We analyzed recordings from two separate experiments (experiment 1: 585 neurons in the hippocampus, 16 participants, 7 female; age mean = 36.125 years, from 26-53 years; experiment 2: 86 neurons in the hippocampus, 8 participants, 4 female; age mean = 34.375 years, from 19-58 years) where patients were implanted with stereotactic Behnke-Fried depth electrodes in the hippocampus (Figure S1), while they performed a memory association task (Figure 1B & Figure 1C).
During encoding of experiment 1 participants created a vivid mental story consisting of an animal cue and two associate images (two faces, two places or a face and a place). By contrast, experiment 2 consisted of one cue and one associate image (both either an animal, face, or place). During recall, the animal cue was presented again and participants were asked to retrieve the associate image(s). The experiments were self-paced and every episode was learned and retrieved only once.
Identifying Episode Specific Neurons (ESNs)
For every neuron, we determined the firing rate during each correctly remembered episode at encoding and retrieval. We then z-scored the firing rate across all encoding and retrieval episodes. This was done independently for encoding and retrieval to account for general differences in firing rates. We measured episode-specific firing reinstatement as the product of the standardized firing rates at encoding and retrieval (Figure 2A). Due to the self-paced nature of the experiment, we expected no consistent temporal firing pattern to emerge and instead assumed that a rate code contained enough information to represent an event.
Using a trial-shuffling procedure, we generated a distribution of reinstatement values expected by chance. A neuron was considered an ESN if (i) the empirical reinstatement value exceeded the 99th-percentile of the shuffled distribution for at least one episode and (ii) the standardized firing rate for encoding and retrieval of that episode each exceeded 1.645 (≙ pright-tailed < 0.05). The second criterion prevented the identification of ESNs which would excessively fire at only one phase of the task (i.e., encoding or retrieval). It could be argued that ESNs identified in this manner could reflect the firing of cells tuned to the image of the animal cue, rather than the conjunction of all elements since the cue is episode-unique and presented during encoding and retrieval. To address this issue, in experiment 1, we excluded neurons that showed a significant firing increase during the first second after the encoding of the animal cue for episodes that were later reinstated (see Methods). This procedure has traditionally been used to identify putative concept neurons (Quiroga et al., 2005; Mormann et al., 2008; Mormann et al., 2011). This approach revealed that in experiment 1 a significant number of hippocampal neurons were ESNs (141 out of 585 neurons ≙ 24.10%; p < 0.001; permutation test; Figure 2B). In experiment 1 117 out of 141 ESNs (≙82.98%) coded for a single episode. Two example ESNs are shown in Figure 3. These ESNs are unlikely to be concept cells tuned to the animal cue as the firing rate during encoding reaches its maximum only after the presentation of the associate stimulus (see Figure 4A). It is of note that the proportion of neurons that can be classified as ESNs is proportional to the number of events learned and retrieved (the same is the case for Concept Neurons). As such it is not suitable to determine the sparseness of the hippocampal code. The proportion of ESNs of all recorded neurons is useful as an estimation of how many ESNs we can expect in future analyses.
ESNs do not code for the (content of the) associate images
To rule out that ESNs are firing in response to a stimulus category (such as faces or places) we repeated the ESN identification analysis, but only included ESNs that showed reinstatement in at least one episode with two face associate images and at least one episode with two place associate images. Such neurons would be very unlikely to code for a specific stimulus category or concept because the stimulus material in these episodes does not overlap. Again, we found a significant number of ESNs (7 out of 585 single units ≙ 0.012%; p = 0.0004). It is important to note, that this is a very conservative analysis, as only a few ESNs reinstate two or more episodes (24 out of 141 ESNs ≙ 17.02%).
ESNs do not code for the content/visual properties of the cue or associate image
Traditionally, visually responsive neurons have been identified using the repeated presentation of a stimulus. In the above analysis, we only present the animal cue once, which is suboptimal for ruling out Concept Neurons tuned to the animal cue. To ameliorate this shortcoming, in experiment 2 we added a visual tuning task (Figure 1C) after the memory association task. During the visual tuning task, images from the memory task were repeatedly shown in quick succession. This approach has typically been used to identify putative Concept Neurons that respond to one of the images independently of any memory processes (Mormann et al., 2008; Mormann et al., 2011; Quiroga et al., 2005). When excluding Concept Neuron activity in this independent dataset, we replicated our previous results and identified a significant number of ESNs (17 out of 86 neurons ≙ 19.77%; p = 0.0202; permutation test; Figure 2C). In experiment 2 16 out of 17 ESNs (≙94.12%%) coded for a single episode.
ESNs are limited to later remembered episodes
We have so far demonstrated that ESNs reinstate their firing rate when remembering a unique episode. This reinstatement cannot be explained by the semantic content or visual properties of the used image, which strengthens the notion that ESNs code for memories. Consequently, we expected to not find a significant number of ESNs when limiting our analysis to later forgotten episodes, which was indeed the case (10 out of 585 neurons ≙ 1.71%; p = 0.4317; permutation test). However, this result could stem from a lower number of forgotten events (see Table S1). To counter this bias, we equalized event numbers between later remembered and later forgotten events for every neuron by randomly sampling (with replacement) later remembered events as many times as participants forgot an event. If any of the sampled events were later reinstated, we considered this single neuron a miss-ESN under the null. By repeating this procedure 10,000 times we generated a distribution of how many miss-ESNs were expected if the number of later remembered and later forgotten events were equal. This analysis revealed a significantly lower empirical number of miss-ESNs compared to hit-ESNs (p < 0.001, bootstrapping test), demonstrating that the likelihood of finding ESNs depended on memory.
ESNs do not code for time
Recent studies in humans show that some hippocampal neurons code specific time points invariant across repetitions, which are referred to as Time Cells (Reddy et al., 2021; Umbach et al., 2020). We investigated whether our dataset contains such Time Cells (TC) using a similar method as employed by (Umbach et al., 2020). Due to the self-paced nature of our experiment, each encoding block varied in length. To accommodate this, we used both the unaltered block length, as well as a normalized block length within one recording session (see Methods). Of all 585 recorded cells, 12 (normalized) and 9 (non-normalized) fulfilled the criteria of TCs, which is below chance level (p values > 0.8; permutation test). Critically, there was no significant overlap between neurons that behaved like TCs and ESNs (p values > 0.8; permutation test) indicating that ESNs cannot be construed as TCs.
ESNs show a wider waveshape than other neurons
Another question is if ESNs are physiologically different from other single neurons. We found some evidence that spike waveshapes of ESNs are wider than those of other units (unpaired t-test; p = 0.019; Figure S3A), possibly indicating that ESNs are physiologically different from other single neurons. In the hippocampus, a wider waveshape has previously been associated with excitatory cells (Prestigio et al., 2019). There was no significant difference in the spike height or Fano factor between ESNs and other neurons (unpaired t-tests; all p values > 0.4; Figures S3B and S4).
Recorded neurons are mostly single neurons and not multi units
Although we tried to separate multi units into single neurons as best as possible during the spike sorting procedure, some units might still represent activity from multiple neurons. We thus employed the method outlined by Tankus and colleagues (Tankus et al., 2009) to classify units into single units and multi units, using the inter-spike interval and spike waveshape variability as objective criteria. In the first experiment, 518 out of 585 units (≙ 88.55%) were classified as single units (133/141 ESNs ≙ 94.32%), while in the second experiment 59 out of 86 units (≙ 68.60%) were classified as single units (12/17 ESNs ≙ 70.59%). Thus, the vast majority (i.e., 91.77%) of the reported ESNs reflect putative single units.