Gamma oscillations during episodic memory processing reveal reversal of information flow between the hippocampus and prefrontal cortex

A critical and emerging question in human episodic memory is how the hippocampus interacts with the prefrontal cortex during the encoding and retrieval of items and their contexts. In the present study, participants performed an episodic memory task (free recall) while intracranial electrodes were simultaneously inserted into the hippocampus and multiple prefrontal locations, allowing the quantification of relative onset times of gamma band activity in the cortex and the hippocampus in the same individual. We observed that in left anterior ventrolateral prefrontal cortex (aVLPFC) gamma band activity onset was significantly later than in the hippocampus during memory encoding, whereas its activity significantly preceded that in the hippocampus during memory retrieval. These findings provide direct evidence to support models of prefrontal-hippocampal interactions derived from studies of rodents, but suggest that in humans, it is the aVLPFC rather than medial prefrontal cortex that demonstrate these reciprocal interactions.


24
Prefrontal monitoring and control during episodic memory processing is thought to be critical for 25 contextually mediated memory retrieval (Miller, 2013;Preston and Eichenbaum, 2013). An influen-26 tial model characterizing one of the mnemonic roles of the prefrontal cortex (PFC) -termed here 27 the reciprocal flow hypothesis -posits that during memory encoding, contextual information flows 28 from the hippocampus to the PFC while during retrieval, the PFC uses this stored information to 29 guide selection of a contextually appropriate hippocampal memory representation (Desimone and  pothesis that the VLPFC is necessary for generating retrieval cues during episodic memory search 36 (Kim, 2019), consistent with rodent findings, and lesion studies suggest that patients with frontal 37 lobe dysfunction have difficulty recalling items when the context is altered between encoding 38 and subsequent retrieval (Chao, 1997;Fletcher, 2001). However, to date there is no direct human 39 1 of 12 Manuscript submitted to eLife electrophysiological evidence of reversed lags in the timing of hippocampal and PFC activation that 40 would be indicative of differential information flow during encoding and retrieval. fMRI studies lack 41 sufficient temporal resolution to identify such an effect, precise source localization of MEG signals 42 to different mesial temporal structures is problematic, and the absence of direct homology between 43 rodent and human prefrontal cortex means that human intracranial EEG studies are necessary to 44 establish whether this phenomenon is characteristic of human episodic memory and to determine 45 in which brain regions it may occur. 46 Encoding Retrieval CAT SAND SAND TREE TREE 3 + 7 + 1 = ? 1 + 4 + 5 = ? 6 + 2 + 5 = ?
. . . . . . . . .    55 Here we sought evidence of reversal of information flow between the hippocampus and pre-56 frontal cortex during the encoding versus the retrieval of episodic memories. We did this by taking 57 advantage of a unique dataset obtained from 77 human patients implanted with stereo EEG elec-58 trodes for seizure mapping purposes who performed a verbal free recall paradigm. During the 59 study and recall phases of the task, we identified activation peaks in gamma oscillations from 40 to 60 120 Hz, using the onset of gamma activation as an estimate of the initial timing of activity in a given 61 brain region. As our data set included subjects with electrodes implanted in both the hippocampus 62 and PFC (in addition to other cortical locations), we were able to directly compare the timing of 63 2 of 12 Manuscript submitted to eLife memory-related gamma activation in the PFC and hippocampus within-subjects.

65
Behavioral Performance 66 Across participants, the average probability of recall for all words was 24.4%. The average percent-67 age of list intrusions (recall errors) per subject was 12.8%. We derived an estimate of temporal 68 clustering (the tendency for items adjacent to each other in the study list to be recalled sequentially) 69 to determine if temporal contextual factors were operating at retrieval (Watrous and Ekstrom, For each memory condition, a black region border indicates that Δ across all electrode pairs is significantly different than zero (t-test, FDR corrected < 0.011 for encoding and < 0.007 for retrieval). An asterisk (*) between the encoding and retrieval conditions indicates that the encoding and retrieval Δ are significantly different when compared with a paired t-test (FDR corrected < 0.019). The left aVLPFC exhibited a mean activation lag relative to the hippocampus during successful encoding of +13.4 msec (FDR corrected < 0.001, t-test of activation times compared to hippocampus), and a reversal of this effect during retrieval, such that the region led the hippocampus by -10.4 msec (FDR corrected = 0.0116). Moreover, the Δ distributions for encoding and retrieval were significantly different (FDR corrected < 0.001) across electrodes when compared with a paired t-test. Of the remaining PFC regions, only one other region, the right DLPFC, exhibited a Δ that was significantly greater than zero during encoding (FDR corrected = 0.0037); however this region did not exhibit a reversal in Δ values during retrieval (with the hippocampus leading during both encoding and retrieval). The left ACC exhibited a significant difference in the distribution of Δ during encoding vs. retrieval (-6.86 msec during encoding, 6.46 msec during retrieval; FDR corrected = 0.0173); but the Δ was not significantly different than zero during neither encoding (FDR corrected = 0.0700) or retrieval (FDR corrected = 0.0700) indicating onset nearly commensurate with that of the hippocampus. The pattern of Δ reversal was evident for the left but not the right aVLPFC. In the latter region, while there was a significant difference between encoding and retrieval (FDR corrected = 0.0080), the values indicated that the cortex led the hippocampus during both phases of the free recall task (lag = -23.5 msec during retrieval, -7.44 msec during encoding).  Figure 3. Subsequent memory effect in Δ for all regions. Z-scores were calculated for each region using a paired t-test between Δ for recalled words and Δ for non-recalled words. The subsequent memory effect for left aVLPFC, left DLPFC, and left ACC was significant (FDR corrected p < 0.007), which is indicated by black borders for those regions. No regions in the right hemisphere show a significant subsequent memory effect. Manuscript submitted to eLife an SME), with a pattern that fits a putative model of the transfer of contextual information to the 104 frontal cortex during successful encoding (significantly positive relative to hippocampal activation) 105 with evidence of a reversal during retrieval (significantly negative relative to the hippocampus). 106 Across all electrode pairs, 38% of aVLPFC electrodes exhibited this pattern of Δ reversal, which 107 was significantly greater than the fraction exhibiting this effect in the DLPFC ( 2 (1, N=630) = 17.160, 108 < 0.001) (Figure 4). Across the subjects who contributed an electrode pair to the left aVLPFC, 59% 109 showed a pattern of Δ reversal in at least one electrode pair.  Δ for all aVLPFC electrodes during the recalled and non-recalled conditions. The Δ for non-recalled words is not significantly different than zero (mean non-recalled lag is +2.91 msec). (B) 38% of aVLPFC electrodes have Δ reversal between conditions, with the hippocampus leading in activation during encoding and the cortex leading during retrieval, and 10% show the opposite pattern of activation, with the cortex leading in activation during encoding and the hippocampus leading during retrieval. 52% of electrodes show no reversal in lag between conditions. (C) Histograms of the mean during encoding (subsequently recalled only) and retrieval for the 38% of aVLPFC -HIPP electrode pairs exhibiting the effect depicts the differences in timing of activation onset for hippocampal and aVLPFC electrodes between memory conditions. Further, we analyzed prior list intrusions (PLI) to test more directly whether Δ reversal is 111 associated with the transmission of contextual information, as hypothesized by the reciprocal 112 flow model. List intrusions represent errors of item-context association (the wrong item for a 113 given context, i.e. the list on which the item was presented, although we discuss caveats to the 114 interpretation of PLI data in the Discussion below). For this analysis, oscillatory activity can be 115 analyzed only during item retrieval. We observed no evidence of information reversal for PLI events, 116 with the onset of left aVLPFC activation not significantly different than for the hippocampus (mean 117 Δ = -3.5 msec, uncorrected = 0.3861). In addition, the Δ during correct retrieval events was 118 significantly less than that of PLI events (uncorrected = 0.0436).

119
In a convergent approach, we looked for evidence of lagged activation using a different method, 120 this time employing the lagged correlation of the gamma power envelope (rather than activation 121 onset), following established methods (Ossandon et al., 2011). With this approach, which has a very 122 different rationale and underlying set of assumptions than the lagged activation analysis described 123 above, the results for the left aVLPFC were highly consistent with those observed previously, with a 124 significant positive lag of 14 msec during encoding (uncorrected = 0.0442) (hippocampus leading) 125 and negative 8 msec lag during retrieval (in the same direction as our initial analysis with the PFC 126 5 of 12 Manuscript submitted to eLife leading, though it did not reach significance -uncorrected = 0.1426). As with the previous method, 127 the encoding versus retrieval lag distributions were significantly different (FDR corrected = 0.0276). specification consistent with onset of activation preceding the hippocampus during retrieval, as 146 we observed. In the setting of the free recall paradigm, cue specification presumably includes 147 specification of temporal contextual information (Sederberg et al., 2010). We acknowledge however 148 that our data by itself does not allow us to make strong claims regarding the content of information 149 characterized by reciprocal flow. 150 Since we did not employ an experimental manipulation that allow a distinction to be made 151 between verbal information and its contextual features, it is possible that our observations could 152 be consistent with a contribution from the aVLPFC at encoding that is specifically related to the 153 verbal features of the items such as cue processing (e.g. selecting a specific meaning for an item), 154 at retrieval, the transmission of a non-specific 'biasing' signal that favors a hippocampal "retrieval 155 mode" (e. g. Lepage et al., 2000). 156 That being said, in the rodent investigations that motivated our analysis, similar observations 157 to ours were interpreted as evidence of the transmission of contextual information, and the lack 158 of a significant reversal in aVLPFC during retrieval of list intrusions is consistent with the proposal 159 that lag reversal in this region is related to the transmission specifically of temporal contextual 160 information. We acknowledge however that this inference is weakened by the concern that list 161 intrusions for freely recalled items, may occur for a variety of reasons other than the failure to Participants preformed a free recall task consisting of multiple study/test cycles. During the study 229 period, 12 words from a pre-selected pool of high-frequency, single-syllable, common nouns were 230 visually presented, one at a time, on a computer screen for a duration of 1.6 s followed by a blank 231 screen of 4 s with 100 msec of random jitter. Subjects were instructed to study each word as it 232 appeared on the screen. The presentation of the last item in a list was followed by a 30 s period 233 during which a math distractor task (A + B + C = ??) was performed to limit rehearsal. Participants 234 were then instructed to verbally recall as many items as possible from the immediately prior list in 235 no particular order. A full session consists of 12 full study/test cycles and 1 practice study/test cycle 236 which was excluded from analysis. One complete session yielded electrophysiological recordings 237 from 144 word encoding epochs (12 lists x 12 words) and a variable number of retrieval epochs. 238 Participants performed between 1 and 9 sessions of the free-recall task over several days (median 239 number 2). 240 We used the temporal clustering factor, which is a measure of temporal contiguity for each recall Stereo-EEG data were recorded using a Nihon Kohden EEG-1200 clinical system. Signals were 248 sampled at 1000 Hz and referenced to a common intracranial contact. Raw signals were subse-249 quently re-referenced to a bipolar montage, with each contact referenced to the superficial adjacent 250 contact. All analyses were conducted using MATLAB with both built-in and custom-made scripts. 251 We employed an automated artifact rejection algorithm to exclude interictal activity and abnormal 252 trials (kurtosis threshold greater than 4). The raw signals were filtered for noise on a session by 253 session basis using the following steps: 1) the power spectral density was estimated across the 254 entire session, 2) a 7th order polynomial was fit to the power spectral density estimate to obtain a 255 trend line, 3) the trend line was subtracted from the power spectral density estimate to identify 256 peaks in the periodogram, and 4) for each peak above 15 dB, the local minima surrounding the 257 peak were used to define the cutoff frequencies for a first-order Butterworth notch filter. The notch 258 filter identified for each peak in the periodogram was applied sequentially to the raw data. Retrieval 259 trials were isolated such that each included trial was isolated from any other retrieval events by 260 at least 1200 msec before the onset of vocalization and 200 msec after the onset of vocalization 261 (this led to the exclusion of 3,258/12,791 [25.5% trials]). Identification of retrieval events followed 262 previously published methods (Burke et al., 2014). 263 To assign bipolar electrode contacts to regions of interest, electrodes were defined as being in 264 the hippocampus or one of the five PFC regions if at least 1 of the bipolar contacts was determined to 265 lie within the structure. To compare activation onset times between electrodes in the hippocampus 266 and the PFC for a given subject, each electrode contact within a given PFC region was paired with all 267 hippocampal contacts for that subject (PFC-hippocampal electrode pairs). This was repeated for all 268 regions.

269
Activation Onset Detection (Calculation of ) 270 We compared the temporal patterns of high gamma band power changes in the hippocampus 271 and frontal cortex in the 1000 msec immediately following study item presentation (encoding) and  (Dastjerdi et al., 2011). This was repeated for each frequency band and trial, and the was 298 averaged across all bands and then across all trials for the recalled epochs, non-recalled epochs, 299 and retrieval epochs to obtain a single time estimate of activation onset, , for each hippocampal 300 and PFC electrode and each condition (see Figure 2). We repeated our analysis using a threshold of Manuscript submitted to eLife band was subtracted to obtain the mean-centered instantaneous amplitude envelope. For each 321 hippocampal-PFC electrode pair, the normalized cross-correlation was calculated on the mean-322 centered amplitude envelope on a trial-by-trial basis using an 800 msec moving window with a 1 323 msec step size and a maximum lag of 150 msec. The initial 800 msec cross-correlation moving 324 window for the encoding epochs was centered 100 msec prior to word presentation and stepped by 325 1 msec until 1700 msec after word presentation. The initial retrieval cross-correlation was centered 326 1100 msec prior to word vocalization and stepped by 1 msec until 100 msec after word vocalization. 327 Thus the resulting matrices for the cross-correlations of a single trial were 301 by 1800 for encoding 328 and a 301 by 1200 for a retrieval. For each electrode pair, the correlation coefficients were Fisher 329 transformed across the recalled, non-recalled, and retrieval trials separately to allow comparison 330 across trials (Cohen et al., 2003). Next   Z-scores were calculated for each region using a paired t-test between gamma power for recalled words and non-recalled words. A positive z-score indicates that the non-recalled gamma band power is greater than recalled gamma band power.