Task, behavior and recording channels
Fourteen patients with drug resistant epilepsy (7 females) (Table 1) performed a modified Sternberg WM task (65 total sessions from 14 subjects) during an invasive presurgical evaluation. In this task, the items were presented simultaneously rather than sequentially, thus separating the encoding period from the maintenance period. In each trial, the subject was asked to memorize a set of 4, 6, or 8 letters presented for 2 s (encoding). After a delay (maintenance) period of 3 s, a probe letter was presented and the subject responded whether the probe letter was identical to one of the letters held in memory (retrieval) (Fig. 2(A)). The average accuracy was 91.9%±3.2% (range 86.1%-97.6%). The mean response time was faster for correct than incorrect trials (1.44 ± 0.36 versus 1.95 ± 0.66 seconds, paired t test: t (13) = -4.15, p = 0.0011). Hence, the subjects performed well in the task. Unless stated otherwise, we report results for correct trials only.
Table 1
Subject | Age | Gender | Recording sites (amygdala) | Recording sites (hippocampus) | Retrieval Accuracy | RT for correct trials (mean ± std) |
1 | 24 | Female | 2 | 6 | 0.925 | 1.26 ± 0.63 |
2 | 39 | Male | 4 | 8 | 0.864 | 1.28 ± 0.65 |
3 | 18 | Female | 2 | 6 | 0.932 | 1.10 ± 0.34 |
4 | 28 | Male | 4 | 8 | 0.949 | 1.48 ± 0.63 |
5 | 20 | Female | 2 | 4 | 0.913 | 1.41 ± 0.90 |
6 | 31 | Male | 4 | 8 | 0.943 | 1.47 ± 0.59 |
7 | 47 | Male | 4 | 8 | 0.949 | 1.44 ± 0.57 |
8 | 56 | Female | 4 | 7 | 0.900 | 1.41 ± 0.51 |
9 | 19 | Female | 4 | 8 | 0.928 | 1.24 ± 0.33 |
10 | 35 | Male | 4 | 6 | 0.930 | 1.64 ± 0.73 |
11 | 51 | Female | 4 | 6 | 0.914 | 1.30 ± 0.58 |
12 | 30 | Male | 4 | 7 | 0.895 | 1.49 ± 0.77 |
13 | 29 | Female | 4 | 4 | 0.976 | 1.05 ± 0.28 |
14 | 56 | Male | 4 | 8 | 0.861 | 2.58 ± 1.22 |
Local field potentials (LFPs) were recorded simultaneously from depth electrodes implanted in the amygdala and hippocampus (Fig. 2(B)). In total from all subjects, we recorded from 94 channels in the hippocampus and 50 channels in the amygdala (see the details in Methods).
Working memory induced oscillatory response in amygdala and hippocampus
We calculated the time-frequency power for each channel within the amygdala and the hippocampus, and the power outputs were z-scored against pretrial (500 ms) baseline distributions to assess the significance of the task-induced power effects per trial. Both the amygdala and the hippocampus showed sustained activity in the frequency range (1–40 Hz), especially during the encoding and the maintenance phase (Fig. 3(A)). We then focused on the temporally resolved changes in amplitudes of 1–40 Hz in the amygdala and the hippocampus (Fig. 3(B)). Both amygdala and hippocampus showed elevated oscillatory activities (1–40 Hz) relative to the baseline (p < 0.05, cluster-based permutation test), suggesting that both regions are engaged in the working memory processing. The results in the present study match the findings reported in our previous study 25.
We next asked whether the engagement of the amygdala and the hippocampus was preferentially associated with a particular task period. We averaged the spectral power during encoding and maintenance and compared the two spectra. The z-scored power in the hippocampus during maintenance was higher than during encoding in the theta-alpha frequency band (3–13 Hz, gray shaded window in Fig. 3(C) right, p < 0.05, cluster-based permutation test). No significant difference between task periods was found in the amygdala (p > 0.05, cluster-based permutation test, Fig. 3(C) left). Next, we separately extracted the theta-alpha power (3–13 Hz) during encoding and maintenance for amygdala and hippocampus and performed a repeated measure ANOVA (RMANOVA) with the extracted z-power as the dependent variable and two within-subject factors, area (amygdala/hippocampus) and task period (encoding/maintenance), and their interaction as independent variables. As shown in Fig. 3(D), we found a significant interaction (p = 0.001, F (1,13) = 19.043). A simple effect analysis further indicated that the hippocampus power in the maintenance was significantly higher than in the encoding (p = 0.0017) while no significant difference was found between encoding and maintenance in the amygdala (p = 0.20); the power in the amygdala was higher than in the hippocampus during encoding (p = 0.017), and the power in the hippocampus was higher than in the amygdala during maintenance but this did not reach statistical significance (p = 0.054). Together, the univariate analyses indicate that the amygdala was preferentially engaged in encoding and the hippocampus was preferentially engaged in maintenance.
Functional specialization: Representation in the amygdala is specific during encoding
We next applied multivariate analysis to investigate how information is represented across multiple channels and different spectral powers at different times. We performed a series of representational similarity analyses (RSA) to investigate two representational properties crucial for memory performance, i.e., the distinctiveness and the stability of neural representations 26.
We first quantified representational dissimilarity separately for the amygdala and hippocampus. Given that both the amygdala and hippocampus showed oscillatory activities across 1–40 Hz (Fig. 3(A)), we used the spectral power in this frequency band as the feature for the RSA. We correlated the power from every two trials across channels and frequencies (1 to 40 Hz in steps of 1 Hz) in consecutive overlapping time windows of 100 ms (step width 10 ms Fig. 2(C)). Then the dissimilarity (1 - similarity) of the representational patterns was averaged across all trial-pairs, resulting in a temporal map of encoding-encoding dissimilarity (EED) across all subjects for the amygdala (Fig. 4(A) left) and the hippocampus (Fig. 4(A) right) separately. Next, we compared the EED map between the amygdala and the hippocampus using cluster-based permutation tests. Across all encoding time windows, a cluster with higher EED values in the amygdala than in the hippocampus appeared (outlined in black in Fig. 4(B), p = 0.007, cluster-based permutation); there was no cluster with higher EED values in the hippocampus than in the amygdala. Further, we averaged the EED values in significant clusters for all subjects and compared them for the amygdala and the hippocampus. This revealed higher EED values in the amygdala than in the hippocampus across subjects (p = 0.027, t(13) = 2.49, paired t test, Fig. 4(C)). As the letter strings were different across trials, these findings indicate that the amygdala represented distinct WM information in a more specific neural pattern during encoding.
Functional specialization: Hippocampal representation is stable from encoding to maintenance
We next examined whether and how stable representational structures were maintained in the absence of stimuli in the amygdala and the hippocampus. Representational stability was indexed by memory reinstatement, an approach borrowed from the long-term memory literature 27. Memory reinstatement was quantified as the correlation between patterns of oscillatory power across channels and frequencies (1 to 40 Hz in steps of 1 Hz) within consecutive overlapping time windows of 100 ms (step width 10 ms) for each combination of the encoding-maintenance time bins in the same trial. The correlation matrix was then averaged across trials, resulting in a temporal map of encoding-maintenance similarity (EMS) for the amygdala and the hippocampus (Fig. 2(D)). The EMS values were higher in the hippocampus than in the amygdala for every encoding-maintenance time pair in the EMS map (Fig. 4(D)). And, the averaged EMS values in the hippocampus was higher than that in the amygdala across subjects (p = 0.0049, t(13) = 3.38, paired t test, Fig. 4(E)). These findings indicated that the hippocampus retained WM information in a more stable representation during maintenance than the amygdala.
Was the functional specialization of the amygdala and the hippocampus correlated? To answer this question, we carried out a correlation analysis between the EEDAmy−Hipp (EEDAmy - EEDHipp) and the EMSHipp−Amy (EMSHipp - EMSAmy) for each subject. We found a significantly positive correlation (Spearman’s correlation, p = 0.0067, r = 0.701, Fig. 4(F), which remained significant after removing an outlier (p = 0.025)). This suggests that in subjects in which the amygdala represented more WM information, the hippocampus maintained the representation more stably.
Functional coordination: Coordination of the amygdala and hippocampus in representational reinstatement
In addition to functional specialization, functional coordination between brain regions is considered to be another important mechanism in WM 28. To test whether the amygdala and hippocampus work independently or interactively in WM, we investigated whether the representational structure during encoding in one structure was reinstated in the other structure during maintenance and vice versa.
We calculated cross-region EMS on each channel pair (one from the amygdala, one from the hippocampus in the same hemisphere in the same subject). The cross-region EMS maps were then averaged across channel pairs and trials, resulting in two cross-region EMS maps, one between the hippocampal encoding-amygdala maintenance combination (Fig. 5(A), left), the other between the amygdala encoding-hippocampal maintenance combination (Fig. 5(A), right). We found substantial cross-region EMS between the amygdala and the hippocampus (Fig. 5(A), rho > 0). Next, we compared the two cross-region EMS maps using cluster-based permutation tests and found a significant cluster (outlined in black in Fig. 5(B), p = 0.015, cluster-based permutation test), showing that stronger hippocampal representational structures during encoding were reinstated by the amygdala during maintenance. The averaged EMS values within the significant clusters were higher in the hippocampal encoding-amygdala maintenance direction than in the opposite direction across subjects (p = 0.0015, t(13) = 4.006, paired t test, Fig. 5(C)). These results provide evidence that the amygdala and the hippocampus coordinate their activity to maintain the representation and that the hippocampus conveys information to the amygdala.
Since we found both functional specialization and coordination between the amygdala and the hippocampus in WM, we then tested for a possible correlation between the functional differentiation and collaboration. Spearman’s correlation was separately carried out between the coordinated EMS and the EEDAmyg−Hipp and between the coordinated EMS and the EMSHipp−Amyg. Intriguingly, we found a significant positive correlation between the coordinated EMS and the EMSHipp−Amy (p = 0.015, r = 0.644, Fig. 5(D)), but the positive correlation between the coordinated EMS and the EEDAmyg−Hipp was not significant (p = 0.120, r = 0.437, Fig. 5(E)). These findings suggest that in subjects with more information maintained in the hippocampus, the activities between the two regions were more coordinated.
Functional coordination: Directional information transfer from the hippocampus to the amygdala tracks WM processing
Cross-region EMS analysis found that WM representation in one region is reinstated in the other. This suggest that WM representations are distributed and coordinated across the amygdala and hippocampus. To further investigate the directionality of information transfer between these regions, we computed the spectral Granger causality (GC) index from the hippocampus to the amygdala and that in the reverse direction, during encoding and maintenance separately. The GC for both directions was significantly above the threshold (see Methods for details, grey lines denote the thresholds in Fig. 6(A)). Moreover, the GC from in the hippocampus to the amygdala was higher than the GC in the reverse direction in the 2–40 Hz band both during encoding (Fig. 6(A) top, black line, p = 0.014, cluster-based permutation test) and during maintenance (Fig. 6(A) bottom, black line, p = 0.004, cluster-based permutation test). The averaged GC values from hippocampus to amygdala were also higher than the reverse direction during encoding (p = 0.0072, t(13) = 2.88, paired t test, Fig. 6(B) top) as well as during maintenance (p = 0.0015, t(13) = 2.58, paired t test, Fig. 6(B) bottom) across subjects. These findings match the coordinated reinstatement and provide converging evidence of hippocampus-driven directional hippocampal-amygdala communication during WM maintenance. Notably, to exclude the volume conduction effect, we used the bipolar rereferencing scheme and obtained a highly similar pattern of hippocampus-driven information flow during encoding and maintenance (Fig. S1).
Functional specialization and coordination in the amygdala-hippocampal circuit support successful WM outcomes
Functional specialization and coordination were found in the amygdala-hippocampal circuit during WM processing for the correct trials. However, it was still unclear whether the two properties were specific to successful WM processing. As a control, we repeated the above analyses for the incorrect trials.
We calculated the z-scored power within the amygdala and the hippocampus across 3–13 Hz during encoding and maintenance for the incorrect trials and compared them using RMANOVA with two within factors: area (amygdala/hippocampus) × phase (encoding/maintenance). No interaction effect was found for the incorrect trials (p = 0.11, Fig. 7(A)). We also calculated the EED map during encoding within the amygdala and the hippocampus for the incorrect trials. The EED map of the amygdala was compared with that of the hippocampus using cluster-based permutation, and no significant clusters were found (Fig. 7(B), p > 0.05, cluster-based permutation test). Next, we computed the EMS for the incorrect trials within the amygdala and the hippocampus separately. Contrasting the within-region EMS between the two regions revealed no difference (Fig. 7(C) left, p = 0.086, t(13) = 1.861, paired t-test). In addition, we did not find any difference in the cross-region EMS based on the error trials (Fig. 7(C) right, p = 0.17, t(13) = 1.47, paired t-test). This suggests that functional specification and coordination support successful WM processing.
To exclude possible bias from having an imbalanced number of correct and incorrect trials, we performed control analyses by bootstrapping 10 times. Specifically, we recomputed the EMS within the amygdala and the hippocampus on a subset of trials that contained the same number of correct trials and incorrect trials. These subsets of trials were randomly selected from all the correct trials, and the process was repeated 10 times. As shown in Fig. S2(A), Pearson’s correlation was computed and revealed high correlations between the EMS value for each subset and those from all the trials both in the amygdala (r = 0.975 ± .013) or the hippocampus (r = 0.976 ± .025). In each subset, the EMS within the hippocampus was higher than that within the amygdala (Fig. S2(B)). We thereby replicated our findings using a small subset of trials. Next, we recomputed the EMS across regions based on 10 subsets. Again, we found significant association between the EMS from each subset and those from all the trials, both in the combination of hippocampal encoding and amygdala maintenance (r = 0.965 ± .015, Pearson’s correlation) and in the amygdala encoding and hippocampal maintenance pair (r = 0.961 ± .014, Pearson’s correlation) (as shown in Fig. S2(C)). Again, for each subset, the EMS between the hippocampal encoding and the amygdala maintenance was higher than the EMS in the opposite direction (Fig. S2(D)). Taken together, our findings ruled out bias from having an imbalanced number of correct and incorrect trials.
To conclude, successful WM performance was promoted both by functional differentiation, including preferred stages of increased power, more distinct representation during encoding in the amygdala, and more stable representation reinstatement during maintenance in the hippocampus and by functional coordination, including biased cross-region reinstatement and directional information flow from the hippocampus to the amygdala.