 6 (1.2±0.5 s) and 8 (1.3±0.6 s), 53 ms/item (Figure 1c). Correct IN/OUT decisions were made more rapidly than incorrect decisions (1.1±0.5 s vs 1.3±0.6 s). These data show that the participants performed well in the task and that the difficulty of the trials increased with the number of letters in the set. In further analysis, we focused on correct trials with set size 6 and 8 letters to assure hippocampal activation and hippocampo-cortical interaction as shown earlier (Boran et al., 2019).

Power spectral density in cortical and hippocampal recordings

To investigate how cortical and hippocampal activity subserves WM processing, we analyzed the LFP recorded in the hippocampus (Figure 1d, Figure 1—figure supplement 1, Supplementary file 1) together with ECoG from cortical strip electrodes (Figure 2a, Figure 3a and f). In the following, we present power spectral density (PSD) time-frequency maps from representative electrode contacts. In an occipital recording of Participant 1 (grid contact H3), strong gamma activity (>40 Hz) in the relative PSD occurred while the participant viewed the letters during encoding (increase >100% with respect to fixation, Figure 2b). Similarly, encoding elicited gamma activity in a temporal recording over auditory cortex (increase >100%, grid contact C2, Figure 2c), similar as in Kumar et al., 2021. Gamma increased significantly only in temporal and occipital-parietal contacts (permutation test with z-score >1.96, Figure 2a).

Figure 2. Encoding and replay of letters in Participant 1.

Figure 2.

(a) Location of the ECoG contacts over temporal and parietal cortex for Participant 1.Relative gamma power spectral density (PSD; [60 80] Hz) during encoding ([−3.5 −3] s) is maximal for contacts over temporal and occipital-parietal cortex. (b) The relative PSD in the occipital contact (contact H3) over visual cortex shows gamma activity (>40 Hz) during encoding ([−5 −3] s) while the subject sees and hears the letters. Sustained low beta activity ([11 14] Hz) appears toward the end of the maintenance period ([–3 0] s). (c) The relative PSD in the temporal contact (contact C2) over auditory cortex shows gamma activity ([60 80] Hz) during the last second of encoding ([−4 −3] s) while the subject sees and hears the letters. (d) Relative beta PSD ([11 14] Hz) during maintenance ([−2 0] s) is maximal for contacts over temporal and occipital cortex. (e) Hippocampal PSD shows sustained beta activity toward the end of maintenance. (f) Phase-locking value (PLV) between hippocampus and auditory cortex (contact C3) during fixation (black), encoding (blue), and maintenance (red). The PLV spectra show a broad frequency distribution. The PLV during maintenance is higher than during fixation. Red bars: frequency ranges of significant PLV difference (p<0.05, cluster-based non-parametric permutation test against a null distribution with scrambled trials during fixation and maintenance). (g) PLV between hippocampus and cortex in theta ([4 8] Hz) during maintenance ([−2 0] s) is highest to contacts over auditory cortex. (h) Spectral Granger causality. During encoding ([−5 −3] s), auditory cortex (contact C2) predicts hippocampus ([6 8] Hz, dark blue curve exceeds light blue curve). During maintenance ([−2 0] s), hippocampus predicts auditory cortex ([5 8] Hz, dark red curve exceeds light red curve). Bars: frequency range of significant ΔGranger (p<0.05), cluster-based non-parametric permutation test against a null distribution with scrambled trials during encoding (blue) and maintenance (red). (i) Net information flow ΔGranger ([4 8] Hz) during encoding ([−5 −3] s). ECoG over auditory cortex predicts hippocampal local field potentials. (j) Net information flow ΔGranger ([4 8] Hz) during maintenance ([−2 0] s). Hippocampus is maximal in predicting auditory cortex (contact C2 and surrounding contacts). (k) Statistical significance of the spatial spread of contacts with high ΔGranger ([4 8] Hz) during maintenance ([−2 0] s). We calculated the scalar product between two spread vectors. We then tested the statistical significance of the scalar product. The true distribution (red) is clearly distinct from the null distribution (gray, blue bar marks 95th percentile). (l) The Granger time-frequency map illustrates the time course of the spectra of panel (h). During encoding, net information (ΔGranger) flows from auditory cortex to hippocampus (blue). During maintenance, the information flow is reversed from hippocampus to auditory cortex (red) indicating the replay of letters in memory. Grid contacts with significant increase are marked with a yellow rim (p<0.05, cluster-based non-parametric permutation test against a null distribution with scrambled trials). The time course in time-frequency maps is shown relative to the fixation period (b, c, e). Colors of Granger spectra indicate information flow: dark blue, cortex to hippocampus during encoding; light blue, hippocampus to cortex during encoding; dark red, hippocampus to cortex during maintenance; light red, cortex to hippocampus during maintenance. ΔGranger is the difference between spectra, where ΔGranger <0 denotes information flow cortex→hippocampus and ΔGranger >0 denotes information flow hippocampus→cortex. Grid contacts are identified by column (anterior A to posterior H) and row (inferior 1 to superior 8).

Figure 3. Encoding and replay of letters in three participants with ECoG.

Figure 3.

(a) Location of the ECoG contacts in Participant 2. The most anterior strip contact records from auditory cortex. Color bar: ΔGranger during maintenance ([4 8] Hz). (b) The relative power spectral density (PSD) in the temporal scalp EEG electrode (T5) shows beta activity ([14 25] Hz) during encoding ([−5 −3] s) while the subject sees and hears the letters. Sustained theta activity ([6 9] Hz) appears toward the end of the maintenance period ([–3 0] s). (c) Hippocampal PSD shows alpha-beta activity (9–18 Hz) toward the end of maintenance. (d) Spectral Granger causality (GC). During encoding, the auditory cortex predicts hippocampus ([6 8] Hz, dark blue curve exceeds light blue curve). During maintenance, hippocampal local field potentials (LFP) predict auditory cortex ([6 10] Hz, dark red curve exceeds light red curve). (e) The time-frequency map illustrates the time course of ΔGranger in Participant 2. (f) Location of the ECoG contacts in Participant 3. The most posterior contact records from visual cortex (yellow rimmed disk). Color bar: ΔGranger during maintenance ([4 8] Hz). (g) The relative PSD in the most posterior contact (yellow rimmed disk, panel (f)) shows gamma during encoding while the subject sees the letters. Sustained alpha activity ([8 11] Hz) appears toward the end of the maintenance period. (h) Hippocampal PSD shows sustained beta activity ([13 21] Hz) toward the end of the maintenance. (i) Spectral GC. During encoding, the occipital ECoG predicts hippocampus (6–9 Hz, dark blue curve exceeds light blue curve). During maintenance, hippocampal LFP predicts ECoG ([6 8] Hz, dark red curve exceeds light red curve). (j) The time-frequency map illustrates the time course of ΔGranger in Participant 3. (k) Location of the ECoG contacts in Participant 4 on right parietal cortex. Color bar: ΔGranger during maintenance ([4 8] Hz). (l) The relative PSD in contact over the right parietal lobule shows gamma during encoding while the subject sees the letters. Sustained alpha activity ([8 11] Hz) appears during the maintenance period. (m) Hippocampal PSD shows sustained beta activity ([13 21] Hz) toward the end of the maintenance. (n) Spectral GC. Task performance does not elicit significant GC to the right parietal cortex in Participant 4. (o) The time-frequency map illustrates the time course of ΔGranger in Participant 4. Task performance does not elicit significant GC to the right parietal cortex in Participant 4. Color bar: ΔGranger during maintenance ([4 8] Hz). Grid contacts with significant increase in ΔGranger are marked with a yellow rim (permutation test p<0.05). The time course in time-frequency maps is shown relative to the fixation period (b, c, g, h, l,m). Colors of Granger spectra indicate information flow: dark blue, cortex to hippocampus during encoding; light blue, hippocampus to cortex during encoding; dark red, hippocampus to cortex during maintenance; light red, cortex to hippocampus during maintenance. ΔGranger is the difference between spectra where ΔGranger <0 denotes information flow cortex→hippocampus and ΔGranger >0 denotes information flow hippocampus→cortex. Bars: frequency range of significant ΔGranger (p<0.05), cluster-based non-parametric permutation test against a null distribution with scrambled trials during encoding and maintenance, respectively.

After the letters disappeared from the screen, activity occurred in the [11 14] Hz range (high alpha/low beta, Figure 2b) toward the end of the maintenance period in temporal and occipital contacts (permutation test p<0.05, Figure 2d). Similarly, the temporal scalp EEG of Participant 2 (black rimmed disk denotes electrode site T3 in Figure 3a) showed activity during encoding and maintenance, albeit at lower frequencies (Figure 3b); this pattern was found only in scalp EEG and not in ECoG, probably because the strip electrode was not located over auditory cortex. In Participant 3, a similar pattern occurred in the PSD of a temporo-parietal recording (most posterior strip electrode contact, Figure 3f), where the appearance of the letters prompted gamma activity and the maintenance period showed alpha activity ([8 11] Hz, Figure 3g). Similarly, in the electrode contacts on right parietal cortex of Participant 4 (Figure 3k), the letter stimulus elicited gamma activity and the maintenance period showed alpha activity (8–11 Hz, Figure 3l).

The site of the participants’ maintenance activity coincides with the generator of scalp EEG that was found in the parietal cortex for the same task (Michels et al., 2008). The PSD thereby confirmed the findings of local synchronization of cortical activity during WM maintenance (Michels et al., 2008; Bidelman et al., 2021; Pavlov and Kotchoubey, 2022).

In the hippocampus of all four participants, we found elevated activity in the beta range ([12 24] Hz) toward the end of the maintenance period (increase >100%, Figure 2e, Figure 3c, h and m), confirming the hippocampal contribution to processing of this task (Boran et al., 2019).

Functional coupling between hippocampus and cortex

To investigate the functional coupling between cortex and hippocampus, we first calculated the phase-locking value (PLV). In Participant 1, we found high PLV over a broad frequency range in contacts over auditory cortex throughout the trial. Compared to encoding, maintenance showed enhanced PLV in the theta range between hippocampal LFP and cortical ECoG (PLV=0.4 in contact C3, permutation test p<0.05, Figure 2f). PLV in the [4 8] Hz theta range increased significantly with several contacts over auditory cortex (permutation test p<0.05, Figure 2g). This speaks for a functional coupling between auditory cortex and hippocampus mediated by synchronized oscillations (Rezayat et al., 2021).

Directed functional coupling between hippocampus and ECoG

What was the directionality of the information flow during encoding and maintenance in a trial? We used spectral Granger causality (GC) as a measure of directed functional connectivity to determine the direction of the information flow between auditory cortex and hippocampus in Participant 1 during the trials. During encoding, the information flow was from auditory cortex to hippocampus with a maximum in the theta frequency range (dark blue curve in Figure 2h). The net information flow ΔGranger (GC hipp→cortex – GC cortex→hipp) during encoding was significant in the [6 8] Hz range (blue bar in Figure 2h, p<0.05 permutation test against a null distribution). During maintenance, the information flow in the theta frequency range was reversed (dark red curve), i.e., from hippocampus to auditory cortex (dark red curve in Figure 2h). The net information flow ΔGranger during maintenance was significant in the [5 8] Hz range (red bar in Figure 2h, p<0.05 permutation test against a null distribution). Concerning the spatial spread of the theta GC, the maximal net information flow ΔGranger (GC hipp→cortex – GC cortex→hipp) during encoding occurred from auditory cortex to hippocampus (p<0.05, permutation test, Figure 2i). During maintenance, the theta ΔGranger was significant from hippocampus to both auditory cortex and occipital cortex (permutation test p<0.05, Figure 2j). Interestingly, in Participant 1, the distribution of high ΔGranger coincides with the distribution of high PLV: both show a spatial maximum to grid contacts over auditory cortex and both appear in the theta frequency range.

We next tested the statistical significance of the spatial spread of contacts with high ΔGranger ([4 8] Hz) during maintenance ([−2 0] s). To provide a sound statistical basis, we tested the spatial distribution of GC on the grid contacts against a null distribution. The activation on grid contacts was reshaped into a grid vector. The spatial collinearity of two grid vectors was captured by their scalar product. We next performed 200 iterations of random trial permutations. For each iteration, we selected two subsets of trials, and we calculated the scalar product between the two vectors corresponding to these subsets. We then tested the statistical significance of the scalar product (Figure 2k). The true distribution (red) is clearly distinct from the null distribution (gray, blue bar marks the 95th percentile). The analogous procedure was performed for PSD (Figure 2a and d), PLV (Figure 2g), and GC during encoding (Figure 2i), which gave equally significant results in all cases.

As a further illustration of the ΔGranger time course, the time-frequency plot (Figure 2l) shows the difference between GC spectra (GC hipp→cortex – GC cortex→hipp) at each time point, where blue indicates net flow from auditory cortex to hippocampus and red indicates net flow from hippocampus to auditory cortex.

Similarly in Participant 2, the time course of GC followed the same pattern between auditory cortex (anterior strip electrode contact in Figure 3a) and hippocampus (Figure 3d and e). Among the participants that had both LFP and temporo-parietal ECoG recordings, Participant 3 had an electrode contact over left visual cortex; the sensory localization was indexed by the strong gamma activity in the most posterior contact of the strip electrode (Figure 3g). The time course of information flow between visual cortex and hippocampus (Figure 3i and j) followed the same pattern as described for the auditory cortex above. Interestingly, the pattern appeared with LFP recorded from right hippocampus in Participant 3 (Supplementary file 1). However, in Participant 4, the recordings from the right cortical hemisphere (Figure 3k) did not show significant GC between LFP and ECoG during task performance (Figure 3n and o).

Thus, we showed in recordings from the left cortical hemisphere that letters were encoded with information flow from sensory cortex to hippocampus; conversely, the information flow from hippocampus to sensory cortex indicated the replay of letters during maintenance.

Source reconstruction of the scalp EEG

We used beamforming (Oostenveld et al., 2011) to reconstruct the EEG sources during encoding and maintenance for each of the 15 participants (Table 1). We tested whether the sources during fixation differed from sources during encoding and during maintenance (non-parametric cluster-based permutation t-test Maris and Oostenveld, 2007; Popov et al., 2018). In each participant, the proportion of significant sources in the left hemisphere exceeded 80% of all significant sources. Across all participants, the spatial activity pattern during both encoding and maintenance showed the highest significance in frontal and temporal areas of the left hemisphere (Figure 4—figure supplement 1).

Directed functional coupling between hippocampus and averaged EEG sources

The synchronization between hippocampal LFP and EEG sources (N=15 participants) confirmed the directed functional coupling found in the three participants with ECoG. We first calculated the GC between hippocampus and the EEG beamforming sources in the auditory cortex. We found that the mean GC spectra resembled the GC spectrum for ECoG in the theta frequency range ([4 8] Hz, Figure 4a). During encoding, the net information flow was from auditory cortex to hippocampus (light blue curve – dark blue curve, blue bar, p<0.05, group cluster-based permutation test). During maintenance, the net information flow was reversed (dark red curve – light red curve, red bar, p<0.05, group cluster-based permutation test), i.e., from hippocampus to auditory cortex. Interestingly, the pattern appeared with LFP recorded from the right hippocampus in several participants (Supplementary file 1). A similar GC pattern emerged when using the signals from left temporal scalp electrodes but was eliminated when using a Laplacian derivation. Thus, both for ECoG and EEG beamforming sources, GC showed the same bidirectional effect in theta between auditory cortex and hippocampus.

Figure 4. Granger causality (GC) between hippocampal local field potentials (LFP) and EEG sources.

(a) Spectral GC between hippocampal LFP and auditory EEG sources, averaged over all N=15 participants. The shaded area indicates the variability across the population. During encoding, the net Granger (ΔGranger) indicates information flow from auditory cortex to hippocampus ([6 10] Hz, blue bar). During maintenance, ΔGranger indicates information flow from hippocampal LFP to auditory cortex (red bars, [6 9] Hz, [13 15] Hz). Bars: frequency range of significant ΔGranger (p<0.05), group cluster-based non-parametric permutation t-test against a null distribution with scrambled trials during encoding and maintenance. Colors of Granger spectra indicate information flow: dark blue, cortex to hippocampus during encoding; light blue, hippocampus to cortex during encoding; dark red, hippocampus to cortex during maintenance; light red, cortex to hippocampus during maintenance. (b) The median net information flow (ΔGranger) in the [4 8] Hz range during encoding is projected onto an inflated brain surface. The maximal ΔGranger appeared from temporal superior gyrus (median ΔGranger=–0.049) indicating information flow from auditory cortex to hippocampus. Negative values of median ΔGranger appeared also in other areas, albeit less intense (ΔGranger>–0.03). (c) The median net information flow (ΔGranger) in the [4 8] Hz range during maintenance is projected onto an inflated brain surface. The maximal ΔGranger appeared from temporal superior gyrus (median ΔGranger=0.034) indicating an information flow from hippocampus to auditory cortex. Positive values of median ΔGranger appeared also in other areas, albeit less intense (ΔGranger <0.02). (d) The maximal ΔGranger in the [4 8] Hz range was negative during encoding (blue, auditory cortex → hippocampus, median ΔGranger=–0.049) and positive during maintenance (red, hippocampus → auditory cortex, median ΔGranger=0.034) for each participant (red and blue connected marker, paired permutation test, correct trials only). The mean values and statistical significance derive only from 10% of the correct trials in order to balance the number of incorrect trials. (e) The net information flow between hippocampal LFP and lateral prefrontal cortex in the [4 8] Hz range has a lower median than to auditory cortex and higher variability (correct trials only, p=0.16, paired permutation test, not significant). (f) For incorrect trials, the maximal ΔGranger in the [4 8] Hz range is highly variable (p=0.37, paired permutation test, not significant). (g) Bidirectional information flow between cortical sites and hippocampus in the working memory network. The GC analysis suggests a surprisingly simple model of information flow during the task. During encoding, letter strings are verbalized as subvocal speech; the incoming information flows from auditory cortex to hippocampus. During maintenance, participants actively recall and rehearse the subvocal speech in the phonological loop; GC indicates an information flow from hippocampus to cortex as the physiological basis for the replay of the memory items.

Figure 4.

Figure 4—figure supplement 1. Spatial activation pattern of EEG beamforming sources.

Figure 4—figure supplement 1.

(a) The area of significant activation (t-value>8) during encoding compared to fixation is averaged for the group of participants and is projected onto an inflated brain surface. The most significant increase appears on sources over the left lateral prefrontal cortex. The spatial activation pattern at the cortical level spreads mostly over the left hemisphere (left frontal area, temporal pole, temporal superior gyrus, and Heschl gyrus). On the right hemisphere, there is only a small orbitofrontal activation. (b) The area of significant activation (t-value>8) during maintenance compared to fixation is projected onto an inflated brain surface. The most significant increase appears on sources over the left temporal superior gyrus (auditory cortex). The spatial activation pattern at the cortical level spreads mostly over the left hemisphere (left frontal area, temporal pole, temporal superior gyrus, and Heschl gyrus). On the right hemisphere, an activation appears on premotor/motor cortex. The spatial activation pattern derives from a non-parametric cluster-based permutation t-test (N=1000 permutations, significance established at t>1.96 and p<0.05). The activation map is thresholded at the 80% of the maximal t-value. Blue colorbar: encoding, red colorbar: maintenance.

To explore the spatial distribution, we computed GC also for other areas of cortex. We averaged the net information flow (ΔGranger) in the theta range across the participants and projected it onto the inflated brain surface (Figure 4b and c). During encoding, the mean information flow was strongest from auditory cortex to hippocampus (ΔGranger=−0.049, p=0.0009, Kruskal-Wallis test, Figure 4b). For all other areas, the mean ΔGranger was also from cortex to hippocampus but the effect was weaker (mean ΔGranger = [–0.03 0], Dunn’s test, Bonferroni corrected). During maintenance (Figure 4c) the information flow was reversed. While all areas had information flow from hippocampus to cortex (ΔGranger = [0.02], Dunn’s test, Bonferroni corrected), the strongest flow appeared from hippocampus to auditory cortex (ΔGranger=0.034, p=0.001, Kruskal-Wallis test).

Directed functional coupling and the participants’ performance

The reversal of ΔGranger appeared in all 15 participants individually (Figure 4d). We averaged ΔGranger for each participant in the [4 8] Hz theta frequency range. The ΔGranger between hippocampus and auditory cortex, was negative during encoding and was positive during maintenance in the theta frequency range (p=4.1e-10, paired permutation test). The directionality and its reversal was missing for all other areas, e.g., lateral prefrontal cortex (p=0.16, paired permutation test, Figure 4e). Of note, all analyses up to here were performed on correct trials only.

Finally, we established a link between the participants’ performance and ΔGranger. For incorrect trials, the net information flow ΔGranger from auditory cortex to hippocampus did not show the same directionality in all participants and did not reverse in direction (p=0.37, paired permutation test, Figure 4f). Since participants performed well (median performance 91%), we balanced the numbers of correct and incorrect trials. We calculated the GC in a subset of correct trials (median of 200 permutations of a number of correct trials that equals the mean percentage of incorrect trials=10%); the effect was equally present for the subset of correct trials (p<0.0005). This suggests that timely information flow, as indexed by GC, is relevant for producing a correct response.

Discussion

WM describes our capacity to represent sensory input for prospective use. Our findings suggest that this cognitive function is subserved by bidirectional oscillatory interactions between memory neurons in the hippocampus and sensory neurons in the auditory cortex as indicated by phase synchrony and GC. In our verbal WM task, the encoding of letter items is isolated from the maintenance period in which the active rehearsal of memory items is central to achieve correct performance. First, analysis of task-induced power showed sustained oscillatory activity in cortical and hippocampal sites during the maintenance period. Second, analysis of the inter-electrode phase synchrony and the directional information flow showed task-induced interactions in the theta band between cortical and hippocampal sites. Third, the directional information flow was from auditory cortex to hippocampus during encoding, and during maintenance, the reverse flow occurred from hippocampus to auditory cortex. This pattern was found only to the left cortical hemisphere, as expected for a language-related task. Fourth, the comparison between correct and incorrect trials suggests that the participants relied on timely information flow to produce a correct response. Our data suggests a surprisingly simple model of information flow within a network that involves sensory cortices and hippocampus (Figure 4g): during encoding, letter strings are verbalized as subvocal speech. The incoming information flows from sensory cortex to hippocampus (bottom-up). During maintenance, participants actively recall and rehearse the subvocal speech in their phonological loop (Baddeley, 2003; Christophel et al., 2017). The GC indicates the information flow from hippocampus to cortex (top-down) as the physiological basis for the replay of the memory items, which finally guides action.

The current study is embedded in previous studies using the same or similar tasks. Persistent firing of hippocampal neurons indicated hippocampal involvement in the maintenance of memory items (Boran et al., 2019; Kamiński et al., 2017; Kornblith et al., 2017). An fMRI study reports salient activity in the auditory cortex during maintenance in an auditory WM task (Kumar et al., 2016), which indicates that sensory cortical areas are involved in the maintenance of WM items. During encoding, the activity of local assemblies was associated with gamma frequencies and local processing (Figure 2a b c, Figure 3g l) while GC inter-areal interactions took place in theta frequencies, in line with previous reports (Solomon et al., 2017; von Stein and Sarnthein, 2000). Parietal generators of theta-alpha EEG indicated involvement of parietal cortex in WM maintenance (Michels et al., 2008; Tuladhar et al., 2007; Näpflin et al., 2008; Boran et al., 2019; Boran et al., 2020). The hippocampo-cortical phase synchrony (PLV) was high during maintenance of the high workload trials (Boran et al., 2019). Building on these previous studies, the current study focused on high workload trials and extended them by the analysis of directional information flow.

In the design of the task, we aimed to separate in time the encoding of memory items from their maintenance. In the choice of the 2 s duration for the encoding period were guided by the magic number 7±2, which may correspond to ‘how many items we can utter in 2 s’(Baddeley, 2003; Christophel et al., 2017). The median Cowan’s K=6.1 shows that high workload trials were indeed demanding for the participants, where both encoding and maintenance may limit performance. We therefore presented the letters both as a visual and an auditory stimulus. Certainly, maintenance processes are likely to appear already during the encoding period as maintenance neurons ramp up their activity already during encoding (Boran et al., 2019). Furthermore, encoding may extend past the visual stimulus (t=–3 s). We therefore focused our analysis on the last 2 s of maintenance [–2 0] s. With this task design, we found patterns of GC that were clearly distinct between encoding and maintenance.

Our study capitalizes on a unique dataset. We first benefitted from direct cortical recordings that assured the neuronal origin of the signals. We then confirmed the GC results by using the wide spatial coverage of scalp EEG, where we used beamforming to localize the cortical sources that generate the scalp EEG. The interaction between recordings from different brain regions has to be discussed with respect to volume conduction (Trongnetrpunya et al., 2015). On the recording level, the choice of two separate references for LFP and ECoG has been shown to avoid spurious effects in GC (Bastos and Schoffelen, 2015). On the level of scalp EEG analysis, we used beamforming as a source reconstruction technique (Popov et al., 2018) to characterize the primary neuronal generators that were localized specifically in left auditory cortex. A similar GC pattern emerged when using the signals from left temporal scalp electrodes but it was eliminated when using a Laplacian derivation. When looking at the GC frequency spectra, there was a strong frequency dependence of GC from hippocampus to ECoG (Figure 2h, Figure 3d i). Likewise, GC to EEG sources showed a strong frequency dependence (Figure 4a). This speaks against volume conduction because the transfer of signal through tissue by volume conduction is independent of frequency in the range of interest here (Miceli et al., 2017). Finally, there was a strong task dependence of GC (Figure 2h, Figure 3d i, Figure 4a d), again speaking against a strong contribution of volume conduction.

In the literature, there are several studies investigating the WM network. However, only few report directional interactions. One of these (Johnson et al., 2018a) reports cross-spectral directionality between intracranial recordings in frontal cortex and the medial temporal lobe in theta frequencies. One study on episodic memory suggests directional information flow to and from hippocampus (Griffiths et al., 2019). Within hippocampus, directional information flow from posterior to anterior hippocampus indicated successful WM maintenance (Li et al., 2022). The frequencies of GC found in the current study were in the ([4 8] Hz) theta band, in line with scalp EEG findings during WM tasks (Sarnthein et al., 1998; Polanía et al., 2012) and other tasks (Solomon et al., 2017) that activate oscillations in long-range recurrent connections (Fries, 2015; Pesaran et al., 2018).

Taken together, our results corroborated earlier findings on the WM network and extended them by providing a physiological mechanism for the active replay of memory items.

Materials and methods

Task

We used a modified Sternberg task in which the encoding of memory items and their maintenance was temporally separated (Figure 1a). Each trial started with a fixation period ([−6, –5] s), followed by the stimulus ([−5, –3] s). The stimulus consisted of a set of eight consonants at the center of the screen. The middle four, six, or eight letters were the memory items, which determined the set size for the trial (4, 6, or 8 respectively). The outer positions were filled with ‘X’, which was never a memory item. The participants read the letters and heard them spoken at the same time. After the stimulus, the letters disappeared from the screen, and the maintenance interval started ([−3, 0] s). Since the auditory encoding may have extended beyond the 2 s period, we restrict our analysis to the last 2 s of the maintenance period ([−2, 0] s). A fixation square was shown throughout fixation, encoding, and maintenance. After maintenance, a probe was presented. The participants responded with a button press to indicate whether the probe was part of the stimulus. The participants were instructed to respond as rapidly as possible without making errors. After the response, the probe was turned off, and the participants received acoustic feedback regarding whether the response was correct or incorrect. The participants performed sessions of 50 trials in total, which lasted approximately 10 min each. Trials with different set sizes were presented in a random order, with the single exception that a trial with an incorrect response was always followed by a trial with a set size of 4. The task can be downloaded at http://www.neurobs.com/ex_files/expt_view?id=266.

Participants

The participants in the study were patients with drug resistant focal epilepsy. To investigate a potential surgical treatment of epilepsy, the patients were implanted with intracranial electrodes. Electrodes were placed according to the findings of the non-invasive presurgical evaluation, where the epileptologists hypothesized the epileptic foci to be localized (Zijlmans et al., 2019). Since the presumed epileptic foci included the hippocampus in all patients, electrodes were placed in the hippocampus. In four patients, additional electrodes were placed on the cortex because an epileptic focus in the cerebral cortex was considered. The participants provided written informed consent for the study, which was approved by the institutional ethics review board (PB 2016–02055). The participants were right handed and had normal or corrected-to-normal vision. For nine participants (5–14), the PSD and PLV have been reported in an earlier study (Boran et al., 2019).

Electrodes for LFP, ECoG, and EEG

The depth electrodes (1.3 mm diameter, eight contacts of 1.6 mm length, spacing between contact centers 5 mm, Ad-Tech, adtechmedical.com) were stereotactically implanted into the hippocampus for LFP recording. Subdural grid and strip electrodes (platinum electrode contacts with 4 mm2 contact surface and 1 cm inter-contact distance, Ad-Tech) were placed directly on the cortex for ECoG recordings. For scalp EEG recording, cup electrodes (Ag/AgCl) were placed on the scalp and filled with electrolyte gel (Signagel, Parker Laboratories) to obtain an impedance <5 kΩ.

Electrode localization

The stereotactic depth electrodes were localized using post-implantation CT and post-implantation structural T1-weighted MRI scans. The CT scan was registered to the post-implantation scan as implemented in FieldTrip (Stolk et al., 2018). A fused image of CT and MRI scans was produced and the electrode contacts were marked visually. The position of the most distal hippocampal contact was projected in a hippocampal surface (Figure 1d, Figure S1).

To localize the ECoG grids and strips, we used the participants’ postoperative MRI, aligned to CT and produced a 3D reconstruction of the participants’ pial brain surface. Grid and strip electrode coordinates were projected on the pial surface as described in Groppe et al., 2017; Figure 2a and Figure 3a and f.

The scalp EEG electrodes were placed at the sites of the 10–20 system by experienced technicians and no further localization was performed. While the 10–20 standard is 21 scalp electrodes, in some patients some electrode sites stayed vacant to assure the sterility of the leads to the intracranial electrodes, resulting in a median of 17 scalp sites per patient.

Some of the intracranial electrode contacts were found in tissue that was deemed to be epileptogenic and that was later resected. Still, neurons in this tissue have been found to participate in task performance in an earlier study (Boran et al., 2019).

Recording setup, re-referencing, and preprocessing

All recordings (LFP, ECoG, and scalp EEG) were performed with the Neuralynx ATLAS system (sampling rate 4000 Hz, 0.5 1000 Hz passband, Neuralynx, neuralynx.com). ECoG and LFP were recorded against a common intracranial reference. Signals were analyzed in MATLAB (Mathworks, Natick MA, USA). We re-referenced the hippocampal LFP against the signal of a depth electrode contact in white matter. We re-referenced the cortical ECoG against a different depth electrode contact. The choice of two separate references for LFP and ECoG has been shown to avoid spurious functional connectivity estimates (Bastos and Schoffelen, 2015). The scalp EEG was recorded against an electrode near the vertex and was then re-referenced to the averaged mastoid channels. All signals were downsampled to 500 Hz. All recordings were done at least 6 hr from a seizure. Trials with large unitary artifacts in the scalp EEG were rejected. We focused on the trials with high workload (set sizes 6 and 8) for further analysis. We used the FieldTrip toolbox for data processing and analysis (Oostenveld et al., 2011).

Power spectral density

We first calculated the relative PSD in the time-frequency domain (Figure 2b). Time-frequency maps for all trials were averaged. We used 3 multitapers with a window width of 10 cycles per frequency point, smoothed with 0.2×frequency. We computed power in the frequency range [4 100] Hz with a time resolution of 0.1 s. The PSD during fixation ([−6 –5] s) served as a baseline for the baseline correction (PSD[t] – PSD[fixation])/ PSD(fixation) for each time-frequency point.

Phase-locking value

To evaluate the functional connectivity of hippocampus and cortex, we calculated the PLV between hippocampal LFP channels and ECoG grid (multitaper frequency transformation with two tapers based on Fourier transform, frequency range [4 100] Hz with frequency resolution of 1 Hz).

PLVi,j(f)=1N|n=1NXi(f)(Xj(f))|Xi(f)||Xj(f)||

where PLVi,j is the PLV between channels i and j, N is the number of trials, X(f) is the Fourier transform of x(t), and (∙)* represents the complex conjugate.

Using the spectra of the 2-s epochs, phase differences were calculated for each electrode pair (i,j) to quantify the inter-electrode phase coupling. The phase difference between the two signals indexes the coherence between each electrode pair and is expressed as the PLV. The PLV ranges between 0 and 1, with values approaching 1 if the two signals show a constant phase relationship over all trials.

In our description of EEG frequency bands, we used theta [4 8] Hz, alpha [8 12] Hz, beta [12 24] Hz, and gamma >40 Hz, while the exact frequencies may differ in individual participants.

Source reconstruction of the EEG sources

We reconstructed the scalp EEG sources using linearly constrained minimum variance (LCMV) beamformers in the time domain. To solve the forward problem, we used a precomputed head model template and aligned the EEG electrodes of each participant to the scalp compartment of the model via interactive scaling, translation, and rotation (ft_electrode_realign.m). We then computed the source grid model and the leadfield matrix, wherein we determined the grid locations according to the brain parcels of the automated anatomical atlas (AAL) (Tzourio-Mazoyer et al., 2002). We solved the inverse problem by scanning the grid locations using the LCMV filters separately for encoding and maintenance. The EEG sources were baselined with respect to the fixation period and presented as a percent of change from the pre-stimulus baseline. We defined cortical areas from multiple parcels since AAL is a parcellation based on sulci and gyri. We performed all the steps of the source reconstruction with FieldTrip (Oostenveld et al., 2011) and projected the sources onto an inflated brain surface.

Spectral Granger causality

In order to evaluate the direction of information flow between the hippocampus and the cortex, we calculated spectral non-parametric GC as a measure of directed functional connectivity analysis (Oostenveld et al., 2011). We evaluated the direction of information flow in the (Sarnthein et al., 1998; Li et al., 2022; [4 20]) Hz frequency range. To compute the GC, we first downsampled the signals to the Nyquist frequency=40 Hz. We then computed the GC between hippocampal contacts and ECoG grid contacts. We also computed GC between the same hippocampal contacts and EEG sources located over the regions of interest. GC examines if the activity on one channel can forecast activity in the target channel. In the spectral domain, GC measures the fraction of the total power that is contributed by the source to the target. We transformed signals to the frequency domain using the multitaper frequency transformation method (two Hann tapers, frequency range [4 20] Hz with 20 s padding) to reduce spectral leakage and control the frequency smoothing.

We used a non-parametric spectral approach to measure the interaction in the channel pairs at a given interval time (Bastos and Schoffelen, 2015). In this approach, the spectral transfer matrix is obtained from the Fourier transform of the data. We used the FieldTrip toolbox to factorize the transfer function H(f) and the noise covariance matrix Σ. The transfer function and the noise covariance matrix were then employed to calculate the total and the intrinsic power, S(f)=H(f)ΣH×(f), through which we calculated the Granger interaction in terms of power fractions contributed from the source to the target.

GCYX→=lnSxx(f)Sxx(f)

where Sxx(f) is the total power and S~xx(f) is the instantaneous power. To average over the group of participants, we calculated the Granger spectra for the selected channel pairs and averaged these spectra over participants (Figure 4a).

To illustrate the time course of GC over time, we calculated time-frequency maps with the multitaper convolution method of Fieldtrip (Oostenveld et al., 2011).

Statistics

To analyze statistical significance, we used cluster-based non-parametric permutation tests. To assess the significance of the difference of the Granger between different directions, we compared the difference of the true values to a null distribution of differences. We recomputed GC after switching directions randomly across trials, while keeping the trial numbers for both channels constant. Then we computed the difference of GC for the two conditions. We repeated this n=200 times to create a null distribution of differences. The null distribution was exploited to calculate the percentile threshold p=0.05. In this way, we compare the difference of the dark and light spectra against a null distribution of differences. We mark the frequency range of significant GC with a blue bar for encoding (dark blue spectrum exceeds light blue spectrum, information flow from cortex to hippocampus) and with a red bar for maintenance (dark red spectrum exceeds light red spectrum, information flow from hippocampus to cortex).

To test the statistical significance of the spatial spread of contacts with high PSD, PLV, or ΔGranger, we calculated the spatial collinearity on the grid contacts against a null distribution. First, we transform the activation on grid contacts into a grid vector. We then performed 200 iterations of random trial permutations. For each iteration, we selected two subsets (50%) of trials and we calculated the scalar product between the vectors corresponding to the two subsets. The null distribution was created by randomly mixing trials from the two task periods fixation and encoding. We finally tested the statistical significance of the scalar product. The true distribution was established to be statistically distinct from the null distribution if it exceeded the 95th percentile of the null distribution.

We assess if the reconstructed EEG sources during encoding and maintenance are significantly different from the pre-stimulus baseline (fixation). We use the FieldTrip’s method ft_sourcestatistics (Oostenveld et al., 2011), wherein we apply a non-parametric permutation approach to quantify the spatial activation pattern during the encoding of the memory items and their active replay.

Due to high average performance of the participants (91%) the number of correct and incorrect trials is imbalanced. To balance the number of correct trials with the number of incorrect trials, we randomly selected 10% of the correct trials and recomputed the GC spectra and then the net information flow (ΔGranger). We repeated this n=200 times and presented the mean ΔGranger for each participant.

For comparisons between two groups, we used the non-parametric paired cluster-based permutation test. We created a null distribution by performing N=200 random permutations.

To test the directionality of the information flow in the group of the participants, we used the group cluster-based permutation t-test from the FieldTrip toolbox (Oostenveld et al., 2011) with multiple comparison correction using the false discovery rate approach. Statistical significance was established at p<0.05.

Acknowledgements

We thank the physicians and the staff at Schweizerische Epilepsie-Klinik for their assistance and the patients for their participation. We acknowledge grants awarded by the Swiss National Science Foundation (funded by SNSF 204651 to JS) and SNSF Ambizione fellowship (PZ00P3_167836 to PM) and a scholarship by Alexander S Onassis Foundation (to VD). The funders had no role in the design or analysis of the study.

Funding Statement

The funders had no role in study design, data collection and interpretation, or the decision to submit the work for publication.

Contributor Information

Johannes Sarnthein, Email: johannes.sarnthein@usz.ch.

Timothy D Griffiths, University of Newcastle, United Kingdom.

Laura L Colgin, University of Texas at Austin, United States.

Funding Information

This paper was supported by the following grants:

  • Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung funded by SNSF 204651 to Johannes Sarnthein.

  • Schweizerischer Nationalfonds zur Förderung der Wissenschaftlichen Forschung PZ00P3_167836 to Pierre Mégevand.

  • Alexander S. Onassis Public Benefit Foundation - to Vasileios Dimakopoulos.

Additional information

Competing interests

No competing interests declared.

No competing interests declared.

Author contributions

Investigation, Visualization, Methodology, Writing – original draft, Writing – review and editing.

Visualization.

Investigation.

Investigation.

Conceptualization, Supervision, Investigation, Methodology, Writing – original draft, Writing – review and editing.

Ethics

Human subjects: The participants provided written informed consent for the study, which was approved upfront by the institutional ethics review board (PB 2016-02055).

Additional files

MDAR checklist
Supplementary file 1. ECoG and LFP recording locations.

For each participant, we list the coordinates of the tip of the hippocampal electrode for LFP recording. In addition, in Participants 1–4, grid or strip electrodes were placed for ECoG recording. LFP: local field potential; ECoG: electrocorticography.

elife-78677-supp1.xlsx (29.1KB, xlsx)

Data availability

All data needed to evaluate the conclusions in the paper are present in the paper. The codes used to produce the results in the paper are freely available at the repository https://github.com/vdimakopoulos/verbal_working_memory. The task can be downloaded at http://www.neurobs.com/ex_files/expt_view?id=266. Part of the data has been published earlier [36]. Additional data and code are indexed in http://www.hfozuri.ch/.

The following previously published dataset was used:

Boran E, Fedele T, Steiner A, Hilfiker P, Stieglitz L, Grunwald T, Sarnthein J. 2020. Dataset of human medial temporal lobe neurons, scalp and intracranial EEG during a verbal working memory task. Scientific Data.

References

  1. Baddeley A. Working memory: looking back and looking forward. Nature Reviews. Neuroscience. 2003;4:829–839. doi: 10.1038/nrn1201. [DOI] [PubMed] [Google Scholar]
  2. Bastos AM, Schoffelen JM. A tutorial review of functional connectivity analysis methods and their interpretational pitfalls. Frontiers in Systems Neuroscience. 2015;9:175. doi: 10.3389/fnsys.2015.00175. [DOI] [PMC free article] [PubMed] [Google Scholar]
  3. Bidelman GM, Brown JA, Bashivan P. Auditory cortex supports verbal working memory capacity. Neuroreport. 2021;32:163–168. doi: 10.1097/WNR.0000000000001570. [DOI] [PMC free article] [PubMed] [Google Scholar]
  4. Boran E, Fedele T, Klaver P, Hilfiker P, Stieglitz L, Grunwald T, Sarnthein J. Persistent hippocampal neural firing and hippocampal-cortical coupling predict verbal working memory load. Science Advances. 2019;5:eaav3687. doi: 10.1126/sciadv.aav3687. [DOI] [PMC free article] [PubMed] [Google Scholar]
  5. Boran E, Fedele T, Steiner A, Hilfiker P, Stieglitz L, Grunwald T, Sarnthein J. Dataset of human medial temporal lobe neurons, scalp and intracranial EEG during a verbal working memory task. Scientific Data. 2020;7:30. doi: 10.1038/s41597-020-0364-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
  6. Boran E, Hilfiker P, Stieglitz L, Sarnthein J, Klaver P. Persistent neuronal firing in the medial temporal lobe supports performance and workload of visual working memory in humans. NeuroImage. 2022;254:119123. doi: 10.1016/j.neuroimage.2022.119123. [DOI] [PubMed] [Google Scholar]
  7. Christophel TB, Klink PC, Spitzer B, Roelfsema PR, Haynes JD. The distributed nature of working memory. Trends in Cognitive Sciences. 2017;21:111–124. doi: 10.1016/j.tics.2016.12.007. [DOI] [PubMed] [Google Scholar]
  8. Cogan GB, Iyer A, Melloni L, Thesen T, Friedman D, Doyle W, Devinsky O, Pesaran B. Manipulating stored phonological input during verbal working memory. Nature Neuroscience. 2017;20:279–286. doi: 10.1038/nn.4459. [DOI] [PMC free article] [PubMed] [Google Scholar]
  9. Fries P. Rhythms for cognition: Communication through coherence. Neuron. 2015;88:220–235. doi: 10.1016/j.neuron.2015.09.034. [DOI] [PMC free article] [PubMed] [Google Scholar]
  10. Griffiths BJ, Parish G, Roux F, Michelmann S, van der Plas M, Kolibius LD, Chelvarajah R, Rollings DT, Sawlani V, Hamer H, Gollwitzer S, Kreiselmeyer G, Staresina B, Wimber M, Hanslmayr S. Directional coupling of slow and fast hippocampal gamma with neocortical alpha/beta oscillations in human episodic memory. PNAS. 2019;116:21834–21842. doi: 10.1073/pnas.1914180116. [DOI] [PMC free article] [PubMed] [Google Scholar]
  11. Groppe DM, Bickel S, Dykstra AR, Wang X, Mégevand P, Mercier MR, Lado FA, Mehta AD, Honey CJ. iELVis: An open source MATLAB toolbox for localizing and visualizing human intracranial electrode data. Journal of Neuroscience Methods. 2017;281:40–48. doi: 10.1016/j.jneumeth.2017.01.022. [DOI] [PubMed] [Google Scholar]
  12. Hsieh L-T, Ranganath C. Frontal midline theta oscillations during working memory maintenance and episodic encoding and retrieval. NeuroImage. 2014;85 Pt 2:721–729. doi: 10.1016/j.neuroimage.2013.08.003. [DOI] [PMC free article] [PubMed] [Google Scholar]
  13. Johnson EL, Adams JN, Solbakk AK, Endestad T, Larsson PG, Ivanovic J, Meling TR, Lin JJ, Knight RT. Dynamic frontotemporal systems process space and time in working memory. PLOS Biology. 2018a;16:e2004274. doi: 10.1371/journal.pbio.2004274. [DOI] [PMC free article] [PubMed] [Google Scholar]
  14. Johnson EL, King-Stephens D, Weber PB, Laxer KD, Lin JJ, Knight RT. Spectral imprints of working memory for everyday associations in the frontoparietal network. Frontiers in Systems Neuroscience. 2018b;12:65. doi: 10.3389/fnsys.2018.00065. [DOI] [PMC free article] [PubMed] [Google Scholar]
  15. Kamiński J, Sullivan S, Chung JM, Ross IB, Mamelak AN, Rutishauser U. Persistently active neurons in human medial frontal and medial temporal lobe support working memory. Nature Neuroscience. 2017;20:590–601. doi: 10.1038/nn.4509. [DOI] [PMC free article] [PubMed] [Google Scholar]
  16. Kornblith S, Quian Quiroga R, Koch C, Fried I, Mormann F. Persistent single-neuron activity during working memory in the human medial temporal lobe. Current Biology. 2017;27:1026–1032. doi: 10.1016/j.cub.2017.02.013. [DOI] [PMC free article] [PubMed] [Google Scholar]
  17. Kumar S, Joseph S, Gander PE, Barascud N, Halpern AR, Griffiths TD. A brain system for auditory working memory. The Journal of Neuroscience. 2016;36:4492–4505. doi: 10.1523/JNEUROSCI.4341-14.2016. [DOI] [PMC free article] [PubMed] [Google Scholar]
  18. Kumar S, Gander PE, Berger JI, Billig AJ, Nourski KV, Oya H, Kawasaki H, Howard MA, Griffiths TD. Oscillatory correlates of auditory working memory examined with human electrocorticography. Neuropsychologia. 2021;150:107691. doi: 10.1016/j.neuropsychologia.2020.107691. [DOI] [PMC free article] [PubMed] [Google Scholar]
  19. Leszczyński M, Fell J, Axmacher N. Rhythmic working memory activation in the human hippocampus. Cell Reports. 2015;13:1272–1282. doi: 10.1016/j.celrep.2015.09.081. [DOI] [PubMed] [Google Scholar]
  20. Li J, Cao D, Dimakopoulos V, Shi W, Yu S, Fan L, Stieglitz L, Imbach L, Sarnthein J, Jiang T. Anterior-posterior hippocampal dynamics support working memory processing. The Journal of Neuroscience. 2022;42:443–453. doi: 10.1523/JNEUROSCI.1287-21.2021. [DOI] [PMC free article] [PubMed] [Google Scholar]
  21. Maris E, Oostenveld R. Nonparametric statistical testing of EEG- and MEG-data. Journal of Neuroscience Methods. 2007;164:177–190. doi: 10.1016/j.jneumeth.2007.03.024. [DOI] [PubMed] [Google Scholar]
  22. Maris E, van Vugt M, Kahana M. Spatially distributed patterns of oscillatory coupling between high-frequency amplitudes and low-frequency phases in human iEEG. NeuroImage. 2011;54:836–850. doi: 10.1016/j.neuroimage.2010.09.029. [DOI] [PubMed] [Google Scholar]
  23. Miceli S, Ness TV, Einevoll GT, Schubert D. Impedance spectrum in cortical tissue: Implications for propagation of lfp signals on the microscopic level. ENeuro. 2017;4:ENEURO.0291-16.2016. doi: 10.1523/ENEURO.0291-16.2016. [DOI] [PMC free article] [PubMed] [Google Scholar]
  24. Michels L, Moazami-Goudarzi M, Jeanmonod D, Sarnthein J. EEG alpha distinguishes between cuneal and precuneal activation in working memory. NeuroImage. 2008;40:1296–1310. doi: 10.1016/j.neuroimage.2007.12.048. [DOI] [PubMed] [Google Scholar]
  25. Näpflin M, Wildi M, Sarnthein J. Test-retest reliability of EEG spectra during a working memory task. NeuroImage. 2008;43:687–693. doi: 10.1016/j.neuroimage.2008.08.028. [DOI] [PubMed] [Google Scholar]
  26. Oostenveld R, Fries P, Maris E, Schoffelen JM. FieldTrip: Open source software for advanced analysis of MEG, EEG, and invasive electrophysiological data. Computational Intelligence and Neuroscience. 2011;2011:156869. doi: 10.1155/2011/156869. [DOI] [PMC free article] [PubMed] [Google Scholar]
  27. Pavlov YG, Kotchoubey B. Oscillatory brain activity and maintenance of verbal and visual working memory: A systematic review. Psychophysiology. 2022;59:e13735. doi: 10.1111/psyp.13735. [DOI] [PubMed] [Google Scholar]
  28. Pesaran B, Vinck M, Einevoll GT, Sirota A, Fries P, Siegel M, Truccolo W, Schroeder CE, Srinivasan R. Investigating large-scale brain dynamics using field potential recordings: analysis and interpretation. Nature Neuroscience. 2018;21:903–919. doi: 10.1038/s41593-018-0171-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
  29. Polanía R, Nitsche MA, Korman C, Batsikadze G, Paulus W. The importance of timing in segregated theta phase-coupling for cognitive performance. Current Biology. 2012;22:1314–1318. doi: 10.1016/j.cub.2012.05.021. [DOI] [PubMed] [Google Scholar]
  30. Popov T, Oostenveld R, Schoffelen JM. Fieldtrip made easy: An analysis protocol for group analysis of the auditory steady state brain response in time, frequency, and space. Frontiers in Neuroscience. 2018;12:711. doi: 10.3389/fnins.2018.00711. [DOI] [PMC free article] [PubMed] [Google Scholar]
  31. Raghavachari S, Kahana MJ, Rizzuto DS, Caplan JB, Kirschen MP, Bourgeois B, Madsen JR, Lisman JE. Gating of human theta oscillations by a working memory task. The Journal of Neuroscience. 2001;21:3175–3183. doi: 10.1523/JNEUROSCI.21-09-03175.2001. [DOI] [PMC free article] [PubMed] [Google Scholar]
  32. Rezayat E, Clark K, Dehaqani M-RA, Noudoost B. Dependence of working memory on coordinated activity across brain areas. Frontiers in Systems Neuroscience. 2021;15:787316. doi: 10.3389/fnsys.2021.787316. [DOI] [PMC free article] [PubMed] [Google Scholar]
  33. Rizzuto DS, Madsen JR, Bromfield EB, Schulze-Bonhage A, Seelig D, Aschenbrenner-Scheibe R, Kahana MJ. Reset of human neocortical oscillations during a working memory task. PNAS. 2003;100:7931–7936. doi: 10.1073/pnas.0732061100. [DOI] [PMC free article] [PubMed] [Google Scholar]
  34. Rutishauser U, Reddy L, Mormann F, Sarnthein J. The architecture of human memory: Insights from human single-neuron recordings. The Journal of Neuroscience. 2021;41:883–890. doi: 10.1523/JNEUROSCI.1648-20.2020. [DOI] [PMC free article] [PubMed] [Google Scholar]
  35. Sarnthein J, Petsche H, Rappelsberger P, Shaw GL, von Stein A. Synchronization between prefrontal and posterior association cortex during human working memory. PNAS. 1998;95:7092–7096. doi: 10.1073/pnas.95.12.7092. [DOI] [PMC free article] [PubMed] [Google Scholar]
  36. Schwiedrzik CM, Sudmann SS, Thesen T, Wang X, Groppe DM, Mégevand P, Doyle W, Mehta AD, Devinsky O, Melloni L. Medial prefrontal cortex supports perceptual memory. Current Biology. 2018;28:R1094–R1095. doi: 10.1016/j.cub.2018.07.066. [DOI] [PubMed] [Google Scholar]
  37. Solomon EA, Kragel JE, Sperling MR, Sharan A, Worrell G, Kucewicz M, Inman CS, Lega B, Davis KA, Stein JM, Jobst BC, Zaghloul KA, Sheth SA, Rizzuto DS, Kahana MJ. Widespread theta synchrony and high-frequency desynchronization underlies enhanced cognition. Nature Communications. 2017;8:1704. doi: 10.1038/s41467-017-01763-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
  38. Stolk A, Griffin S, van der Meij R, Dewar C, Saez I, Lin JJ, Piantoni G, Schoffelen J-M, Knight RT, Oostenveld R. Integrated analysis of anatomical and electrophysiological human intracranial data. Nature Protocols. 2018;13:1699–1723. doi: 10.1038/s41596-018-0009-6. [DOI] [PMC free article] [PubMed] [Google Scholar]
  39. Trongnetrpunya A, Nandi B, Kang D, Kocsis B, Schroeder CE, Ding M. Assessing granger causality in electrophysiological data: Removing the adverse effects of common signals via bipolar derivations. Frontiers in Systems Neuroscience. 2015;9:189. doi: 10.3389/fnsys.2015.00189. [DOI] [PMC free article] [PubMed] [Google Scholar]
  40. Tuladhar AM, ter Huurne N, Schoffelen J-M, Maris E, Oostenveld R, Jensen O. Parieto-occipital sources account for the increase in alpha activity with working memory load. Human Brain Mapping. 2007;28:785–792. doi: 10.1002/hbm.20306. [DOI] [PMC free article] [PubMed] [Google Scholar]
  41. Tzourio-Mazoyer N, Landeau B, Papathanassiou D, Crivello F, Etard O, Delcroix N, Mazoyer B, Joliot M. Automated anatomical labeling of activations in SPM using a macroscopic anatomical parcellation of the MNI MRI single-subject brain. NeuroImage. 2002;15:273–289. doi: 10.1006/nimg.2001.0978. [DOI] [PubMed] [Google Scholar]
  42. van Vugt MK, Schulze-Bonhage A, Litt B, Brandt A, Kahana MJ. Hippocampal gamma oscillations increase with memory load. The Journal of Neuroscience. 2010;30:2694–2699. doi: 10.1523/JNEUROSCI.0567-09.2010. [DOI] [PMC free article] [PubMed] [Google Scholar]
  43. von Stein A, Sarnthein J. Different frequencies for different scales of cortical integration: from local gamma to long range alpha/theta synchronization. International Journal of Psychophysiology. 2000;38:301–313. doi: 10.1016/s0167-8760(00)00172-0. [DOI] [PubMed] [Google Scholar]
  44. Zijlmans M, Zweiphenning W, van Klink N. Changing concepts in presurgical assessment for epilepsy surgery. Nature Reviews. Neurology. 2019;15:594–606. doi: 10.1038/s41582-019-0224-y. [DOI] [PubMed] [Google Scholar]

Editor's evaluation

Timothy D Griffiths 1

The work provides important information about the communication between the auditory cortex and hippocampus during phonological working memory. The results provide crucial insights into the networks involved in this fundamental process. The results are expected to be of broad interest to readers in the fields of working memory and cognitive neuroscience in general.

Decision letter

Editor: Timothy D Griffiths1

In the interests of transparency, eLife publishes the most substantive revision requests and the accompanying author responses.

Decision letter after peer review:

Thank you for submitting your article "Information flows from hippocampus to cortex during replay of verbal working memory items" for consideration by eLife. Your article has been reviewed by 3 peer reviewers, and the evaluation has been overseen by a Reviewing Editor and Laura Colgin as the Senior Editor. The following individual involved in review of your submission has agreed to reveal their identity: Joel Berger (Reviewer #1).

The reviewers have discussed their reviews with one another, and the Reviewing Editor has drafted this to help you prepare a revised submission.

Essential revisions:

The referees all found the work interesting but raise interpretational issues in their detailed comment that all need to be addressed. We have highlighted a number of these below:

1) The data need to be reported less selectively to give readers an idea of the consistency of the effects emphasised.

2) One referee raises the issue of possible volume conduction which can be addressed with re-referencing of electrodes.

3) Many claims are based on descriptive data as opposed to quantification of the effect claimed and there are statistical issues raised including correction for multiple comparisons.

4) The authors need to comment on a more systematic examination of effects in different frequency ranges including the γ range.

If these issues can be addressed the suggested model will be of broad interest, but an extensive revision will be required.

Reviewer #1:

The authors have shown a unique set of recordings, wherein they have collected intracranial data from parietal cortex and hippocampus, as well as scalp EEG in a number of subjects. With this unique advantage, they have examined directionality of connectivity between various regions during a working memory task. Given the growing evidence for the role of hippocampus in working memory, understanding its connectivity to the rest of the brain provides a crucial insight into the network involved in such a fundamental process. Whilst the existing content is generally of a high standard, and the analyses seem sound, there are some areas of considerable brevity that would benefit from expansion. Below are my comments on the manuscript.

Discussion: This is surprisingly short Discussion section. I feel this should be expanded considerably, such as including some of the information that I have discussed below regarding potential considerations of the task (e.g bimodal nature), discussion of the PLV results in the context of previous non-directional findings, the differences observed between correct and incorrect trials, considering in more detail the other behavioral consequences of these results. These suggestions are not necessarily exhaustive but are all points I believe should be included.

Comments on Results section in general:

– All the results in this section seem to refer to a single electrode for each subject. It would be beneficial to know whether these electrodes were representative of activity from surrounding electrodes or not. That is, how generalizable are the PSD results shown here.

– Also, many of these results are very descriptive. Whilst in some specific scenarios this is unavoidable, for the purposes of reporting results from PSD (for example) it is definitely possible to report details such as the degree of power increase. At present, this reads more like a Discussion section that an informative Results section.

– It would be helpful to see an overlay of the parietal electrode with a topographic map of the scalp EEG recording, to truly appreciate the spatial overlap between the electrode and the generator.

Figures in general: many of the figures appear to refer only to single subjects. It would be useful to have more detailed summary information across subjects to understand how reliable/variable these effects are.

Data availability section: The bit on previously published datasets confused me a little. Is this published dataset included as part of this article? It isn't so clear in the manuscript whether these are previously published data. If they are, this should be made more apparent.

Line 29: Phrasing – I would add the words "rather than sequentially" here to help readers with interpretation of why this separates out encoding from maintenance

Line 64: This can actually extend as low as δ band (see Leszczynski et al., 2015, Cell Reports; Kumar et al., 2021, Neuropsychologia).

Line 88: Do the authors have behavioral data or prior knowledge of how long it takes (on average) to encode 4, 6 or 8 letters? That is, how much of the 'encoding' period is truly encoding, rather than an initial encoding followed by maintenance. Or in a similar manner, how much of maintenance is still residual encoding.

Line 90: Was there a particular reason as to why the encoding phase was bimodal? Do the authors think this may have influenced their results?

Line 94-95: Was this an instruction to the participants? If so, I would put this more explicitly, i.e. "participants were instructed to rehearse…". Of course, one cannot know for certain whether individual subjects employed this strategy.

Figure 2f: Where is this change in Granger relative to? A particular baseline window.

Line 293: Were any electrodes here included in a seizure foci? Was anything done to ensure that seizure activity did not affect recordings (e.g. not recording within xxx hours of a seizure)?

Line 301: Was anything done to deal with artefacts on the ECoG/sEEG electrodes? I.e. were trials with unusually large amplitudes, potentially indicative of muscle artefact (a known contaminant) removed?

Line 325-326: I am confused by this. You say that the individual frequencies may differ between participants – do you mean in terms of the peak frequency, or were different bands used for each subject? If different bands, why?

All power spectral density plots: I assume these are relative to baseline. Are they statistically-thresholded in any way?

Reviewer #2:

Dimakopoulos et al. use intracranial data in humans to ask whether information flow is primarily cortical to hippocampal or the reverse during the encoding and retrieval stages of a working memory task. They find a highly reliable pattern where information in the α/β range flows from auditory cortex to hippocampus during encoding and in the reverse direction during maintenance of items in WM. The authors show this pattern in a sub-selection of ECoG recordings and go on to show it is present in virtually all subjects at the EEG to intracranial hippocampus level. In addition, this directional pattern breaks down during incorrect trials. However, the current analysis suffers from possible contamination by volume conduction.

The study is unique in its data set and provides a valuable look into hippocampal cortical interactions during WM. However, there are multiple technical questions remaining. One of the limitations is that the study investigated primarily interactions in the α/β range when looking at interactions. In contrast, their power spectral results show increases in γ during encoding, and other studies have emphasized a role for γ in feedforward routing. Did the authors perform a granger causality analysis in γ?

1. The reported PLV values (e.g., in figure 2d) of 0.3 – 0.6 are quite high. This level of PLV is usually observed when there is some amount of volume conduction present. Could the authors repeat the PLV and GC analysis using a local bipolar re-referenced signal? For more information on why this is a problem please see Trongnetrpunya, Nandi, Ding, et al., Frontiers in Systems Neuroscience.

2. Did the authors find any detectable GC or PLV in the γ frequency range, particularly during the encoding epoch?

3. Across subjects, the authors report that the Sensory to Hippocampus information flow is dominant during sensory processing and the reverse during delay processing (Figure 5). In Figure 5b, it is impressive how the effect holds in each and every individual subject. My question is how were the data selected in figure 5b? Is that an average of all GC pairs linking all available hippocampus leads to / from all available EEG leads, or was there a data selection that occurred? What would the results look like if only the EEG leads are selected that overly to auditory/temporal cortex? Is the topography the same or different when performing the analysis on e.g., frontal / occipital / parietal EEG leads? In other words, how spatially specific is the effect?

Reviewer #3:

Dimakopoulos and colleagues investigate connectivity and flow of information during encoding and maintenance of Working Memory. They use unique data, which combine human intracranial recordings from depth electrodes with ECOG and EEG. This data, combined with Granger causality analysis (GC), provides interesting results that signal from cortex (mostly from EEG electrodes located over temporal cortex) is flowing to hippocampus during encoding and this flow is reversed during maintenance. Authors interpret this as a sign of bottom-up and top-down processing. I believe that chosen methods for signal analysis are appropriate.

However, paper contains several statements that are unsupported by statistics and there is no clear information about why some decisions in the analysis process were made. This could give an impression that the analysis is built from arbitrarily chosen single case examples. I believe that because of below listed flaws results of the analysis do not support conclusions.

1) Authors do not use correction for multiple comparisons – this cast doubts on the strength of obtained results.

2) There is no criterion given for ECOG electrodes selected to the analysis.

For instance, authors state that for participant 1 for C2 electrode, increased γ power during encoding proves that this electrode was over auditory cortex but there is no systematic analysis of γ power. From the results we can observe that this electrode has the strongest GC with hippocampus what suggests that it was used because of this characteristic what looks like double dipping.

3) Why frequencies observed in PLV and GC are so different? For instance, in supplementary Figure 1 PLV shows significant differences in 18-30 Hz but GC is calculated for 8-18 Hz. Such large differences in frequencies suggest some inconsistencies in the analysis.

4) For analyses depicted in Figure 4 and 5 it is unknown how the highest GC is defined (is it a mean from all frequencies?) Furthermore, there is no systematic measurement or criterion that would support that indeed chosen electrodes have the highest GC.

5) All analysis conducted in the time domain (time to frequency and GC) does not contain any statistics supporting validity of the proposed conclusions.

6) There is no data that supports statement that patients used verbalization. Although material is verbal authors cannot rule out that subject uses different modalities to support information maintenance.

7) In figure 3i significance for encoding (blue) is marked where there are no differences between dark blue and light blue conditions. This suggests that the significance is computed not between conditions but versus null distribution. In general, in the paper, it is unclear if comparison using null distribution or between conditions is being calculated.

8) In Figure 2b – why EEG, and not ECOG, signal is being analyzed?

9) I appreciate that authors present single subject results in Figure 4 but I believe group level analysis should be also performed to support conclusions.

10) Does the quality/magnitude of GC estimation depends on the number of trials? If so

for Figure 5b authors show subsample number of trials in the correct condition to match the number of trials in the incorrect condition.

11) Authors should state how many sessions were used in each comparison.

12) In the statement:

"The GC was lower than for ECoG, as expected for the lower signal amplitude of scalp EEG." I do not know what "lower" means – is it a magnitude or frequency of GC? If frequency, how does it relate to the scalp EEG amplitude?

13) Throughout the paper it seems that "ΔGranger" is used to describe a different thing.

[Editors’ note: the authors submitted for reconsideration following the decision after peer review. What follows is the decision letter after the second round of review.]

Thank you for resubmitting the paper entitled "Information flows from hippocampus to cortex during replay of verbal working memory items" for further consideration by eLife. Your revised article has been evaluated by a Senior Editor and a Reviewing Editor. The revision has also been re-evaluated by three referees. We apologise for the delay in this decision over the holiday period. We are sorry to say that we have decided that this submission will not be considered further for publication by eLife.

We found the approach to working memory based on effective connectivity between sensory cortex and hippocampus during working memory maintenance and retrieval intriguing and novel. All of the referees appreciated the efforts that have been made to improve the manuscript. The principal reason for rejection is the statistical interpretation of the data raised by all three referees, who are concerned whether the analyses allow robust inference about effects in the population. Referee 1 highlights differences between subjects in the frequency bands of effective connectivity. The effects are not robust to re-referencing recommended by referee 2, who does not consider a common reference adequate for this work. Referee 3 is concerned about multiple comparisons.

Reviewer #1:

I appreciate the efforts the authors have taken to improve the manuscript. The additional information does aid interpretation of the study. I still have a couple of remaining core issues to be sorted.

1) In the abstract, it states that information flow from auditory cortex to hippocampus peaked in the range of 8-18 Hz. In the newly inserted (and very helpful) table it seems that the range is anywhere from 4 to 21 Hz. Unless I am missing something, it appears that the authors are still making conclusions based on only Participant 1 (who showed significance in the 8-18 Hz range). I feel that this confusion runs throughout the manuscript and should be corrected – that is, any points where generalizations are made from a single subject that are not truly consistent across subjects should be corrected and the variability should be highlighted/discussed.

2) Similarly, the frequency band of hippocampus > auditory cortex is often not congruent with the AC > hippocampus within subjects, so it is too simplistic to say this was reversed during maintenance. Moreover, this difference in frequencies within subjects should be discussed.

Reviewer #2:

The authors have attempted to answer my concerns by performing a bipolar and Laplacian analysis to the data prior to the connectivity analysis. Unfortunately, this revealed that effects were not preserved after applying local re-referencing. I appreciate the authors' arguments that some effects have a large spatial structure that is removed by bipolar. However, the electrodes used by the authors are quite far apart and a priori I would think that a referencing across a large spatial structure should preserve true effects while removing spurious ones.

Unfortunately, given simulation work as in the article I referred to in my original review (Trongnetrpunya, Nandi, Ding, et al., Frontiers in Systems Neuroscience), the false positive rate is unacceptably high when Granger and related methods are applied to data sharing a common reference. The current results therefore may be incorrect and cannot be interpreted in their present form.

The authors also apparently have cherry picked their data in Figures 4 and 5 by first taking channels that show a change in Granger, and then analyzing those changes across participants. It is like forcing the data to behave as we would like, another example of a flawed approach taken by the present manuscript. Given these technical/methodological issues, I cannot recommend the manuscript for publication in the present state.

I do want to commend the authors for recording this dataset, and performing directed connectivity analysis, which is rare for these types of data. I would encourage the authors to think more deeply about how to assess directed functional connectivity robustly and reformulate their work with this in mind.

Reviewer #3:

The authors have revised their manuscript and attempted to address all of the comments raised by myself and the other reviewers.

I have few remaining issues.

If authors compute the difference against null distribution not between conditions (for example light blue and dark blue) in Figure 3d,i and in Figure 4 – How statement "dark blue spectrum exceeds light blue spectrum, information flow from cortex to hippocampus" is correct? It seems that, for each light colour result (hippocampus to cortex during encoding, cortex to hippocampus during maintenance) there should be a result of a statistical test and it seems none of these tests were performed. The caption of Figure 4 where authors use ΔGranger seems also incorrect as authors compute statistics from null distribution.

Unfortunately, I believe that issue with multiple comparisons was not addressed entirely. For instance, the analysis in the frequency domain does not use any correction for multiple comparisons. Because of that it is unknown if small differences observed in results are not effects of multiple comparisons.

Additionally for Figure 5 b and c analysis it seems that the comparison is not designed correctly. For figure 5b the data is picked using significant reversal of net GC (ΔGranger) – this is not done for Figure 5c analysis so it is hard to draw any conclusions from this contrast.

About the group statistics issue. Figure 5a although presents average results does not show any statistics thus it is unknown if any of those information flows is significant. Figure 4 can give me an intuition that the result is present, but I have no idea about how significant it is on the group level. I believe that paper would benefit from the group statistics.

[Editors’ note: further revisions were suggested prior to acceptance, as described below.]

Thank you for resubmitting your work entitled "Information flows from hippocampus to auditory cortex during replay of verbal working memory items" for further consideration by eLife. Your revised article has been evaluated by Laura Colgin (Senior Editor), a Reviewing Editor, and the original reviewers.

The work provides new insights into the communication between sensory cortex and hippocampus during working memory.

The revised work has been seen by all of the original referees who agree it is substantially improved. They have raised a number of further points about analysis and exposition we would like you to consider when preparing the final article for acceptance.

1. No details are provided for the scalp EEG system. How many electrodes? Was this an active system? Filtering at all? Many of these points are particularly relevant and important for interpreting the beamforming data. It says that the electrodes were aligned for each participant – how was electrode localization performed? Using what system – Polhemus?

2. Was all source reconstruction done with a template MRI? If so, wouldn't it have been appropriate to use each individual's MRI for better localisation, given that you surely had these available? You could then do the group averages, but at least the individual subject data would be more accurately localised.

3. The Granger causality analyses are reported for only three of the subjects between hippocampal LFP and ECoG data. Did the other 12 subjects not have any cortical electrodes then, even without parietal coverage, only hippocampal LFPs? This isn't at all clear from the descriptions in the manuscript and seems surprising that data would only be available from hippocampus, and no other temporal regions, even if parietal coverage was lacking.

4. It would be beneficial to include a display of all the contact locations for hippocampus, for transparency and interpretability purposes, along with (or perhaps instead) a supplement including all MNI coordinates of localised electrodes.

5. L110-110: It would be more consistent with other literature to describe this frequency range as "high α/low β".

6. At a number of points, the authors state that during maintenance, rehearsal is performed as a melody. Where does this assumption come from? That is, nowhere is it clear that these letters are transformed into a melodic representation (i.e. related to a musical sequence). This seems overly speculative.

7. The beamforming shows the area specificity for the effects in auditory cortex. The volume conduction paragraph should take this into account, but should also mention that the PLV and Laplacian based referencing scheme eliminated the effects. There should be a discussion about the possible reasons as to why this was observed in the author's data analysis.

8. On a small technical note, raw Granger should not be multiplied by 100%. Please revert back to the raw Granger values, as this would make the work more comparable to other studies, and these numerical units are meaningful. It is OK to express GC percentage differences in the cortex to hippocampus vs. hippocampus to cortex directions during the task modulation, however, to emphasize the changing direction patterns.

9. Relative power plots are not showing any decreases in power relative to fixation. How were the data baselined? They vary from 0 to 1. What about power decreases from fixation?

10. Lines 200-206 when reporting the change in Granger using a permutation test, but no p-values are given.

Reviewer #2:

I appreciate some of the additions that the authors have included. I do have further suggestions on the manuscript.

No details are provided for the scalp EEG system. How many electrodes? Was this an active system? Filtering at all? Many of these points are particularly relevant and important for interpreting the beamforming data. It says that the electrodes were aligned for each participant – how was electrode localization performed? Using what system – Polhemus?

Was all source reconstruction done with a template MRI? If so, wouldn't it have been appropriate to use each individual's MRI for better localisation, given that you surely had these available? You could then do the group averages, but at least the individual subject data would be more accurately localised.

The granger causality analyses are reported for only three of the subjects between hippocampal LFP and ECoG data. Did the other 12 subjects not have any cortical electrodes then, even without parietal coverage, only hippocampal LFPs? This isn't at all clear from the descriptions in the manuscript and seems surprising that data would only be available from hippocampus, and no other temporal regions, even if parietal coverage was lacking.

It would be beneficial to include a display of all the contact locations for hippocampus, for transparency and interpretability purposes, along with (or perhaps instead) a supplement including all MNI coordinates of localised electrodes.

Reviewer #3:

The authors have quantified patterns of information flow during a working memory task between cortex and hippocampus. They capitalize on a unique dataset where human intracranial data have been collected from the hippocampus, in parallel with human electrocorticography and/or electroencephalography. The authors use a statistical method to infer functional directed connectivity. This reveals information flowing from cortex to hippocampus during encoding but the reverse direction during maintenance. This study is quite unique in its combination of methodology with data that can precisely track the anatomical sources of neural computations during a memory task.

Recommendations for the authors:

In particular, the nice use of beamforming show the area specificity for the effects in auditory cortex. The volume conduction paragraph should take this into account, but should also mention that the PLV and Laplacian based referencing scheme eliminated the effects. There should be a discussion about the possible reasons as to why this was observed in the author's data analysis.

On a small technical note, raw Granger should not be multiplied by 100%. Please revert back to the raw Granger values, as this would make the work more comparable to other studies, and these numerical units are meaningful. It is OK to express GC percentage differences in the cortex to hippocampus vs. hippocampus to cortex directions during the task modulation, however, to emphasize the changing direction patterns.

Relative power plots are not showing any decreases in power relative to fixation. How were the data baselined? They vary from 0 to 1. What about power decreases from fixation?

Lines 200-206 when reporting the change in Granger using a permutation test, but no p-values are given.

eLife. 2022 Aug 12;11:e78677. doi: 10.7554/eLife.78677.sa2

Author response


Reviewer #1:

The authors have shown a unique set of recordings, wherein they have collected intracranial data from parietal cortex and hippocampus, as well as scalp EEG in a number of subjects. With this unique advantage, they have examined directionality of connectivity between various regions during a working memory task. Given the growing evidence for the role of hippocampus in working memory, understanding its connectivity to the rest of the brain provides a crucial insight into the network involved in such a fundamental process. Whilst the existing content is generally of a high standard, and the analyses seem sound, there are some areas of considerable brevity that would benefit from expansion. Below are my comments on the manuscript.

Discussion: This is surprisingly short Discussion section. I feel this should be expanded considerably, such as including some of the information that I have discussed below regarding potential considerations of the task (e.g bimodal nature), discussion of the PLV results in the context of previous non-directional findings, the differences observed between correct and incorrect trials, considering in more detail the other behavioral consequences of these results. These suggestions are not necessarily exhaustive but are all points I believe should be included.

We found muscle artefact in the scalp EEG but not in the ECoG/sEEG. Trials with muscle artefact in the scalp EEG were removed.

Following the Reviewer’s comment, we have now mentioned this in more detail in the Methods (Line 419). Furthermore, we have created a new Table 1 where we list the number of correct trials analyzed for each participant.

Comments on Results section in general:

– All the results in this section seem to refer to a single electrode for each subject. It would be beneficial to know whether these electrodes were representative of activity from surrounding electrodes or not. That is, how generalizable are the PSD results shown here.

Participant 1 had a large grid over posterior cortex and we now present several topographical plots. We were indeed able to see some similarities between adjacent electrodes.

Following the Reviewer’s comment, we have now moved the topographical plots of GC from the supplementary material to the new Figure 2. Furthermore, we have added topographical plots for PSD in the new Figure 2 a, d, for PLV in Figure 2 h, and for ΔGranger in Figure 2 i, j. Regarding the GC to scalp EEG, we have created Table 2 that highlights all scalp electrode sites where ΔGranger was significant during maintenance. It turned out that ΔGranger was always > 0 (net information flow from hippocampus to cortex) and that the majority of these scalp electrodes were on central or temporal sites, suggesting involvement of auditory cortex.

– Also, many of these results are very descriptive. Whilst in some specific scenarios this is unavoidable, for the purposes of reporting results from PSD (for example) it is definitely possible to report details such as the degree of power increase. At present, this reads more like a Discussion section that an informative Results section.

We show PSD in time-frequency plots. As the reviewer rightly requests, these show the degree of power increase with respect to fixation [-6 -5] s as a baseline according to (PSD(t) – PSD(fixation))/ PSD(fixation). For example, in Figure 1b, γ PSD during encoding has increased more than twice with respect to PSD during fixation.

Following the Reviewer’s comment, we have now entered the number of maximal increases in the Results (Line 124, 126, 142).

We focus on changes in GC as the main novel message of our study. Here we perform extensive computational statistics. In earlier studies with the same task, we have extensively analyzed scalp EEG (Michels 2008), PLV between hippocampus and scalp EEG (Boran 2019), and single neuron firing in the hippocampus (Boran 2019). We validate our results by statistical testing to ascertain that the observed changes in GC were indeed associated with the task.

– It would be helpful to see an overlay of the parietal electrode with a topographic map of the scalp EEG recording, to truly appreciate the spatial overlap between the electrode and the generator.

We thank the reviewer for this suggestion.

We have now added the scalp EEG recording sites to the corresponding panels in Figure 2 (Line 524-525) and Figure 3 (Line 581-582).

Figures in general: many of the figures appear to refer only to single subjects. It would be useful to have more detailed summary information across subjects to understand how reliable/variable these effects are.

Figure 1 and Figure 5 aggregate information across subjects. The panels of Figure 2 and Figure 3 are arranged to facilitate comparison between results of participants 1,2,3. The panels of Figure 4 allow comparison across all participants.

We have created a new Table 1 and Table 2 to facilitate the understanding of how reliable the effects are across subjects.

Data availability section: The bit on previously published datasets confused me a little. Is this published dataset included as part of this article? It isn't so clear in the manuscript whether these are previously published data. If they are, this should be made more apparent.

The data of our earlier study (Boran 2019) has been published (Boran 2020). In that study we had analyzed the PLV between scalp EEG and hippocampus sEEG in 9 participants. We include these participants here. As a new analysis, we calculate GC between scalp EEG and hippocampus sEEG.

We have created a table listing the data that has been published previously and the data that is newly analyzed for this manuscript.

Line 29: Phrasing – I would add the words "rather than sequentially" here to help readers with interpretation of why this separates out encoding from maintenance

Following the Reviewer’s comment, we have now added "rather than sequentially" (Line 31).

Line 64: This can actually extend as low as δ band (see Leszczynski et al., 2015, Cell Reports; Kumar et al., 2021, Neuropsychologia).

Following the Reviewer’s comment, we have now added this information and cited (Kumar 2021), (Line 65).

Line 88: Do the authors have behavioral data or prior knowledge of how long it takes (on average) to encode 4, 6 or 8 letters? That is, how much of the 'encoding' period is truly encoding, rather than an initial encoding followed by maintenance. Or in a similar manner, how much of maintenance is still residual encoding.

In the design of the task we were guided by the magic number 7±2, which may correspond to “how many items we can utter in 2 seconds” (Baddeley 2003). Certainly, maintenance neurons ramp up their activity already during encoding (Figure 2 in (Boran 2019)). In a similar manner, encoding may extend past the visual stimulus (t = -3 s). We therefore analyze only the last two seconds of maintenance [-2 0] s. With this design, we found patterns of GC that were distinct between encoding and maintenance.

We have now added this information in the discussion (Line 285 – 296)

Line 90: Was there a particular reason as to why the encoding phase was bimodal? Do the authors think this may have influenced their results?

Several of the patients were not able to read 8 letters within the encoding period of 2 s. Reading the letters to them greatly improved their performance. At the same time, the auditory encoding may have extended beyond the 2 s period. We therefore restrict our analysis to the last 2 s of the maintenance period.

Following the Reviewer’s comment, we have now entered this in Methods (Line 362-364).

Line 94-95: Was this an instruction to the participants? If so, I would put this more explicitly, i.e. "participants were instructed to rehearse…". Of course, one cannot know for certain whether individual subjects employed this strategy.

We instructed participants to rehearse. We further asked them whether they had in fact employed this strategy. The answer was yes in all subjects.

Following the Reviewer’s comment, we have now entered the following in Results (Line 100-101).

Figure 2f: Where is this change in Granger relative to? A particular baseline window.

ΔGranger denotes the net information flow (GChippcortex – GCcortexhipp). A positive ΔGranger indicates predominant information flow from hippocampus to cortex (dark red spectrum exceeds light red spectrum), e.g. during maintenance (new Figure 2h). The new Figure 2l illustrates ΔGranger for each time point for contact C2.

We have now described this in more detail in Results and the figure caption (Line 170 -172, 179-180, 199, 576-578)

Line 293: Were any electrodes here included in a seizure foci? Was anything done to ensure that seizure activity did not affect recordings (e.g. not recording within xxx hours of a seizure)?

Some of the electrodes were in tissue that was deemed to be epileptogenic and that was later resected. Still, neurons in this tissue have been found to participate in task performance (Boran 2019). All recordings were done at least 6 h from a seizure.

Following the Reviewer’s comment, we have now entered this in the Methods and we now report whether the hippocampal recording site was part of the seizure onset zone in Table 1 (Line 402-404).

Line 301: Was anything done to deal with artefacts on the ECoG/sEEG electrodes? I.e. were trials with unusually large amplitudes, potentially indicative of muscle artefact (a known contaminant) removed?

We found muscle artefact in the scalp EEG but not in the EcoG/sEEG. Trials with muscle artefact in the scalp EEG were removed.

Following the Reviewer’s comment, we have now mentioned this in more detail in the Methods (Line 419). Furthermore, we have created a new Table 1 where we list the number of correct trials analyzed for each participant.

Line 325-326: I am confused by this. You say that the individual frequencies may differ between participants – do you mean in terms of the peak frequency, or were different bands used for each subject? If different bands, why?

The frequency ranges result from the statistical analysis. The frequency ranges turned out to vary across participants. The variability of frequency bands across participants has been shown earlier for this task (Boran 2019, Michels 2008) and is a general phenomenon in cognitive EEG research.

We now clarify this phrase and extend the discussion (Line 315 – 319).

All power spectral density plots: I assume these are relative to baseline. Are they statistically-thresholded in any way?

Correct, the power spectral density plots are relative to the fixation period [-6 -5] s. They are not thresholded.

We now add this information to the Figure captions (Line 569- 570, 610 – 611).

Reviewer #2:

Dimakopoulos et al. use intracranial data in humans to ask whether information flow is primarily cortical to hippocampal or the reverse during the encoding and retrieval stages of a working memory task. They find a highly reliable pattern where information in the α/β range flows from auditory cortex to hippocampus during encoding and in the reverse direction during maintenance of items in WM. The authors show this pattern in a sub-selection of EcoG recordings and go on to show it is present in virtually all subjects at the EEG to intracranial hippocampus level. In addition, this directional pattern breaks down during incorrect trials. However, the current analysis suffers from possible contamination by volume conduction.

The study is unique in Its data set and provides a valuable look into hippocampal cortical interactions during WM. However, there are multiple technical questions remaining. One of the limitations is that the study investigated primarily interactions in the α/β range when looking at interactions. In contrast, their power spectral results show increases in γ during encoding, and other studies have emphasized a role for γ in feedforward routing. Did the authors perform a granger causality analysis in γ?

Please find our detailed response below to Point 2.

1. The reported PLV values (e.g., in figure 2d) of 0.3 – 0.6 are quite high. This level of PLV is usually observed when there is some amount of volume conduction present. Could the authors repeat the PLV and GC analysis using a local bipolar re-referenced signal? For more information on why this is a problem please see Trongnetrpunya, Nandi, Ding, et al., Frontiers in Systems Neuroscience.

We apologize for presenting PLV in Figure 2d with a color scale only in the range [0.3 0.6]. In the new Figure 2g (Line 598), we now show the full range of PLV < 0.6. The PLV to parietal recording sites is < 0.1, i.e. with high anisotropy across contacts. Similarly, GC is highly anisotropic both for eCoG (Figure 2) and for scalp EEG (Table 2). This anisotropy speaks against a significant amount of volume conduction present.

Following the Reviewer’s suggestion, we have repeated PLV and GC analysis with a Laplacian montage for the ECoG of Participant 1 and a bipolar montage for the ECoG of all three Participants with ECoG. There was neither PLV nor GC. However, Laplacian and bipolar montages focus on highly localized neuronal assemblies and provides somewhat complementary information to the referential montage (Nunez 1995). The activity of local assemblies is rather associated with higher frequencies and local processing. Inter-areal interactions tend to take place rather in lower frequencies (Solomon 2017, von Stein 2000), which require a referential montage to be appreciable in the signal.

There was a strong frequency dependence of PLV to ECoG (Participant 1, Figure 2g) and to scalp EEG (Figure 6 in (Boran 2019)). Likewise, GC to ECoG and to scalp EEG showed a strong frequency dependence (Figures 2, 3, 4). This speaks against volume conduction because the transfer of signal through tissue by volume conduction is independent of frequency in the range of the EEG studied here (Miceli 2017, Nunez 1995).

There was a strong task dependence of both PLV (Boran 2019) and GC (Figures 2, 3, 4). This also speaks against a strong contribution of volume conduction.

In conclusion, while we cannot rule out a contribution of volume conduction in our data, it seems to be small compared to the effects in PLV and GC that were induced by the processing of the cognitive task.

We have now added a paragraph on volume conduction in the discussion (Line 297 – 313). We have now reproduced the PLV results of (Boran 2019) in a new supplementary Table S1, in which we have added the PLV results of the additional participants of this study.

2. Did the authors find any detectable GC or PLV in the γ frequency range, particularly during the encoding epoch?

We agree that there was a strong effect in γ during encoding. Therefore, we always extended our analyses to γ frequencies. Unfortunately, in our GC analysis >30 Hz we never found significant difference in directionality. Please see Author response image 1 for the analysis up to 100 Hz.

Author response image 1.

Author response image 1.

We therefore presented spectra only <30 Hz. As can be seen in all GC plots, the four GC spectra show little difference already around 30 Hz.

We now report this null-finding in Results (Line 178)

3. Across subjects, the authors report that the Sensory to Hippocampus information flow is dominant during sensory processing and the reverse during delay processing (Figure 5). In Figure 5b, it is impressive how the effect holds in each and every individual subject. My question is how were the data selected in figure 5b? Is that an average of all GC pairs linking all available hippocampus leads to / from all available EEG leads, or was there a data selection that occurred? What would the results look like if only the EEG leads are selected that overly to auditory/temporal cortex? Is the topography the same or different when performing the analysis on e.g., frontal / occipital / parietal EEG leads? In other words, how spatially specific is the effect?

The spatial sampling of scalp EEG electrodes is very sparse when compared with the anatomical specification of the cortex. We can not a priori assume that adjacent electrodes show similar behavior. For Figure 4 and Figure 5 we therefore selected the electrode with the highest GC value where the reversal of net GC (ΔGranger) was statistically significant. Figure 4 and Figure 5 present data from the same scalp electrodes, respectively.

We have created a new Table 2 where we list the maximal GC during maintenance in the frequency band where the reversal of net GC (ΔGranger) was statistically significant. In 14/15 participants, the maximal GC appeared in electrodes over central or temporal cortex.

We have now mentioned this finding in the Results (Line 216 – 219) and in the Discussion (Line 332 – 337).

Reviewer #3:

Dimakopoulos and colleagues investigate connectivity and flow of information during encoding and maintenance of Working Memory. They use unique data, which combine human intracranial recordings from depth electrodes with ECOG and EEG. This data, combined with Granger causality analysis (GC), provides interesting results that signal from cortex (mostly from EEG electrodes located over temporal cortex) is flowing to hippocampus during encoding and this flow is reversed during maintenance. Authors interpret this as a sign of bottom-up and top-down processing. I believe that chosen methods for signal analysis are appropriate.

However, paper contains several statements that are unsupported by statistics and there is no clear information about why some decisions in the analysis process were made. This could give an impression that the analysis is built from arbitrarily chosen single case examples. I believe that because of below listed flaws results of the analysis do not support conclusions.

1) Authors do not use correction for multiple comparisons – this cast doubts on the strength of obtained results.

Following the Reviewer’s comment, we add data from more electrodes and test the spatial distribution of the results on the grid electrode.

We have now added further analyses and panels to Figure 2 and Figure 3 and created a new Table 2. In Figure 2, the new panels show that the grid contacts C2 and H3, which we mainly present, are surrounded by contacts with similar behavior (Figure 2 a, d, g, i, j). To provide a sound statistical basis, we have now tested the spatial distribution of PSD, PLV and GC on the grid contacts against a null distribution. The activation on grid contacts was transformed into a grid vector. The spatial collinearity of two grid vectors is captured by their scalar product. We tested the statistical significance of the spatial spread of contacts with high ΔGranger (8-14 Hz) during maintenance ([-2 0] s). We first reshaped the ΔGranger values on the grid into a vector over 200 iterations of random trial permutations. For each iteration we selected two subsets of trials and we calculated the scalar product between the two subsets.The true distribution of scalar products occurs only above the 95th percentile of the null distribution and is thereby statistically significant (Figure 2 k).

For the scalp electrodes of all participants, the Table 2 shows that significant GC occurred specifically to central/temporal scalp EEG electrodes in 14/15 participants (p=0.005, two sided non-parametric permutation test).

We have now added this information to the Results (Line 186 – 197) and the Discussion (Line 327 – 334) sections.

2) There is no criterion given for ECOG electrodes selected to the analysis.

For instance, authors state that for participant 1 for C2 electrode, increased γ power during encoding proves that this electrode was over auditory cortex but there is no systematic analysis of γ power. From the results we can observe that this electrode has the strongest GC with hippocampus what suggests that it was used because of this characteristic what looks like double dipping.

Following the Reviewer’s comment, we have added further analyses and panels to Figure 2. Encoding elicited γ power increase at several temporal contacts (among them C2) and occipital contacts. We tested the statistical significance of the increase for each contact. Figure 2a shows the z-score of those contacts in red where the increase reached statistical significance of the increase (z-score >1.96, permutation t-test). In a separate test (NOT double-dipping!), we identified the grid contacts with significant increase in PLV and GC in Figure 2 g, i, j. These turn out to be on temporal cortex as well.

3) Why frequencies observed in PLV and GC are so different? For instance, in supplementary Figure 1 PLV shows significant differences in 18-30 Hz but GC is calculated for 8-18 Hz. Such large differences in frequencies suggest some inconsistencies in the analysis.

In a mechanistic view of communication theory, one assumes that a node displays high PSD, high PLV, high GC, and all of this in the same time window and in the same frequency band. Our data come close to this view but the frequency bands do not match. Since this mismatch follows directly from the data, we would rather call it an inconsistency between the available analysis methods and our underlying assumptions.

Following the Reviewer’s comment, we now highlight these considerations in the Discussion (Line 315 – 323).

4) For analyses depicted in Figure 4 and 5 it is unknown how the highest GC is defined (is it a mean from all frequencies?) Furthermore, there is no systematic measurement or criterion that would support that indeed chosen electrodes have the highest GC.

The highest GC is defined as the mean in the band of significant GC change. We present here the electrode pairs with the highest ΔGranger values. It turned out that in 14/15 participants, the highest GC occurred in temporal electrodes over auditory cortex or in participants where temporal sites were not recorded from, the highest GC occurred at the neighboring electrode sites C3 or C4 (Table 2, p=0.005, two sided non-parametric permutation test).

We have now prepared a Table 2 that lists all electrodes and GC values. We now mention our criterion not only in the Methods (Line 466, 482 – 491) but also in the Results sections (Line 242 – 244).

5) All analysis conducted in the time domain (time to frequency and GC) does not contain any statistics supporting validity of the proposed conclusions.

We agree with the Reviewer that the time frequency maps are illustrative only. Statistical testing was performed for the GC frequency spectra in the time windows encoding [-5 -3] s and maintenance [-2 0] s. The time-frequency plots are labelled as mere illustrations throughout the text.

6) There is no data that supports statement that patients used verbalization. Although material is verbal authors cannot rule out that subject uses different modalities to support information maintenance.

During the maintenance period, participants rehearsed the verbal representation of the letter strings subvocally, i.e. mentally replayed the memory items. Participants had been instructed to employ this strategy and they confirmed after the sessions that they had indeed employed this strategy.

We now add this information more prominently in Results (Line 100).

7) In figure 3i significance for encoding (blue) is marked where there are no differences between dark blue and light blue conditions. This suggests that the significance is computed not between conditions but versus null distribution. In general, in the paper, it is unclear if comparison using null distribution or between conditions is being calculated.

The Reviewer is correct. We compute significance against a null distribution as stated in the Methods: “To assess the significance of GC, we compared the true values to a null distribution. We recomputed GC after shuffling the trial number for a single channel in the pair, while keeping the trial number of the other channel constant. We repeated this n = 200 times to create a null distribution of GC. The null distribution was exploited to calculate the percentile threshold P = 0.05.” We mark the frequency range of significant GC with a blue bar for encoding (dark blue spectrum exceeds light blue spectrum, information flow from cortex to hippocampus) and with a red bar for maintenance (dark red spectrum exceeds light red spectrum, information flow from hippocampus to cortex).

We now mention the null distribution also in Results (Line 170 – 172, 175 – 176). The net information flow ΔGranger (GChippcortex – GCcortexhipp) during encoding was significant in the 8-18 Hz range (blue bar in Figure 2 i, p<0.05 permutation test against a null distribution).

8) In Figure 2b – why EEG, and not ECOG, signal is being analyzed?

We now highlight the difference in recording modality in Results (Line 133 – 135).

9) I appreciate that authors present single subject results in Figure 4 but I believe group level analysis should be also performed to support conclusions.

Figure 5a shows a group level analysis of the single participant results of Figure 4. The panels of Figure 4 allow comparison across all participants.

We have created the new Table 2 to facilitate the understanding of how reliable the effects are across subjects.

10) Does the quality/magnitude of GC estimation depends on the number of trials? If so

for Figure 5b authors show subsample number of trials in the correct condition to match the number of trials in the incorrect condition.

We thank the reviewer for asking this question.

To balance trial numbers, we calculated the GC in a subset of correct trials (median of 200 permutations of a number of correct trials that equals the number of incorrect trials for each participant). The distributions between encoding and maintenance differed (P = 1e-4, paired cluster based permutation test, Figure 5 b). It thus turned out that the GC estimation does not deteriorate when balancing the numbers of correct and incorrect trials. We now present the balanced finding in the Results (Figure 5 b, Line 245 – 248).

11) Authors should state how many sessions were used in each comparison.

We have now created Table 1 that lists the number of correct trials for each participant.

12) In the statement:

"The GC was lower than for ECoG, as expected for the lower signal amplitude of scalp EEG." I do not know what "lower" means – is it a magnitude or frequency of GC? If frequency, how does it relate to the scalp EEG amplitude?

We apologize for the unclear wording. The scalp EEG signal is smaller in amplitude than the ECoG signal. We therefore assume that derived quantities like the GC are smaller, too.

We have now reworded the phrase for clarity (Line 223).

13) Throughout the paper it seems that "ΔGranger" is used to describe a different thing.

We used the term ΔGranger throughout this paper to describe differences in directionality of the Granger. Subtracting the Granger spectrum of one direction (cortex hippocampus) from the other (hippocampus cortex) we aim at showing the bidirectionality of information flow by means of sign reversal. This approach is common in the literature, e.g. (Jenison 2014).

We assume that the Reviewer is referring to the label of the color bars. This is not the percentage of change but simply the difference of GC (where we use the unit % for better legibility). We have now removed the % from the label for clarity (Line 576).

References

Baddeley A. 2003. Working memory: looking back and looking forward. Nat Rev Neurosci; 4:829-839. doi 10.1038/nrn1201.

Boran E, Fedele T, Klaver P, Hilfiker P, Stieglitz L, Grunwald T, Sarnthein J. 2019. Persistent hippocampal neural firing and hippocampal-cortical coupling predict verbal working memory load. Sci Adv; 5:eaav3687. doi 10.1126/sciadv.aav3687.

Boran E, Fedele T, Steiner A, Hilfiker P, Stieglitz L, Grunwald T, Sarnthein J. 2020. Dataset of human medial temporal lobe neurons, scalp and intracranial EEG during a verbal working memory task. Sci Data; 7:30. doi 10.1038/s41597-020-0364-3.

Jenison RL. 2014. Directional Influence between the Human Amygdala and Orbitofrontal Cortex at the Time of Decision-Making. PloS one; 9:e109689. doi 10.1371/journal.pone.0109689.

Kumar S, Gander PE, Berger JI, Billig AJ, Nourski KV, Oya H, Kawasaki H, Howard MA, Griffiths TD. 2021. Oscillatory correlates of auditory working memory examined with human electrocorticography. Neuropsychologia; 150:107691. doi https://doi.org/10.1016/j.neuropsychologia.2020.107691.

Miceli S, Ness TV, Einevoll GT, Schubert D. 2017. Impedance Spectrum in Cortical Tissue: Implications for Propagation of LFP Signals on the Microscopic Level. eNeuro; 4. doi 10.1523/ENEURO.0291-16.2016.

Michels L, Moazami-Goudarzi M, Jeanmonod D, Sarnthein J. 2008. EEG α distinguishes between cuneal and precuneal activation in working memory. NeuroImage; 40:1296-1310. doi 10.1016/j.neuroimage.2007.12.048.

Nunez P. 1995. Neocortical dynamics and human EEG rhythms: Oxford University Press.

Solomon EA, Kragel JE, Sperling MR, Sharan A, Worrell G, Kucewicz M, Inman CS, Lega B, Davis KA, Stein JM, et al. 2017. Widespread theta synchrony and high-frequency desynchronization underlies enhanced cognition. Nat Commun; 8:1704. doi 10.1038/s41467-017-01763-2.

Trongnetrpunya A, Nandi B, Kang D, Kocsis B, Schroeder CE, Ding M. 2015. Assessing Granger Causality in Electrophysiological Data: Removing the Adverse Effects of Common Signals via Bipolar Derivations. Front Syst Neurosci; 9:189. doi 10.3389/fnsys.2015.00189.

von Stein A, Sarnthein J. 2000. Different frequencies for different scales of cortical integration: from local γ to long range α/theta synchronization. Int J Psychophysiol; 38:301-313. doi.

[Editors’ note: the authors resubmitted a revised version of the paper for consideration. What follows is the authors’ response to the second round of review.]

We found the approach to working memory based on effective connectivity between sensory cortex and hippocampus during working memory maintenance and retrieval intriguing and novel. All of the referees appreciated the efforts that have been made to improve the manuscript. The principal reason for rejection is the statistical interpretation of the data raised by all three referees, who are concerned whether the analyses allow robust inference about effects in the population. Referee 1 highlights differences between subjects in the frequency bands of effective connectivity. The effects are not robust to re-referencing recommended by referee 2, who does not consider a common reference adequate for this work. Referee 3 is concerned about multiple comparisons.

Reviewer #1:

I appreciate the efforts the authors have taken to improve the manuscript. The additional information does aid interpretation of the study. I still have a couple of remaining core issues to be sorted.

We thank the reviewer for this appreciation.

1) In the abstract, it states that information flow from auditory cortex to hippocampus peaked in the range of 8-18 Hz. In the newly inserted (and very helpful) table it seems that the range is anywhere from 4 to 21 Hz. Unless I am missing something, it appears that the authors are still making conclusions based on only Participant 1 (who showed significance in the 8-18 Hz range). I feel that this confusion runs throughout the manuscript and should be corrected – that is, any points where generalizations are made from a single subject that are not truly consistent across subjects should be corrected and the variability should be highlighted/discussed.

2) Similarly, the frequency band of hippocampus > auditory cortex is often not congruent with the AC > hippocampus within subjects, so it is too simplistic to say this was reversed during maintenance. Moreover, this difference in frequencies within subjects should be discussed.

We have now completely revised our data analysis. The results addressed by the reviewer (Figures 4 and 5 of the previous version) are now less relevant and no longer reported in the manuscript.

Following the reviewers’ comments, we have now improved the quality of our data. In particular, we have now

  • used separate references for hippocampal LFP and ECoG recordings,

  • determined the cortical sources of our EEG recordings by beamforming,

  • applied cluster-based permutation tests on the population of participants.

We have then improved our GC analysis by down sampling the data to the Nyquist frequency. While new GC analysis confirmed the reversal of the net information flow between encoding and maintenance, the significant differences now appeared in theta band in all analyses.

We have now described the results of our improved GC analysis in the Results sections 2.4, 2.5 and 2.6. We have created several new figures and panels. We now present the GC spectra for Participant 1 (Figure 2h), for Participants 2,3 (Figure 3 d,i) and for the beamforming sources of all participants in Figure 4a. Significant net information flow (ΔGranger) appeared in the theta frequency range for all pairs of signals that we analysed. We therefore selected the 4-8 Hz band to test the reversal of ΔGranger in all participants individually (Figure 4d) and showed that the reversal appears for correct trials only (Figure 4f).

Reviewer #2:

The authors have attempted to answer my concerns by performing a bipolar and Laplacian analysis to the data prior to the connectivity analysis. Unfortunately, this revealed that effects were not preserved after applying local re-referencing. I appreciate the authors' arguments that some effects have a large spatial structure that is removed by bipolar. However, the electrodes used by the authors are quite far apart and a priori I would think that a referencing across a large spatial structure should preserve true effects while removing spurious ones.

Unfortunately, given simulation work as in the article I referred to in my original review (Trongnetrpunya, Nandi, Ding, et al., Frontiers in Systems Neuroscience), the false positive rate is unacceptably high when Granger and related methods are applied to data sharing a common reference. The current results therefore may be incorrect and cannot be interpreted in their present form.

To address the reviewer’s concern, we have revised major parts of our data analysis.

We have re-analyzed the LFP and ECoG data with two separate references:

1) We re-referenced the hippocampal LFP against the signal of a depth electrode contact in white matter.

2) We re-referenced the cortical ECoG against a different depth electro de contact.

The choice of two separate references for LFP and ECoG has been shown to avoid spurious functional connectivity estimates (Bastos and Schoffelen, 2015).

Our main findings remain unchanged, apart from minor changes in the topography and that GC now appears in the theta frequency band. This new analysis has corroborated our previous findings of

1) net information flow from auditory cortex to hippocampus during encoding.

2) net information flow from hippocampus towards auditory cortex during maintenance.

We have now described the new re-referencing scheme in the Methods Section 4.5 “Recording setup, re-referencing, and preprocessing”. We have adapted the results of our analysis throughout the manuscript.

The authors also apparently have cherry picked their data in Figures 4 and 5 by first taking channels that show a change in Granger, and then analyzing those changes across participants. It is like forcing the data to behave as we would like, another example of a flawed approach taken by the present manuscript. Given these technical/methodological issues, I cannot recommend the manuscript for publication in the present state.

We have certainly not cherry picked our data. Nevertheless, as stated above we have thoroughly re-analyzed our data following the reviewer’s suggestion. This extensive re- analysis corroborated the earlier findings with improved statistics. The results addressed by the reviewer (Figures 4 and 5 of the previous manuscript) are now less relevant and no longer reported in the manuscript.

We have now used beamforming to determine the cortical sources of our EEG data. We then analyzed the Granger causality (GC) between the cortical sources and the hippocampal local field potentials (LFP). The statistics was then performed with cluster based permutation tests of the effects in the subject population. We now describe the new analysis of the GC in the source space in the Methods Sections 4.8 and 4.9. We have created a new Figure 4, a new supplementary figure (Figure S1) and modified the Table 1. We have now described the new results of our analysis throughout the manuscript.

I do want to commend the authors for recording this dataset, and performing directed connectivity analysis, which is rare for these types of data. I would encourage the authors to think more deeply about how to assess directed functional connectivity robustly and reformulate their work with this in mind.

We thank the reviewer for this encouragement. We have now analyzed the directed functional connectivity robustly on the beamforming source level, applied cluster based permutation tests and thereby reformulated our work accordingly.

The new analyses have corroborated and improved our earlier results.

Reviewer #3:

The authors have revised their manuscript and attempted to address all of the comments raised by myself and the other reviewers.

We thank the reviewer for this appreciation.

I have few remaining issues.

If authors compute the difference against null distribution not between conditions (for example light blue and dark blue) in Figure 3d,i and in Figure 4 – How statement "dark blue spectrum exceeds light blue spectrum, information flow from cortex to hippocampus" is correct? It seems that, for each light colour result (hippocampus to cortex during encoding, cortex to hippocampus during maintenance) there should be a result of a statistical test and it seems none of these tests were performed. The caption of Figure 4 where authors use ΔGranger seems also incorrect as authors compute statistics from null distribution.

We performed the analysis exactly as the reviewer suggested. We are sorry if this was not clear. We do not compute statistics from a null distribution but from a null distribution of differences.

Following the reviewer’s comment, we have now clarified the description of statistics in the Methods section 4.10:

“To assess the significance of the difference of the Granger between different directions, we compared the difference of the true values to a null distribution of differences.

We recomputed GC after switching directions randomly across trials, while keeping the trial numbers for both channels constant. Then we computed the difference of GC for the two conditions. We repeated this n = 200 times to create a null distribution of differences. The null distribution was exploited to calculate the percentile threshold p = 0.05. In this way, we compare the difference of the dark and light spectra against a null distribution of differences”.

Unfortunately, I believe that issue with multiple comparisons was not addressed entirely. For instance, the analysis in the frequency domain does not use any correction for multiple comparisons. Because of that it is unknown if small differences observed in results are not effects of multiple comparisons.

Following the reviewers’ comments, we have now thoroughly reanalyzed our data.

We have now used beamforming to determine the cortical sources of our EEG data (Methods 4.8). We then analyzed the Granger causality (GC, Methods 4.9) between the cortical sources and the hippocampal local field potentials (LFP). The statistics was then performed with cluster based permutation tests of the effects in the subject population (Methods 4.10). This corrects for multiple comparisons (false-rate-discovery) (Maris and Oostenveld, 2007). We statistically compared the spectral GC in EEG sources between the two directions (hipp->cortex, cortex-> hipp) at the group level. We first selected a task period (encoding or maintenance). We then selected a cortical source and calculated the GC to hippocampus. For auditory cortex, the cluster-based permutation tests on the group level (Figure 4a) revealed

1) net information flow from auditory cortex to hippocampus during encoding,

2) net information flow from hippocampus towards auditory cortex during maintenance.

This was also true for theta GC to other cortical areas on the group level, but with less significance (Figure 4b,c).

On the level of individual participants, the reversal of the net information flow between encoding and maintenance was significant (paired permutation test) only to auditory cortex (Figure 4d). There were statistically significant differences among the cortical areas regarding net information flow both during encoding (p = 0.0009, Kruskal-Wallis test) and maintenance (p = 0.001, Kruskal-Wallis test). We confirmed that the net information flow in auditory cortex was significantly different from any other area (Dunn’s test, Bonferroni corrected).

Additionally for Figure 5 b and c analysis it seems that the comparison is not designed correctly. For figure 5b the data is picked using significant reversal of net GC (ΔGranger) – this is not done for Figure 5c analysis so it is hard to draw any conclusions from this contrast.

We have now completely revised our data analysis. The results addressed by the reviewer (Figures 4 and 5 of the previous version) are now less relevant and no longer reported in the manuscript.

We have now used beamforming to determine the cortical sources of our EEG data (Methods 4.8) and performed cluster-based permutation tests (Methods 4.10). We have thereby replaced the old Figures 4 and 5 with a new Figure 4.

About the group statistics issue. Figure 5a although presents average results does not show any statistics thus it is unknown if any of those information flows is significant. Figure 4 can give me an intuition that the result is present, but I have no idea about how significant it is on the group level. I believe that paper would benefit from the group statistics.

We have now completely revised our data analysis. The results addressed by the reviewer (Figures 4 and 5 of the previous version) are now less relevant and no longer reported in the manuscript.

We have now used beamforming to determine the cortical sources of our EEG data (Methods 4.8) and performed cluster-based permutation tests (Methods 4.10). We used group cluster based permutation with multiple comparison correction on the group level to determine the frequency ranges where the GC differs significantly between the two directions for the two task periods. We have now added bars in new Figure 4a to indicate the frequency ranges where spectral GC differs on the group level. As stated above, for auditory cortex the cluster-based permutation tests on the group level (Figure 4a) revealed

1) net information flow from auditory cortex to hippocampus during encoding,

2) net information flow from hippocampus towards auditory cortex during maintenance.

References

Bastos AM, Schoffelen JM (2015) A Tutorial Review of Functional Connectivity Analysis Methods and Their Interpretational Pitfalls. Front Syst Neurosci 9:175.

Maris E, Oostenveld R (2007) Nonparametric statistical testing of EEGand MEG-data. Journal of Neuroscience Methods 164:177-190.

[Editors’ note: further revisions were suggested prior to acceptance, as described below.]

The work provides new insights into the communication between sensory cortex and hippocampus during working memory.

The revised work has been seen by all of the original referees who agree it is substantially improved. They have raised a number of further points about analysis and exposition we would like you to consider when preparing the final article for acceptance.

Reviewer #2:

I appreciate some of the additions that the authors have included. I do have further suggestions on the manuscript.

We thank the reviewer for this appreciation.

No details are provided for the scalp EEG system. How many electrodes? Was this an active system? Filtering at all? Many of these points are particularly relevant and important for interpreting the beamforming data. It says that the electrodes were aligned for each participant – how was electrode localization performed? Using what system – Polhemus?

We thank the Reviewer for pointing us to this omission.

We have now described the scalp EEG recordings in more detail in the Methods.

Section 4.3

“For scalp EEG recording, cup electrodes (Ag/AgCl) were placed on the scalp and filled with electrolyte gel (Signagel, Parker Laboratories, parkerlabs.com) to obtain an impedance < 5 kΩ.”

Section 4.4

“The scalp EEG electrodes were placed at the sites of the 10-20 system by experienced technicians and no further localization was performed. While the 10-20 standard is 21 scalp electrodes, in some patients some electrode sites stayed vacant to assure the sterility of the leads to the intracranial electrodes, resulting in a median of 17 scalp sites per patient.”

Section 4.5

“All recordings (LFP, ECoG, and scalp EEG) were performed with the Neuralynx ATLAS system (sampling rate 4000 Hz, 0.5-1000 Hz passband, Neuralynx, neuralynx.com).”

Section 4.8.

“To solve the forward problem we used a precomputed head model template and aligned the EEG electrodes of each participant to the scalp compartment of the model via interactive scaling, translation and rotation (ft_electrode_realign.m).”

Was all source reconstruction done with a template MRI? If so, wouldn't it have been appropriate to use each individual's MRI for better localisation, given that you surely had these available? You could then do the group averages, but at least the individual subject data would be more accurately localised.

The reviewer is right. We used precomputed head models based on a template MRI. We agree that subject-specific head models yield source localization that may be more accurate. However, there are several factors that affect the variability in spatial localization and the use of a template MRI is only one of them. In our study, the localization accuracy at the group level appeared sufficient to support our main finding: information flows from auditory cortex and hippocampus during encoding and reverses direction during replay of memory items. We envisage higher localization accuracy in future studies.

The granger causality analyses are reported for only three of the subjects between hippocampal LFP and ECoG data. Did the other 12 subjects not have any cortical electrodes then, even without parietal coverage, only hippocampal LFPs? This isn't at all clear from the descriptions in the manuscript and seems surprising that data would only be available from hippocampus, and no other temporal regions, even if parietal coverage was lacking.

Planning of the intracranial electrode placement is a clinical decision. It reflects where the epileptologists hypothesize the epileptic foci to be localized (Zijlmans et al., 2019). Since the presumed epileptic foci included the hippocampus in all patients, electrodes were placed in the hippocampus. Only in four patients, additional ECoG electrodes were placed because an epileptic focus in the cerebral cortex was considered. In three of these patients, the ECoG electrodes were placed on the left hemisphere and the data analysis had been presented in Figure 2 and Figure 3. In the fourth patient, the ECoG electrodes were placed on the right parietal cortex and had not been presented before. While the spectral power was comparable to that of Participants 1-3, there was no significant Granger information flow from or to the right cortical hemisphere. This adds to the evidence from the scalp EEG sources that the phonological loop recruits predominantly the left cortical hemisphere.

Following the Reviewer’s comment, we now renumbered the participants so the participant with right parietal electrodes gets number 4. We added the findings of Participant 4 in five panels to Figure 3 and to the Results.

Section 2.2

“Similarly in the electrode contacts on right parietal cortex of Participant 4 (Figure 3 k), the letter stimulus elicited γ activity and the maintenance period showed α activity (8-11 Hz, Figure 3 l).”

Section 2.4

“However, in Participant 4, the recordings from the right cortical hemisphere (Figure 3 k), did not show significant Granger causality between LFP and ECoG during task performance (Figure 3 n,o).”

We now elaborate on the rationale for electrode placement in Methods Section 4.2 and have added the reference (Zijlmans et al., 2019).

“Electrodes were placed according to the findings of the non-invasive presurgical evaluation where the epileptologists hypothesized the epileptic foci to be localized (Zijlmans et al., 2019). Since the presumed epileptic foci included the hippocampus in all patients, electrodes were placed in the hippocampus. In four patients, additional ECoG electrodes were placed on the cortex because an epileptic focus in the cerebral cortex was considered.”

It would be beneficial to include a display of all the contact locations for hippocampus, for transparency and interpretability purposes, along with (or perhaps instead) a supplement including all MNI coordinates of localised electrodes.

We thank the reviewer for this suggestion. Note that recordings from both left and right hippocampus showed significant Granger information flow to the left cortical hemisphere.

Following the Reviewer’s suggestion, we have created two new supplements. In the new Supplementary File 1 we list all MNI coordinates of localized electrodes for hippocampus. In the new Figure 1 —figure supplement 1 we show the coordinates of localized electrodes for hippocampus projected in a left hippocampal surface.

Further, we have now prepared a new graphical representation of the hippocampal electrode locations in Figure 1d and Figure 4g.

We have now clarified that several participants had recordings in the right hippocampus.

Results section 2.6:

“Interestingly, the pattern appeared with LFP recorded from the right hippocampus in several participants (Supplementary File 1).”

Reviewer #3:

The authors have quantified patterns of information flow during a working memory task between cortex and hippocampus. They capitalize on a unique dataset where human intracranial data have been collected from the hippocampus, in parallel with human electrocorticography and/or electroencephalography. The authors use a statistical method to infer functional directed connectivity. This reveals information flowing from cortex to hippocampus during encoding but the reverse direction during maintenance. This study is quite unique in its combination of methodology with data that can precisely track the anatomical sources of neural computations during a memory task.

Recommendations for the authors:

In particular, the nice use of beamforming show the area specificity for the effects in auditory cortex. The volume conduction paragraph should take this into account, but should also mention that the PLV and Laplacian based referencing scheme eliminated the effects. There should be a discussion about the possible reasons as to why this was observed in the author's data analysis.

We thank Reviewer #3 for suggesting to expand our discussion on volume conduction. In following this comment, we have added a phrase in the Results section that mentions our analysis on the scalp electrode level that we had presented in a previous version of the manuscript. However, we fear that readers would be confused if we discuss why previous analyses – that are not detailed in the present manuscript – were criticized. We now do discuss the implications of the beamforming technique.

Following the Reviewer’s comments, we have now mentioned in the Results section 2.6:

“A similar GC pattern emerged when using the signals from left temporal scalp electrodes but was eliminated when using a Laplacian derivation.”

We restructured the discussion paragraph on volume conduction and added the phrase:

“On the level of scalp EEG analysis, we used beamforming as a source reconstruction technique [33] to characterize the primary neuronal generators that were localized specifically in left auditory cortex.”

On a small technical note, raw Granger should not be multiplied by 100%. Please revert back to the raw Granger values, as this would make the work more comparable to other studies, and these numerical units are meaningful. It is OK to express GC percentage differences in the cortex to hippocampus vs. hippocampus to cortex directions during the task modulation, however, to emphasize the changing direction patterns.

Following the Reviewer’s comment, we have now reverted to raw Granger values in the figures, the figure captions, and the text. At the same time, we have removed the leading zero from all decimal numbers <1 to improve legibility and to comply with the standard in psychology publications.

Relative power plots are not showing any decreases in power relative to fixation. How were the data baselined? They vary from 0 to 1. What about power decreases from fixation?

Following the Reviewer’s comment, we have now extended the color scale to negative values in all panels that report relative power.

Lines 200-206 when reporting the change in Granger using a permutation test, but no p-values are given.

We thank the reviewer for pointing this omission. As stated in the Methods Section 4.10, statistical significance is established at p<.05.

Following the Reviewer’s comment, we have now mentioned in lines 210-213 ”blue bar, p<.05,” and “red bar, p<.05”.

References

Baddeley A (2003) Working memory: looking back and looking forward. Nat Rev Neurosci 4:829-839.

Zijlmans M, Zweiphenning W, van Klink N (2019) Changing concepts in presurgical assessment for epilepsy surgery. Nature Reviews Neurology 15:594-606.

Associated Data

    This section collects any data citations, data availability statements, or supplementary materials included in this article.

    Data Citations

    1. Boran E, Fedele T, Steiner A, Hilfiker P, Stieglitz L, Grunwald T, Sarnthein J. 2020. Dataset of human medial temporal lobe neurons, scalp and intracranial EEG during a verbal working memory task. Scientific Data. [DOI] [PMC free article] [PubMed]

    Supplementary Materials

    MDAR checklist
    Supplementary file 1. ECoG and LFP recording locations.

    For each participant, we list the coordinates of the tip of the hippocampal electrode for LFP recording. In addition, in Participants 1–4, grid or strip electrodes were placed for ECoG recording. LFP: local field potential; ECoG: electrocorticography.

    elife-78677-supp1.xlsx (29.1KB, xlsx)

    Data Availability Statement

    All data needed to evaluate the conclusions in the paper are present in the paper. The codes used to produce the results in the paper are freely available at the repository https://github.com/vdimakopoulos/verbal_working_memory. The task can be downloaded at http://www.neurobs.com/ex_files/expt_view?id=266. Part of the data has been published earlier [36]. Additional data and code are indexed in http://www.hfozuri.ch/.

    The following previously published dataset was used:

    Boran E, Fedele T, Steiner A, Hilfiker P, Stieglitz L, Grunwald T, Sarnthein J. 2020. Dataset of human medial temporal lobe neurons, scalp and intracranial EEG during a verbal working memory task. Scientific Data.


    Articles from eLife are provided here courtesy of eLife Sciences Publications, Ltd

    RESOURCES