Persistent hippocampal neural firing and hippocampal-cortical coupling predict verbal working memory load

Abstract

The maintenance of items in working memory relies on persistent neural activity in a widespread network of brain areas. To investigate the influence of load on working memory, we asked human subjects to maintain sets of letters in memory while we recorded single neurons and intracranial encephalography (EEG) in the medial temporal lobe and scalp EEG. Along the periods of a trial, hippocampal neural firing differentiated between success and error trials during stimulus encoding, predicted workload during memory maintenance, and predicted the subjects' behavior during retrieval. During maintenance, neuronal firing was synchronized with intracranial hippocampal EEG. On the network level, synchronization between hippocampal and scalp EEG in the theta-alpha frequency range showed workload dependent oscillatory coupling between hippocampus and cortex. Thus, we found that persistent neural activity in the hippocampus participated in working memory processing that is specific to memory maintenance, load sensitive and synchronized to the cortex.

Fig. 1. Task, behavioral results, and recording sites.

(A) In the task, sets of consonants were presented and had to be memorized. The set size (4, 6, or 8 letters) determined WM workload. In each trial, presentation of a letter string (encoding period, 2 s) was followed by a delay (maintenance period, 3 s). After the delay, a probe letter was shown, and subjects indicated whether the probe was presented during the encoding period. (B and C) Behavioral results. (B) Accuracy of all sessions. The jitter reflects the rank order of the accuracy in trials with set size 8. Each dotted line connects an individual session. The nine subjects performed a total of 36 sessions. (C) Median reaction time (relative to onset of the probe letter) as a function of workload. The jitter reflects the rank order of the reaction time in trials with set size 8. Each dotted line connects an individual session. In (B) and (C), thick and light blue lines represent the means and SEM across all sessions, respectively. (D) Location of the microelectrodes at the tip of the depth electrodes in MNI’s MNI152 space (see Methods). Recording locations are projected on the parasagittal plane x = −25.2 mm and are color-coded (cyan, hippocampus; magenta, entorhinal cortex; yellow, amygdala).

Fig. 2. Persistent activity of maintenance neurons.

(A) Example of a maintenance neuron recorded from hippocampus. Top: Peristimulus time histogram (bin size, 500 ms; step size, 20 ms). Shaded areas represent ± SEM across trials of all spikes associated with the neuron. Inset: Mean extracellular waveform ± SEM. Middle: Periods of significance (black) between low-workload trials (set size 4, 45 trials) and high-workload trials (set sizes 6 and 8, 46 trials; P < 0.05, cluster-based nonparametric permutation test). Bottom: Raster plot of trials reordered to set size and RT for plotting purposes only. Compared to set size 4 (blue), the neuron fires more for set size 6 (green) and set size 8 (red). Figure S1 provides further examples of single neurons. (B) Percentage of all recorded neurons identified as maintenance neurons in the hippocampus (Hipp), entorhinal cortex (Ent), and amygdala (Amg). The number of maintenance neurons is provided below the area label. The hippocampus contained significantly higher proportions of maintenance neurons than the two other areas. (C) Percentage of maintenance neurons for which the firing rate during maintenance differed as a function of load. The hippocampus stood out as containing a significant percentage of neurons that increased their firing rate under high workloads. For (B) and (C), significance was assessed by comparing with a null distribution with permuted labels. ***P = 0.002.

Fig. 3. The activity of probe neurons is related to WM retrieval.

(A) Example probe neuron recorded from the entorhinal cortex. Firing rates are shown separately for trials in which the probe was held (IN, cyan) or not held (OUT, magenta) in memory. Top: Peristimulus time histogram (bin size, 200 ms; step size, 20 ms; shaded areas represent ± SEM). Bar: Time with significant differences between IN and OUT trials (P < 0.05, cluster-based nonparametric permutation test). Bottom: Raster plot of reordered trials, black marker at button press. Button press was at median RT 1.1390 s after probe letter presentation. (B) Same neuron as in (A), with trials aligned to button press. The peak response was reduced (27 Hz versus 18 Hz; permuted t test, P = 0.005). (C) Firing rates of an example probe neuron recorded from the entorhinal cortex, shown separately for low-workload trials (set size 4, blue) or high-workload trials (set size 6, green, and set size 8, red). Bar: Time with significant differences between trials of low and high workloads (P < 0.05, cluster-based nonparametric permutation test). Bottom: Raster plot of reordered trials, black marker at button press. (D) Same neuron as in (C), with trials aligned to button press (button press was at median RT 1.1390 s after probe letter presentation). The peak response was reduced (15 Hz versus 10 Hz; permuted t test, P = 0.005). (E) Percentage of probe neurons in each area. Probe neurons were most frequent in the entorhinal cortex. Significance was assessed by comparing with a null distribution with permuted labels. ***P = 0.001.

Fig. 4. Population firing predicts behavior.

(A) Mean trajectories in the neuronal state space constituted by the three largest dPCs during fixation (starting at the origin), encoding, maintenance, and retrieval. Set sizes are color-coded (4, blue; 6, green; 8, red) (see also fig. S3 and movie S1). (B) Multidimensional speed of the population in the four periods of a trial. The speed during maintenance was slowest. (C) Multidimensional pairwise distance between all possible pairs of attractors during the different periods of the task. The distance during maintenance was largest. The combination of low speed and large mutual distance provides evidence for attractors in state space during maintenance. The analysis in (B) and (C) included the first 15 dPCs that explain 79% of the signal variance. Significance was assessed by permutation t test. ***P = 0.0005. (D) During the encoding period of trials with high workloads (set sizes 6 and 8), the activity of neurons in the hippocampus (but not entorhinal cortex or amygdala) carried information about whether the subject would later respond correctly or not. (E) During maintenance, workload (set size 4 versus set sizes 6 and 8) could be decoded from hippocampal neurons. Sufficient decoding was possible even when using only those hippocampal neurons that we had identified as maintenance neurons. (F) During retrieval (−500 to 0 ms relative to button press), the subpopulations in all recorded areas were predictive of the response IN or OUT. In (B) to (F), boxplots represent quartiles (25%, 75%); horizontal lines represent medians; whiskers show ranges up to 1.5 times the interquartile range; dots outside whiskers show outliers. Markers below bars indicate significance versus chance performance. Significance was assessed on the basis of a null distribution with scrambled labels (gray boxes). ***P = 0.002.

Fig. 5. Neuronal responses in the WM network in subject 1.

(A) The task induced high alpha power in the scalp EEG at electrode site Pz, predominantly during maintenance (only trials with set size 8 in this analysis; n = 55). (B and C) During the last 2 s of the maintenance period, EEG power was higher for high-workload trials (set size 8, red line) than for low-workload trials (set size 4, blue line; n = 61 trials) in (B) the band (8.5 to 16 Hz, black bar) at electrode site Pz and (C) the band (8 to 12.5 Hz, black bar) at electrode site Fz. (D) The task increased the PLV between the hippocampal iEEG and the scalp EEG at electrode site Pz specifically in the alpha band and during the last 2 s of the maintenance period (−2 to 0 s, only trials with set size 8). (E) During this period, the PLV between the hippocampus and Pz was higher for trials with high workloads (set size 4 versus set size 8) in alpha (9.0 to 12.0 Hz, black bar), and (F) the alpha PLV was maximal to occipitoparietal scalp electrodes. (G) Theta-alpha power (7.5 to 10.0 Hz) in the left hippocampal iEEG increased with workload during maintenance. (H) During maintenance, the hippocampal iEEG power for set size 8 (red) exceeded that for other set sizes in the 7.5- to 10.0-Hz band (normalized to fixation, corrected for multiple comparisons). (I) The average firing rate of maintenance neurons in this subject was higher for a high workload during encoding and maintenance (n = 5 hippocampal neurons with more than 100 spikes during both encoding and maintenance). (J) Spike-field coherence between the LFP and the neurons from (I) appears in the alpha frequency range during maintenance (thin line, individual neurons; thick cyan line, average across neurons) but not during encoding (thick gray line, average across neurons). (B, C, E, and G to I) Black bar: Frequency range of increase with workload (set size 8 versus set size 4); magenta bar: frequency range of significant PLV; golden bar: frequency range of gating (maintenance versus encoding). P < 0.05, cluster-based nonparametric permutation test.

Fig. 6. Maintenance-induced hippocampal-cortical PLV.

(A to H) PLV spectra for the electrode pair of maximal PLV for subjects 2 to 9; for subject 1, refer to Fig. 5E [set sizes: 4 (blue), 6 (green), and 8 (red)]. High PLV between hippocampus iEEG and scalp EEG appeared for set size 8 in all subjects during the last 2 s of the maintenance period (magenta bar: frequency range of significant PLV with P < 0.05, randomization test against a null distribution with scrambled trials). PLV was higher for high-workload trials (set size 8) than for low-workload trials (set size 4) in eight of nine subjects (black bar: frequency range of elevated PLV with P < 0.05, randomization test against a null distribution with scrambled labels). The PLV was higher during the last 2 s of the maintenance period than during encoding (golden bar: frequency range of elevated PLV with P < 0.05, randomization test against a null distribution with scrambled labels). (I) Maximal PLV for each subject during retention plotted against the location of the electrode site along the anterior-posterior hippocampal axis. Electrodes in left (right) hippocampus are marked by circles (squares). Marker face colors denote subjects. The linear fit shows a gradual increase in the PLV toward the posterior end of the hippocampal axis (34 electrode pairs; R² = 0.118, P = 0.0467, permutation t test).