Cognotyping by What–Where–When Retrieval Reveals the Potential Role of Temporal Memory and Its Neural Correlates in Understanding Individual Differences across Aging and Alzheimer Disease

Episodic memories allow us to mentally re-experience past events by binding together what happened, where it happened, and when it happened (Clayton & Dickinson, 1998; Nyberg et al., 1996; Tulving, 1972, 1993). Although our memory of various episodic components do share common cortico-hippocampal neural correlates (Nyhus & Curran, 2010; Ergorul & Eichenbaum, 2004; Shastri, 2002), previous studies have also reported functional specificity in the brain across these memory components (i.e., what, where, and when), both in the cortex and hippocampal formation. For instance, in addition to traditional studies that showed dissociations in ventral and dorsal streams in the processing of what and where information (e.g., Knierim, Neunuebel, & Deshmukh, 2014; Mishkin, Ungerleider, & Macko, 1983), studies have reported the involvement of a distinct set of regions such as the precuneus, SMA, and parietal cortex for temporal processing (D'Argembeau, Jeunehomme, Majerus, Bastin, & Salmon, 2015; Bueti & Walsh, 2009). Several studies have also directly compared across what, where, and when components of episodic memory (Cheke, Bonnici, Clayton, & Simons, 2017; Kwok & Macaluso, 2015; Holland & Smulders, 2011; Hayes, Ryan, Schnyer, & Nadel, 2004; Nyberg et al., 1996). In one study, the right dorsolateral pFC was more activated during where and when memory retrieval (Hayes et al., 2004). Another study found selective involvement of the precuneus, superior parietal cortex, and medial frontal cortex for temporal order judgment, spatial configuration judgment, and scene recognition, respectively (Kwok & Macaluso, 2015). Given the dissociability of its components, it is possible that relative strengths or weaknesses in their underlying neural mechanisms may give rise to individual variability in memory profile in both healthy populations and in neurological disorders. However, despite the numerous studies on episodic memory, the correspondence between individual episodic memory and the neural activities underlying the three components (i.e., neurocognitive mapping) has yet to be constructed.

One way in which an association between neural and cognitive function can be identified is by comparing specific neural activities across groups of participants with different behavioral patterns (e.g., good vs. bad performing groups; Goyal et al., 2018; Cerliani, Thomas, Aquino, Contarino, & Bizzi, 2017; Gevins & Smith, 2000). An alternative method is to search for such a pattern in every individual across a large number of trials to improve signal to noise ratio (Braga & Buckner, 2017; Gordon et al., 2017). Both methods have their pros and cons; the former method can overgeneralize effects by missing individual characteristics, whereas the latter can be too specific to the individual, thereby hampering an interpretation based on general principles of how the brain works.

In this study, we aimed to balance these two approaches in investigating episodic memory profiles by clustering participants into groups based on their what, where, and when components, so as to not lose individual variability while maintaining generalizability. A similar approach has been taken in clinical research in which subtyping a disorder into three or four groups of patients revealed new neural markers associated with characteristic pathological symptoms of each subgroup (Vogel et al., 2021; Young et al., 2018). Likewise, we aimed to group participants into “cognotypes” that can be traced back to specific neural correlates guiding their memory of each component. In addition, we explored the possibility that particular strengths and weaknesses could signal vulnerability or resistance to age-related decline in memory. In particular, given previous studies that suggest the particularly fragile nature of temporal order memory (compared with what or where memory) in Alzheimer disease (J. H. Park & Lee, 2021; El Haj & Antoine, 2018; Pirogovsky et al., 2013; Bellassen, Iglói, de Souza, Dubois, & Rondi-Reig, 2012), finding specific behavioral and neural markers of individual differences in temporal memory may be useful for both scientific and diagnostic purposes.

Considering the complexity of cognition and the variety of factors that influence one's performance on a behavioral task, it is important to identify easily accessible neural correlates of cognition that give us biological grounds for interpreting patterns of behavior. With its wide accessibility for aging and patient participant groups, scalp EEG can be a powerful tool for providing such neural markers of episodic memory. Theta power (2.5–8.5 Hz) from EEG is widely known for its important role in communication between cortical and deep brain regions, which can contribute to promoting episodic memory performance (Wang, Schmitt, Seger, Davila, & Lega, 2021; Eichenbaum, 2017; Backus, Schoffelen, Szebényi, Hanslmayr, & Doeller, 2016; Hasselmo & Stern, 2014; Lega, Jacobs, & Kahana, 2012; Nyhus & Curran, 2010; von Stein & Sarnthein, 2000). In addition, previous studies showed that discriminability based on spatial and nonspatial “cognitive distances” (e.g., semantic and temporal distance [TD]) may be represented through the modulation of both hippocampal theta from intracranial EEG (Herweg, Solomon, & Kahana, 2020; Solomon, Lega, Sperling, & Kahana, 2019) and cortical theta power from scalp EEG (S.-E. Park, Lee, & Lee, 2023; Liang, Zheng, Isham, & Ekstrom, 2021). One study showed that hippocampal theta-mediated TD coding was associated with better verbal free recall (Solomon et al., 2019). Therefore, both theta power and theta-modulation by cognitive distance can be informative neural markers of episodic memory that may vary across individuals, depending on their behavioral memory performance.

In Experiment 1, a total of 47 college students (age 21–30 years, 21 women) were recruited to perform a scene-based episodic memory task while their scalp EEG signals were recorded. The number of participants was decided across previous studies that defined cognitive style based on participants with different behavioral patterns (n = 45; Cerliani et al., 2017; Miller, Donovan, Bennett, Aminoff, & Mayer, 2012). In Experiment 2: (1) 42 healthy older adults (age 66–85 years, 29 women) who had no history of neurological or psychiatric disease records and (2) 16 Alzheimer disease (AD) patients (age 80–90 years, 12 women) who were diagnosed by a physician according to the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition performed the scene-based episodic memory task (data were adapted from the previous study J. H. Park & Lee, 2021). The results of neurophysiological assessments (Mini-Mental State Exam, Seoul Verbal Learning Test, and Korean version of Boston Naming Test) showed significantly lower scores for AD patients compared with the healthy older adults (see Appendix). The average years of education were significantly higher for the healthy aging group (mean = 8.286) compared with the AD group (mean = 5.313). No EEG data were collected.

All participants were right-handed and had either normal vision or corrected-to-normal vision. Informed consent was gained from all participants. All experimental procedures received the approval of the institutional review board of Korea Advanced Institute of Science and Technology and Seoul National University and were conducted in accordance with ethical guidelines for research on human participants.

A scene-based episodic memory task tested three different conditions (i.e., what, where, and when) in a blocked design. There were four sets of 10 scenes: one consisted of real-life indoor photographs, and the other three were virtual scenes captured in Unreal Engine (Epic Games) with different themes (apartment, outdoor, and museum). There was no difference in accuracy between the real-life indoor set and the other sets (78.4% vs. 80.6%, t(46) = 1.503 and p = .14). Counterbalanced combinations of condition and scene sets were randomly assigned for each block (e.g., Encoding-set2 / where retrieval → Encoding-set4 / what-retrieval → Encoding-set1 / when retrieval → …). During the encoding phase, 10 scenes were shown on the screen for 3 sec each (Figure 1A top-left). Participants were informed of the test condition (what, where, or when) at the start of the retrieval phase. Each retrieval phase followed the encoding phase and consisted of 10 forced-choice questions. In the what condition, participants were shown a scene with an object missing and asked to choose between two objects: the correct object that was in the scene during encoding and a lure object (Figure 1A top-right). In the where condition, participants were asked to choose between the correct scene shown during the encoding from a lure scene with an object displaced from its original location (Figure 1A bottom-left). In the when condition, participants were asked to choose from two scenes the one that was seen first during encoding (Figure 1A bottom-right). Participants were given 5 sec to respond at will, and trials in which they failed to respond within 5 sec were marked as incorrect. After three test blocks were completed, participants were asked to fixate on a cross at the center of the screen for 1 min. The EEG signal from this period was used as a baseline for the following three blocks. Throughout the whole experiment, each participant performed 120 trials, 40 for each condition, to enhance the reliability of the EEG neural markers across three conditions.

In considering potential issues of fatigue and/or difficulty with virtual scenes, the version of the scene-based episodic memory task for older adults consisted of only one set of stimuli consisting of 10 real-life indoor scenes was used, and the order of three conditions was randomized for each participant to avoid the interference effects induced by lures shown during the retrieval. The aging and AD participants were given unlimited time to respond during the retrieval period. The unlimited time was later justified by the results that the average RT for the healthy aging participants and AD patients was 8.81 and 10.82 sec whereas RT was longer than 5 sec in only 140 out of 5640 total trials (2.48%) from 47 young participants. As a control analysis, when only the real-life indoor set was considered, the young group significantly outperformed the aging and AD groups in every condition (what: t(103) = 3.399, where: t(103) = 4.353, when: t(102) = 11.037, and all p < 10⁻³).

We recorded EEG signals from 32 passive electrodes mounted on an elastic cap (Neuroscan Grael, FP1, FP2, F11, F7, F3, FZ, F4, F8, F12, FT11, FC3, FCZ, FC4, FT12, T7, C3, CZ, C4, T8, CP3, CPZ, CP4, P7, P3, PZ, P4, P8, O1, OZ, O2, M1, and M2; Figure 1B); signals from four channels (F11, F12, FT11, and FT12) were discarded because of artifacts. EEG signals were originally collected with 1024-Hz sampling rate while keeping impedance below 10 kΩ, and then band-pass filtered between 0.1 Hz and 100 Hz. All sites were initially referenced by a reference electrode (positioned between Cz and CPz) and rereferenced later to the algebraic average of the left and right mastoids (M1 and M2). Eyeblink signals were detected by VEOG, which was recorded from a bipolar pair of electrodes placed above and below the left eye. VEOG signals above the threshold (50 ∼100 μV) after band-pass filtering between 0.1 and 50 Hz was labeled as eyeblinks. Eyeblink artifacts were detected and removed by using the PCA-based algorithm provided by Curry 7 (Neuroscan). Finally, we applied an independent component analysis-based algorithm provided by the EEGLAB toolbox to remove remaining artifacts such as horizontal eye movement, excessive muscle noise, changes in skin potentials, and so forth (Pion-Tonachini, Kreutz-Delgado, & Makeig, 2019).

We calculated band power for frequencies between 2.5 and 100 Hz during the encoding and the retrieval period (by applying a Morlet wavelet transformation), and calculated z scores with respect to the preceding fixation period. To explore the temporal dynamics of theta oscillations (2.5–8.5 Hz) as shown in a previous study (S.-E. Park et al., 2023), we divided the period from the onset to the response into 80 equal time bins. Theta power from the early and late retrieval was calculated by averaging the first and second half of the time bins (40 bins each), respectively, and used for further analyses. Alpha/beta (8.5–30 Hz) and gamma power (30–100 Hz) were calculated in the same way. Pearson's correlation was calculated between accuracy of what, where, and when retrieval and theta power in every channel or the frontal region (FP1, FP2, F7, F3, FZ, F4, F8, FC3, FCZ, and FC4) after averaging (Figure 2E–J). Outliers at the participant level were not included. To directly compare the strength of the correlations across the three components, we applied a percentile bootstrap method that measures how confidently two correlation coefficients are different (Wilcox, 2009).

Given that semantic dissimilarity contributes to mnemonic discrimination (Naspi et al., 2021; Cann, Mcrae, & Katz, 2011), semantic distance (SD) was calculated by the Wordnet-based Wu & Palmer similarity algorithm (Wu & Palmer, 1994). A semantic similarity value was calculated for each pair of the target and lure objects in the what retrieval, and SD was assigned by ranking the similarity across the 10 trials within a block. For where retrieval, Euclidean spatial distance (ED) indicated the distance of an object in the scene between the original (target) and the moved (lure) position. EDs were also assigned by ranking across the 10 trials within a block. For the when condition, the temporal gaps between episodic events in a sequence affects the difficulty of temporal order judgments (St. Jacques et al., 2008); we defined TD as the difference in the order of presentation during encoding; because 10 scenes were shown during encoding, TD ranged from one to nine. The selection of scene pairs and TDs were pseudorandomly decided, such that TDs were evenly distributed from one to nine across each of the four blocks of when retrieval trials. The repeated presentation of scenes was also varied and counterbalanced across individuals, such that, across the entire group, all scenes were presented equally.

We calculated Kendall's tau correlations between the cognitive distances (SD, ED, and TD) and theta power for what, where, and when conditions during the early and late retrieval periods (Figure 3). Statistical significance was tested by performing one-sample t tests against zero on both channel-averaged and channel-wise Kendall's tau correlation coefficients from every participant (n = 47). All t tests in this study are two-tailed. Outliers at the participant level were not included. The effect size was measured by Cohen's d.

The K-means clustering (Lloyd, 1982) method was used for dividing participants into three groups (cognotypes) based on the task performance. Two-dimensional behavioral features were considered: when accuracy and a mean of what and where accuracy (based on the high correlation found between the two; Figure 4B). Each participant was assigned to one of three cognotypes based on the squared Euclidean distance to the centroid of the group. The hyperparameters of the clustering (squared Euclidean distance, three clusters) were selected to maximally separate the clusters based on the post hoc silhouette analysis (Kaufman & Rousseeuw, 1990).

Cognotypes were compared in both their what, where, or when accuracy and corresponding neural markers (Figure 5). A one-way ANOVA was applied to theta power across the cognotypes in every EEG channel during the early and late retrieval. On the basis of the ANOVA results showing across-cognotype differences in the frontal channels, we further applied two-sample t tests to averaged theta power from those channels in the early and late retrieval periods to compare across pairs of cognotypes.

Cognitive distance (SD, ED, and TD) coding represented by the theta-distance correlation (Kendall's tau) was also compared across the three cognotypes by a one-way ANOVA test in every channel. Because differences were pronounced in the posterior regions, two-sample t tests were used for pairwise comparisons in distance coding in posterior channels. Outliers at the participant level were not included. The effect size was measured by Cohen's d.

The same K-means clustering algorithm was applied to behavioral data from healthy older adults (Figure 7A). Behavioral similarity between each cognotype among aging participants and the AD group was measured by calculating the Euclidean distance between all participant pairs on the 2-D plane. Lastly, we tested whether any of the cognotypes showed particularly lower accuracy in when memory—a characteristic of AD patients shown in previous studies (J. H. Park & Lee, 2021; El Haj & Antoine, 2018; Pirogovsky et al., 2013; Bellassen et al., 2012). To account for intrinsic difficulty differences across what, where, and when trials, accuracy was normalized by the average accuracy of the corresponding task. The performance of AD patients was normalized by the mean accuracy of the healthy aging participants.

EEG data from healthy, young participants (Experiment 1) were analyzed separately for what, where, and when memory components (Figure 1). Compared with the fixation period, theta power (2.5–8.5 Hz) was increased during the retrieval phase of the task (Figure 2A and 2B; one-sample t tests against zero for retrieval of what: t(46) = 5.58, p < 10⁻³ false discovery rate [FDR] corrected, and Cohen's d = 0.8, where: t(46) = 6.87, p < 10⁻³ FDR corrected, and d = 0.99, when: t(46) = 7.55, p < 10⁻³ FDR corrected, and d = 1.08). In contrast, gamma power (30–100 Hz) changed marginally (Figure 2D), and alpha/beta power (8.5–30 Hz) decreased during retrieval (Figure 2C). Theta power showed different patterns between the early and late retrieval periods, increasing to its peak in the early period and gradually decreasing in the late period (Figure 2E). Differences in theta power across the three components also showed an early-late distinction, with two separate peaks in difference across conditions in the early and late periods (Figure 2E).

A correlation between theta power and accuracy of what, where, and when retrieval was computed for every channel. Frontal theta power, which was reported as a neural marker of episodic memory performance in previous studies (Nyhus & Curran, 2010; Summerfield & Mangels, 2005; Sederberg, Kahana, Howard, Donner, & Madsen, 2003), was correlated with performance in the what and where conditions in the late retrieval period (Figure 2F–I, Pearson correlation what: r = .34, p = .021; where: r = .3, p = .044); this correlation was not found in the when condition in any channel (Figure 2J and K; r = .05). In a comparison of correlation coefficients between frontal theta power and retrieval accuracy for each condition (i.e., what, where, or when), we observed a nonsignificant trend for a stronger correlation between frontal theta power and what and where accuracies compared with when accuracy (what vs. when: 95% confidence interval [CI] of difference = [−0.04, 0.59] and p = .082; where vs. when: 95% CI of difference = [−0.16, 0.60] and p = .229). These results suggest that what and where share a common neural correlate involving frontal theta power, whereas when memory performance may be driven by a different neurocognitive process.

As reported in previous studies (Naspi et al., 2021; Cann et al., 2011; St. Jacques et al., 2008), an increase in cognitive distance was associated with higher episodic memory accuracy for all three distances. SD and TD were positively correlated with accuracy in most of the participants (Figure 3A and E; one-sample t tests on Kendall's tau of 47 participants against zero SD: t(46) = 7.11, p < 10⁻³, and d = 1.02; TD: t(46) = 18.86, p < 10⁻³, and d = 2.71). ED showed a positive correlation with accuracy in trials with the target and lure objects placed on the same side of the screen (Figure 3C; t(46) = 3.67, p < 10⁻³, and d = 0.53). When the target and lure objects were on opposite sides, however, accuracy was generally lower, despite the higher ED between them (see S.-E. Park et al., 2023), which may be related to previous reports of representational invariance across left-right symmetry in scene-processing regions of the brain (Dilks, Julian, Kubilius, Spelke, & Kanwisher, 2011).

Theta power was distinctly modulated across the three cognitive distances during episodic memory retrieval (Figure 3B, D, and F), with a significant positive correlation with both SD and ED during the late retrieval period (Figure 3B and D; two-tailed t tests on Kendall's tau correlation coefficient of 47 participants against zero SD: t(46) = 3.24, p = .006 FDR corrected, and d = 0.47; ED: t(46) = 2.58, p = .022 FDR corrected, and d = 0.37). This correlation was not observed in the early retrieval period (Figure 3B and D; SD: t(46) = 0.702, p = .486, and d = 0.1; ED: t(46) = −0.311, p = .758, and d = −0.04). For when memory, theta power showed two distinct patterns in relation to TD. The first was a negative correlation during the early retrieval period (Figure 3F; two-tailed t tests against zero t(46) = −3.24, p = .007 FDR corrected, and d = −0.46), and the second pattern was a positive relationship during the late retrieval period (Figure 3F; t(46) = 2.54, p = .022 FDR corrected, and d = 0.36). These results suggest that cognitive distance-modulated cortical theta power is a context-dependent neural marker for guiding episodic memory retrieval.

Individual differences in performance were observed for all episodic memory components (Figure 4A; accuracy ranging from 0.55 to 0.975). We found that individual performance in the what and where conditions was positively correlated (Figure 4B; Pearson correlation r = .45 and p = .005 FDR corrected), indicating that good performers in the what condition were also good at the where condition; this is consistent with the finding that there is a shared neural marker for performance in the two conditions (i.e., frontal theta power correlated with performance in the what and where conditions but not the when condition). In contrast, when retrieval performance was independent from what and where performance (Figure 4B; Pearson correlation between what and when: r = .23 and p = .128; where and when: r = .15 and p = .304). In a direct comparison of correlation coefficients across each pair of three conditions by applying a percentile bootstrapping method (Wilcox, 2009), we found a significant difference in the strength of the correlations between what–where versus where–when (95% CI of difference = [0.03, 0.57] and p = .016) and a nonsignificant trend in comparing what–where versus what–when (95% CI of difference = [−0.14, 0.55] and p = .116). These results were summarized in a performance similarity map where the distances across the three components indicate the inverse of the correlation coefficient (Figure 4C), with the when component distant from the other two components.

Cognotyping was performed using unsupervised K-means clustering applied to the 2-D space consisting of the average of what and where accuracy on one axis (given their high correlation) and when accuracy on the other axis (Figure 4D and E). Each cognotype showed a distinct profile of episodic memory performance (Figure 4F and G); they are hereafter referred to as Type G (good overall performance, n = 17), Type TG (only good at temporal order memory, n = 14), and Type B (overall bad performance, n = 16). The overall accuracy was highest in Type G followed by Type TG and B (Figure 4F; two-sample t tests Type G vs. TG: t(29) = 4.9, p < 10⁻³, and d = 1.72; Type TG vs. B: t(28) = 4.01, p < 10⁻³, and d = 1.43). Type TG was significantly better at when memory compared with the other components (Figure 4F and 4G; paired t tests when vs. what: t(13) = 6.36, p < 10⁻³, and d = 2.54; when vs. where: t(13) = 14.6, p < 10⁻³, and d = 3.43), contributing to the significantly large variance across what, where, and when accuracies in Type TG compared with the other cognotypes (Figure 4F; two-sample t tests Type TG vs. G: t(29) = 6.75, p < 10⁻³, and d = 2.37; Type TG vs. B: t(28) = 5.13, p < 10⁻³, and d = 1.82). Cognotyping in 3-D space (what, where, and, when accuracy on each axis) produced the same three clusters of participants as above.

Type G outperformed Type TG and B in the what and where conditions (Figure 5A and 5F; two-sample t tests Type G vs. TG what: t(29) = 4.39, p < 10⁻³, and d = 1.54; where: t(29) = 5.74, p < 10⁻³, and d = 2.02; Type G vs. B what: t(31) = 7.56, p < 10⁻³, and d = 2.57; where: t(31) = 4, p ≤ 10⁻³, and d = 1.36), whereas no differences were found between Type TG and B (two-sample t tests what: t(28) = 1.63, p = .114, and d = 0.58; where: t(28) = −0.99, p = .33, and d = −0.35). To test whether these behaviors can be explained by their underlying neural correlates, we first compared theta power across three cognotypes. Frontal theta power was different across cognotypes for what/where retrieval according to a one-way ANOVA (Figure 5C and H), and Type G showed higher frontal theta power than the other low-performing cognotypes during the late retrieval period (Figure 5B and G, two-sample t tests Type G vs. TG what: t(29) = 2.06, p = .048, and d = 0.73; where: t(29) = 1.76, p = .089, and d = 0.62. Type G vs. B: what: t(31) = 2.49, p = .018, and d = 0.85; where: t(31) = 2.5, p = .018, and d = 0.85) whereas no across-cognotype differences were observed in the early retrieval period (see Appendix).

Type TG outperformed the other two cognotypes in the when condition (Figure 5K; Type TG vs. Type G: t(29) = 3.16, p = .004, and d = 1.11; Type TG vs. Type B: t(28) = 9.32, p < 10⁻³, and d = 2.49). However, Type TG's theta power was lower than Type G (Figure 5L; Type TG vs. G: t(29) = −1.71, p = .097, and d = −0.6) and similar to Type B (Type TG vs. Type B: t(28) = 0.51, p = .617, and d = 0.18), suggesting that frontal theta power was not a useful indicator of when memory. Therefore, we tested for an alternative neural marker of when memory performance.

We investigated how the modulation of theta power by cognitive distance (SD, ED, and TD) affected episodic memory performance by comparing the correlation coefficients between accuracy and each cognitive distance across the three cognotypes. One-way ANOVA results showed that the across-cognotype difference, although marginal in the SD-theta and ED-theta correlation (Figure 5D–E and I–J), was significant in when memory especially in the posterior channels (Figure 5O). Two-sample t tests showed that cognotypes with higher when memory performance (Type G and TG) showed a significantly higher TD-theta correlation during the late retrieval period compared with the low-performing Type B participants (Figure 5N; two-sample t tests Type G vs B: t(30) = 2.66, p = .012, and d = 0.92, Type TG vs. B: t(25) = 4.26, p < 10⁻³, and d = 1.6). When TD-theta correlations were tested against zero for each channel, we found that Type B did not show a significant correlation in any of the channels (Figure 6C) whereas Type G and TG showed a positive TD-theta correlation mostly in the posterior region during late retrieval (Figure 6A and B). These results imply that individual performance in when memory depended on how precisely the theta power was modulated for processing temporal order information. In the early retrieval period, we found no theta-mediated distance coding for SD and ED (Figure 3B and 3D) and no across-cognotype differences (one-way ANOVA of TD-theta power correlation in the posterior region across three types: F(2, 45) = 0.326 and p = .724; see Appendix).

Overall, the memory differences across cognotypes were mapped onto specific neural markers for episodic memory components. First, higher theta power in the frontal region during the late retrieval period was correlated with in better performance in what and where memory. Second, the modulation of theta power by TD theta power was associated with better performance in when memory.

Thus far, we have shown that cognotyping provided a more intuitive comparison of individual performance and neural activity patterns in what, where, and when memory in younger adults in their 20s. The same K-means clustering method was applied to assign cognotypes to older participants between 65 and 85 years of age. Remarkably, we found that they clustered into cognotypes with the same characteristics as the younger cognotypes (Figure 7A): Type G – good overall performance (mean age = 73.1), Type B – bad overall performance (mean age = 77.6), and Type TG – good temporal order memory only (mean age = 78). This similarity in pattern across the data from two age groups, although cross-sectional in nature, suggests that cognotypes based on what, where, and when performance may be preserved in healthy aging. We applied the same analytic method to the participants over 80 years old to match with the AD group and observed the same cognotyping pattern (Figure 7C). The number of years of education of the age-matched group was marginally higher than the AD group (two-sample t tests, t(29) = 1.95 and p = .062; AD median = 6) but not different across three cognotypes (One-way ANOVA: F(2, 12) = 0.23 and p = .8; Type G median = 9, Type TG = 12, and Type B = 9).

Interestingly, the behavioral pattern found in AD patients performing the same task mostly overlapped with Type B participants from the healthy aging group (Figure 7A and 7B; two-sample t tests comparing Type B vs. Type G's Euclidean distance from the AD cluster: t(430) = 19.73 and p < 10⁻³; Type B vs. TG: t(446) = 9.89 and p < 10⁻³). When only participants over 80 years of age were considered, Type B still showed the most similar behavioral pattern to the AD patients (Figure 7D; two-sample t tests for distance from AD to Type B vs. G: t(110) = 5.389 and p < 10⁻³; Type B vs. TG: t(190) = 8.676 and p < 10⁻³). Furthermore, comparing the accuracy normalized by the average of the healthy participants (see Methods section for details) showed that Type B from both young and aging groups was especially worse in when memory compared with what and where memory, similar to the pattern found in AD participants (Figure 7E–H; paired t tests for normalized when vs. other accuracy in young participants: t(15) = 2.57, p = .028 FDR corrected, and d = 0.85; aging participants: t(12) = 8.75, p < 10⁻³ FDR corrected, and d = 2.19; AD patients: t(15) = 4.37, p = .001 FDR corrected, and d = 1.09; age over 80: t(3) = 3.872, p = .031 FDR corrected, and d = 2.67). This common behavioral pattern revealing a fragility of when memory may implicate a neurocognitive profile shared across Type B and AD patients.

Understanding individual differences in a complex cognitive ability like memory is sometimes not straightforward as it may be subserved by multiple neural mechanisms that vary across participants and conditions. Furthermore, although a variety of neural activities may be induced during memory retrieval, only some may be critical to explaining individual performance. In this study, we explored the cognotyping approach as a way to provide a better way of mapping individual differences between brain and behavior in as much detail as possible without losing generalizability.

Three cognotypes with distinctive behavioral profiles across what, where, and when memory were identified, and the behavioral dissociation of each cognotype was explained by corresponding neural markers (i.e., neuro-cognitive mapping). The high performance in what and where memory observed for Type G was explained by how effectively frontal theta power was induced, consistent with previous studies supporting the importance of frontal theta in episodic memory (Eschmann, Bader, & Mecklinger, 2020; Staudigl & Hanslmayr, 2013) including its role in mediating long-range communication between brain regions (Herweg et al., 2020; Hasselmo & Stern, 2014; Nyhus & Curran, 2010; von Stein & Sarnthein, 2000). We observed a unique neural marker of when memory, which was dissociable from what and where components. Although the retrieval of when information generally induced theta power, higher individual performance in when memory (i.e., in Types TG and G) was mapped uniquely onto the modulation of theta across TDs (i.e., TD-theta correlation). These results demonstrate that differences in the underlying neural markers of episodic memory components are important for explaining individual memory performance.

We found that the three episodic memory cognotypes were preserved in healthy aging between younger and aging groups. In addition, one specific cognotype (Type B) showed a behavioral pattern similar to AD patients who have been reported to be particularly impaired in when memory (J. H. Park & Lee, 2021; El Haj & Antoine, 2018; Pirogovsky et al., 2013; Bellassen et al., 2012). What could explain the differences in neurocognitive mapping of when compared with what and where memory and its vulnerability in AD? When memory requires heavy contextual processing for “replaying” temporal sequences (and perhaps relatively less perceptual processing) compared with the what and where conditions. Although both spatial (Burgess, Maguire, Spiers, & O'Keefe, 2001) and temporal memory (Hayes et al., 2004) requires contextual processing, spatial memory can be additionally supported by various perceptual mechanisms, such as the recognition of the spatial structure (e.g., closed vs. open space), or semantic properties of the scene (e.g., kitchen vs. bathroom). In contrast, temporal contexts—particularly in the coding of temporal sequence—are defined relative to one another (e.g., Tuesday occurs before Wednesday but after Monday) and, therefore, can be said to be more heavily dependent on contextual processing compared with both the what and where conditions. Temporal-distance modulation of theta oscillations may be an electrophysiological marker for cortico-hippocampal communication required to replay temporal sequences in memory. The fact that, among the three cognotypes, the behavioral profile of Type B was similar to AD patients may indicate an under-utilization of cortico-hippocampal systems that may come to be pathological in AD patients.

The present study, although suggestive of a link between memory profiles and AD, is limited in the following ways. First, there is a possibility that, although the behavioral cognotypes were found in both younger and older participants, the mapping between brain and behavior identified in younger participants may not apply to older participants and AD. The purpose of this study was to validate the sensitivity of our task in relating the cognotypes to neural markers by finding the representation of the corresponding “intact” neural correlates for what, where, and when in each cognotype of healthy young adults. Indeed, our goal for the next study is to investigate the altered neural correlates in aging and disease to bring us closer to a mechanistic understanding of the vulnerability of episodic memory (particularly for the when component). Furthermore, because ours was a cross-sectional study, it was not possible to track whether cognotypes were stable across time. Longitudinal studies can test the prevalence of between-cognotypes conversions (e.g., Type TG to Type B) or the progression of age-related memory impairments across cognotypes.

Another potential limitation of this study is that the task design for the younger adults' EEG experiment was slightly different from the aging/AD experiment. For instance, the aging/AD experiment only involved real life photos, had fewer trials, and gave participants unlimited time to respond—all to make the task feasible for the elderly and patient population. Although we do not believe that these differences played a role in determining the results, particularly given the similar cognotypes that emerged in the two data sets, future studies may benefit from implementing identical tasks across young and older participants.

Early diagnosis or detection of high risk for AD is important for early intervention and the maintenance of a high quality of life in old age (Crous-Bou, Minguillón, Gramunt, & Molinuevo, 2017). Combined with cognitive training targeting specific performance-related neural markers, it may open new doors for precision therapeutics for prevention and treatment.

The cognotyping approach can be applied to various cognitive neuroscience studies and has potential for finding new neurocognitive markers that explain individual variability. We demonstrated this in the case of episodic memory, characterized by its what, where, and when components. Although further experiments are needed to investigate the alterations in these neural markers in aging and AD, we believe that the present findings provide both empirical and theoretical insights that may guide our understanding of individual differences in age-related cognitive decline.

This study was supported by the Creative-Pioneering Researchers Program at Seoul National University and the National Research Foundation of Korea (2021M3E5D2A01023891, 2021M3A9E4080780 and 2020K1A3A1A19088932).

Reprint requests should be sent to Sang Ah Lee, Department of Brain and Cognitive Sciences, Seoul National University, Seoul, Republic of Korea, or via e-mail: [email protected].

Requests for the data can be sent to the corresponding author.

Sang-Eon Park: Conceptualization; Formal analysis; Methodology; Writing—Original draft. Jeonghyun Lee: Formal analysis; Methodology; Writing—Original draft. Jin-Hyuck Park: Conceptualization; Validation; Writing—Review & editing. Maria Jieun Hwang: Conceptualization; Validation; Writing—Review & editing. Sang Ah Lee: Conceptualization; Funding acquisition; Supervision; Writing—Original draft.

Sang Ah Lee, National Research Foundation of Korea (https://dx.doi.org/10.13039/501100003725), grant number: 2020K1A3A1A19088932, grant number: 2021M3A9E4080780, grant number: 2021M3E5D2A01023891.

Retrospective analysis of the citations in every article published in this journal from 2010 to 2021 reveals a persistent pattern of gender imbalance: Although the proportions of authorship teams (categorized by estimated gender identification of first author/last author) publishing in the Journal of Cognitive Neuroscience (JoCN) during this period were M(an)/M = .407, W(oman)/M = .32, M/W = .115, and W/W = .159, the comparable proportions for the articles that these authorship teams cited were M/M = .549, W/M = .257, M/W = .109, and W/W = .085 (Postle and Fulvio, JoCN, 34:1, pp. 1–3). Consequently, JoCN encourages all authors to consider gender balance explicitly when selecting which articles to cite and gives them the opportunity to report their article's gender citation balance. The authors of this article report its proportions of citations by gender category to be as follows: M/M = .531; W/M = .25; M/W = .188; W/W = .031.