Encoding-related Brain Activity Predicts Subsequent Trial-level Control of Proactive Interference in Working Memory

Proactive interference (PI) occurs when previously encountered information disrupts the retention of new goal-relevant information. In working memory (WM), PI is commonly investigated using n-back or recent-probes tasks with embedded lure trials that are familiar but occurs in a task-irrelevant position or target set. One proposed way of handling PI in memory is to associate an item to its unique spatial and temporal context (Bramão, Jiang, Wagner, & Johansson, 2022; Beukers, Buschman, Cohen, & Norman, 2021). Such associations provide strong retrieval cues that help differentiate between similar memories; contextual information is therefore critical for placing an item in its appropriate temporal context. For example, in the commonly used 2-back (2B) task, context information is critical for differentiating between a familiar 2B target trial and a familiar, but task-irrelevant, 3-back lure trial.

Previous neuroimaging and patient studies have helped identify brain regions that are critical for PI control in WM. Such key regions include the left inferior frontal gyrus (IFG; Samrani & Persson, 2022; Samrani, Bäckman, & Persson, 2019; Nelson, Reuter-Lorenz, Persson, Sylvester, & Jonides, 2009; Nee, Wager, & Jonides, 2007), the striatum (Schmidt et al., 2020; Schmidt et al., 2018; Ziaei, Peira, & Persson, 2014), and more recently the hippocampus (HC; Samrani & Persson, 2022; Öztekin, Curtis, & McElree, 2009). Although these regions are included in a canonical network underlying interference control, the exact contribution of these regions in resolving PI is currently not known. One proposed role for these regions is in providing temporal context information or a “temporal tag” that supports an individual to decide whether a familiar item is presented in the correct (i.e., task-relevant) or incorrect temporal position. Indeed, there is a great overlap between regions activated by high interference conditions/trials and those involved in temporal context memory (van de Ven et al., 2020; Clewett, DuBrow, & Davachi, 2019; Eichenbaum, 2013). One possibility is that these regions play an important role in interference control by supporting temporal context information that mitigates PI by distinguishing between target information and familiar, but task-irrelevant lures that are presented in an incorrect temporal context.

Given that contextual information is established during encoding, it is likely the efficiency by which item–context associations are formed also predicts the extent of PI at retrieval. If a robust item–context association is formed during encoding, less effort is needed for handling PI at retrieval. In line with context reinstatement theories (Yonelinas, Ranganath, Ekstrom, & Wiltgen, 2019; Howard, 2017; Polyn, Norman, & Kahana, 2009; Howard & Kahana, 2002), one proposed view is that contextual information can be used for reducing the effects of interference in WM (Beukers et al., 2021). Thus, and similar to studies using the subsequent memory paradigm in episodic memory, items could be selected based on the PI related to that specific trial and then use this information to investigate encoding-related activation associated with high and low interference trials.

Previously used WM tasks are typically characterized by a very limited stimulus set that is recycled over multiple trials. In such tasks, highly similar or identical items accumulate through the buildup of traces and critically reduces the ability to distinguish a current target trial from their presence in previous trials. Such tasks are specifically designed to minimize the involvement of long-term memory by inducing high levels of PI. More recently, it has been proposed that performing a WM task may rely more heavily on long-term memory than was previously thought. For example, it has recently been proposed that WM maintenance in the absence of concurrent brain activation (i.e., activity silent WM) could be explained by a transfer of information from WM to long-term memory (Beukers et al., 2021). By this view, control of interference in WM can be obtained by encoding of contextual information that persists over many trials and that can be subsequently retrieved along with item information when needed. Thus, at encoding, contextual information that are active in WM is integrated with the item representation to establish a longer-lasting episodic memory trace. The neural representation of binding item and contextual representations likely involve brain regions critical for associative learning, such as the pFC and the HC.

We have recently demonstrated that including unique stimuli in the context of an in-scanner fMRI n-back WM task generates long-lasting episodic memory traces in a subsequent offline recognition paradigm (Samrani & Persson, 2022). In the current study, participants were scanned with event-related fMRI while performing a 2B WM task with embedded 3-back lures designed to induce PI. To examine trial-specific binding, and in contrast to most studies using the n-back task, unique verbal stimuli (words) were used. The n-back task places high demand on the ability to encode, retain, and retrieve temporal information to correctly dissociate a correct 2B target response from a familiar but incorrect 3-back lure trial (Figure 1). Failed encoding of temporal context information would lead to trial ambiguity from the inability to place the familiar item in its correct temporal position. Here, we test the hypothesis that encoding of temporal context information, as indicated by brain activation in regions previously associated with these processes such as the IFG, HC, and striatum, predict subsequent level of PI at retrieval.

Twenty-four healthy, right-handed participants between 18 and 35 years (M = 28.5 years, SD = 4.0 years; 12 women) took part in the study. Neurologic and psychiatric health status was assessed by a self-report questionnaire that included questions of whether participants were currently being treated for psychiatric or neurological conditions like depression, epilepsy, diabetes, trauma, and Parkinson disease. Individuals with neurological or psychiatric conditions, or with implants and other surgery unfit for MRI scanning were excluded. One participant was excluded from the analysis because of excessive movement and technical problems during scanning, leaving 23 participants for all analyzes. Our sample size was based on a previous study with the same sample size (Samrani & Persson, 2022), in which we found a robust effect of both behavioral and brain activation measures on the manipulation of PI. Participants received financial compensation of 600 Swedish krona for their participation. The study was approved by the Regional Ethical Review Board in Stockholm, and written consent was obtained from all participants.

PI was measured using a verbal 2B WM task, which included familiar lure items (Marklund & Persson, 2012; Gray, Chabris, & Braver, 2003) occurring either at 3, 5, 6, 7, 8, 9, or 10 trial(s) after first item presentation (i.e., 3- to 10-back lures; Samrani, Bäckman, & Persson, 2017). These are referred to as 3B, 5B, and so forth, up to 10B. The task was divided into two equal blocks of 105 trials each, with a 1-min break between blocks. The proportion and number of trials consisted of (1) nonfamiliar words presented for the first time (new trials), (2) words presented for the second time at the correct 2B position (target trials), (3) words presented for the second time at an incorrect position, one word after the target position (3B; proximal lures), and (4) words presented a third time, 3–10 trials from the target position (3B to 10B; distant lures). Two hundred ten trials were included in total, of which 78 (37.1%) were new trials, 44 (21%) were target trials, 40 (19%) were proximal 3B lure trials, and 48 (22.8%) were distant lure trials.

Lure trials consisted of stimuli already presented 3–10 trials earlier, and required a “no” response, and new trials were nonfamiliar trials that had never been presented previously, which also required a “no” response. Target trials were 2B trials and required a “yes” response. For each presented word, participants were instructed to press with the right index finger on a MRI compatible response box, which corresponds to “yes” (“Yes, the word I now see has been shown two words ago”) and the button on the middle finger for “no” (“No, the word I now see has not been shown two words ago”). Distant lures were all recycled from either previous target trials or proximal lures with the aim to lower the total amount of new trials, consequently the proportion of “no” answers. In the current study, only 3B lures were used in all analyses.

Stimuli and trial conditions were presented in the same fixed order for all participants. Stimuli consisted of common English nouns with a maximum of two syllables and were presented one at a time for 2.5 sec, with an intertrial jittering of either 2, 2.66, 3.33, or 3.99 sec. The words included in the n-back task were not selected, or controlled for, semantic of phonological similarity. There was an even distribution of the jitter-timings, and the timing from a new trial to the target position always added up to a total intertrial time of 5.99 sec to avoid any differences in encoding time between target items. Participants were instructed to answer as quickly and accurately as possible and were also instructed not to overtly repeat the words in the scanner by speaking them out loudly or moving the mouth. Relative difference scores were calculated as the relative proportional difference in RTs between nonfamiliar trials and familiar lure trials (Samrani et al., 2017, 2019). Interference was estimated as the difference in percent between lure trials (high interference trials) and new (non-interference trials). A relative difference score represents a more salient measure of executive control, as it considers baseline individual differences in RT. Median RTs were used to reduce the influence of extreme values.

To ensure that changes in PI was not related to forgetting, each participant also performed a recognition task immediately after the scanning session. In the recognition memory task, 105 one- or two-syllable common English nouns were presented one at a time for 5 sec. Participants were instructed to indicate if the word had been presented previously as part of the in-scanner WM task. Thirty-five of these words (1/3 of total) had not been presented inside the scanner (i.e., lures), and 70 words (2/3 of total) had previously been presented in the fMRI 2B task. Positive responses to previously presented items were scored as hits and positive responses to lure trials were scored as false alarms. Hit rate minus false alarm rate was used as a measure of recognition memory performance.

RTs were calculated for correct trials only. Median RTs were extracted for target trials, new trials, and lure trials for each condition. Words presented for the first time (new trials) and later presented as a 3-back lure were of special interest in this study. Lure trial RTs were divided with new trials to get a relative difference score lure trial $lure trialnewtrials−1×100$ for each trial. A higher relative difference score indicates more interference. Thus, single lure trials were compared against the median RT of new trials for each participant, creating a relative score for each trial to be used in the general linear model. This calculation included two steps. First, we determined a fixed median value to reflect each participant's average RT for new trials. Second, we calculated the relative difference between that median RT value and each lure trial RT, thus obtaining an estimate of PI for each lure trial.

Participants were scanned with an eight-channel phased array receiving head coil (Discovery MR750 3.0 T scanner, General Electric). T1-weighted 3-D spoiled gradient recall images were obtained with the following MRI scanner parameters: repetition time = 8.2 msec, echo time = 3.2 msec, field of view = 25 cm, 176 axial slices, flip angle = 12°. Task-related fMRI data were acquired using a gradient-EPI sequence with the following MRI scanner parameters: repetition time = 2000 msec, echo time = 30 msec, flip angle = 70°, field of view = 25 cm. Forty-two transaxial slices with a thickness of 3 mm (0.35-mm gap) were acquired. Ten initial dummy scans were collected to allow for the fMRI signal to reach equilibrium. The stimuli were presented on a computer screen seen through a tilted mirror. E-Prime software (Psychology Software Tools, Inc.; Version 2.0) was used for stimulus presentation and recordings. Headphones and earplugs were used to dampen scanner noise, and cushions inside the head-coil helped to minimize head movements.

All fMRI data were preprocessed using the statistical parametric mapping software (SPM12; Wellcome Department of Cognitive Neurology) implemented in MATLAB 9.3 (MathWorks Inc.). Before analysis, the data were preprocessed in the following way: slice timing correction, movement correction by unwarping and realignment to the first image of each volume, co-registration, normalization to a sample-specific template using DARTEL (Ashburner, 2007), and affine alignment to Montreal Neurological Institute standard space and smoothing with an 8-mm FWHM Gaussian kernel. Following the co-registration step, the T1 image was segmented into gray matter and white matter. The final voxel size was 1.5 × 1.5 × 1.5 mm.

We tested the hypothesis that encoded new trials are accompanied by increasing engagement in a network of regions previously identified as relevant for positive for both interference control in WM and temporal context memory. These regions included the IFG (Samrani & Persson, 2022; Samrani et al., 2019; Nelson et al., 2009; Nee et al., 2007), the striatum (Schmidt et al., 2018, 2020; Ziaei et al., 2014), and the HC (Samrani & Persson, 2022; Öztekin et al., 2009).

We estimated one first-level model in the current study. The correct response rate for 3B trials were 95.8% (SD = 1.6%) on average across all participants. Forty of the new trials later appeared as 3B lure trials and were included in the analyses if a correct response was provided. Thus, the average number of new trials included in the model was approximately 38.3 trials. Regressors for each task condition were convolved with a canonical hemodynamic response function. Motion regressors from previous SPM preprocessing steps were added to each design matrix as regressors of no interest.

The general linear model examined to what extent encoding-related brain activation was modulated as a function of increased PI at retrieval. We hypothesized that activation during encoding of new trials in regions with a hypothesized role in associative/contextual memory during encoding would reflect the extent of PI when these trials were subsequently presented as lure trials. Thus, encoding-related activation for these items would predict the amount of interference measured at retrieval, with subsequent high interference trials being associated with less activation in HC, IFG, and striatum. The parametric modulator of interference RT was added to these new trials. A contrast image was made for each participant, weighting the parametric modulator of interest by 1, and then compared on a group level, using a one-sample t test.

At the whole-brain level, we set a FWE corrected threshold of p < .05 at the cluster level for the SPM analyses using a cluster-defining peak threshold of p < .001. Because regions with a known role in familiarity-based interference control, including the IFG, caudate nucleus, and HC, were of particular interest in the study, a more liberal threshold was used for determining the involvement of these regions in interference control. Within these ROIs, we controlled for multiple comparisons using Gaussian random field theory for small volumes (Worsley et al., 1996) as implemented in SPM. This was done at the second-level statistical analysis by using an uncorrected threshold of p < .001, which were followed-up using small-volume correction (SVC) at a FWE corrected threshold of p < .05.

Behavioral results have been reported in detail previously (Samrani & Persson, 2022). In short, participants were accurate in remembering target 2B items (M^{target accuracy} = 97.4%) and made few errors overall (M^Hits-FA = 94.6%), indicating high compliance and 2B task performance. Corroborating previous results, familiar 3B lures were associated with longer RTs and lower accuracy compared with new trials (RT: M^{new trials} = 844 msec, M^{3B lure trials} = 1054 msec, t(22) = −8.01, p < .001, d = 1.06; accuracy: M^{new trials} = 99.1%, M^{3B lure trials} = 90.6%, t(22) = −4.45, p < .001, d = 1.27). As previously reported (Samrani & Persson, 2022), the offline recognition task showed that memory performance was high (M^Hits-FA = 0.77, SD^Hits-FA = 0.14). That is, a majority of the words used in the 2B task were well recognized up to 45 min after the fMRI 2B task. Using a paired-samples t test, we found that the difference in error rate between high- (M = 11.3%, SD = 11.3%) and low-interference trials (M = 7.5%, SD = 8.0%) was nonsignificant, t(22) = −1.94, p = .069, d = 0.39. It should be noted that on a group level, RTs for new trials were negatively correlated with PI whereas RTs for 3B trials were positively correlated with PI (new trials – PI: r = −.521, p < .001; 3B trials – PI: r = .288, p = .008) indicating that encoding-related activation for new trials that were later associated with higher PI was not related to these trials having generally longer RTs.

Results from the whole-brain analysis revealed no significant associations between brain activity at encoding as a function of subsequent level of PI at retrieval. Focusing on the predefined ROIs, we found a significant negative relationship between encoding-related activation and subsequent PI at retrieval in the left IFG (x, y, z = −46, 36, −16, t/Z = 3.72/3.24, P^FWE(svc) = 0.018), left HC (x, y, z = −38, −21, −14, t/Z = 3.89/3.36, P^FWE(svc) = 0.016) and bilateral caudate nucleus (left: x, y, z = −15, 8, 16, t/Z = 3.81/3.30, P^FWE(svc) = 0.014; right: x, y, z = 9, 18, 4, t/Z = 3.87/3.34, P^FWE(svc) = 0.014). Thus, stronger activation in these regions at encoding predicted less PI at retrieval. A test of the opposite contrast, which would indicate that greater activity at encoding predicted more PI at retrieval, revealed nothing above threshold for neither the whole-brain analysis nor the ROI analysis. The results are illustrated in Figure 2. Note that only trials with correct responses were included in the analysis and the results are therefore not driven by subsequent memory effects (i.e., errors vs. correct responses). This was further confirmed by the finding that accuracy for trials with high and low interference did not differ in the offline recognition test.

The current study aimed to investigate to what extent neural engagement during WM encoding predicts subsequent level of PI at retrieval. We did this by using a trial-level parametric modulation of PI to model encoding-related activation using fMRI. We demonstrate that the ability to control interference in WM was modulated by level of activation in IFG, HC, and the caudate nucleus during encoding. These results suggest that efficient interference control relies on encoding of information critical for making decisions on whether a familiar trial is task-relevant or not. Such information may consist of contextual information that binds a representation in WM to a specific temporal context.

Current results support the view that control of PI is critically dependent on the ability to encode, retain, and utilize contextual information to make decisions on whether a familiar trial is task-relevant or not. Efficient encoding of contextual information, such as temporal or spatial positions, thus help to reduce the amount of PI during subsequent retrieval. These results are in line with recent demonstrations that reinstatement of encoding contexts in episodic memory can mitigate PI during competitive retrieval (Bramão et al., 2022). The view of contextual information as critical for resolving PI is not new (Beukers et al., 2021). Indeed, previous behavioral (Samrani et al., 2017; Szmalec, Verbruggen, Vandierendonck, & Kemps, 2011; Oberauer, 2005) and brain imaging studies (Samrani & Persson, 2022; Öztekin et al., 2009) have suggested that item–context associations are important for interference control in both WM and episodic memory. Here, we extend these observations by linking encoding-related activity in regions previously implicated in associative learning and item–context binding to subsequent levels of PI at retrieval.

Our results are also in line with recent proposals that information can be simultaneously retained in WM and episodic memory (Bartsch & Oberauer, 2023; Cowan, 2019). Episodic memory involves, by definition, binding of item information to its context and could therefore provide a means for resolving PI from familiar, but out of context, information in WM. The use of nonrecycled stimuli in the n-back task enables encoding of contextual information at the trial level, prevents the accumulation of PI over the task, and does not require removal of information from WM. The involvement of episodic memory during performance in the n-back task is further supported that participants retained high levels of recognition performance well up to 45 min after the main experiment. This also indicates that variability in trial-level PI could not be attributed to forgetting.

Collective evidence from animal (Kesner, Hunsaker, & Gilbert, 2005; Fortin, Agster, & Eichenbaum, 2002; Mitchell & Laiacona, 1998), patient (Konkel, Warren, Duff, Tranel, & Cohen, 2008; Mayes et al., 2001; Milner, Corsi, & Leonard, 1991; Milner, Petrides, & Smith, 1985), and human brain imaging (Thavabalasingam, O'Neil, Tay, Nestor, & Lee, 2019; Nielson, Smith, Sreekumar, Dennis, & Sederberg, 2015; Ezzyat & Davachi, 2014; Hsieh, Gruber, Jenkins, & Ranganath, 2014) studies implicate pFC and HC in long-term temporal context memory. Most neuroimaging studies have focused on brain activation during the retrieval phase. There are a few exceptions, however. For example, Jenkins and Ranganath (2010) demonstrated that brain activation in pFC and HC during encoding in the context of a WM task was related to successful recall of trials with high coarse temporal accuracy, that is, a small deviance between a participant's estimate of when a trial occurred and when it actually occurred. Similarly, Tubridy and Davachi (2011) found that within the medial temporal lobe, activations in bilateral HC predicted subsequent order memory. Also, in a recent WM study, Roberts, Libby, Inhoff, and Ranganath (2018) found that the posterior HC, along with the dorsolateral PFC was engaged during encoding and maintenance of visual temporal information. The current observations suggest that efficient interference control is accompanied by accessibility of the encoding context. This is well in line with a view in which the HC binds together slowly drifting contextual changes with item information to form an event (Yonelinas et al., 2019; Howard, 2017; Polyn et al., 2009). This would support differentiation between familiar target and lure items and thereby reduce memory interference.

While less is known about the involvement of the striatum in temporal context memory, this region, including the CN has been implicated in memory for temporal sequences (Hélie, Ell, & Ashby, 2015; Packard & Knowlton, 2002). Many brain imaging studies have also demonstrated that caudate nucleus is activated during WM tasks (McNab & Klingberg, 2008; O'Reilly & Frank, 2006) and associative memory more generally (Bauer, Toepper, Gebhardt, Gallhofer, & Sammer, 2015; Liljeholm & O'Doherty, 2012). There are some recent evidence for a role of the striatum in associative memory for temporal information. For example, in a recent fMRI study (van de Ven et al., 2020) participants performed a visual associative task with different time intervals between the cue and target, thereby manipulating the temporal context, and instructed to learn these cue-target associative pairs. Results showed increased functional connectivity between the HC and striatum when temporal context during retrieval matched the previously encoded temporal context. In addition, encoding related activation in the striatum, as well as HC–striatal interaction has been demonstrated for items that were subsequently remembered compared with forgotten (Sadeh, Shohamy, Levy, Reggev, & Maril, 2011), further suggesting that the striatum is critical for successful (contextual) memory processing. Alternatively, in line with theories of striatum as critical for input gating into WM (McNab & Klingberg, 2008; O'Reilly & Frank, 2006), striatal activation during the encoding phase might support subsequent interference control through the selective gating of task-relevant information.

It should be noted that no explicit instruction to encode temporal information was provided. Rather, embedding familiar trials in the incorrect position implicitly requires the participants to encode temporal information to accurately perform the task. Moreover, performing the task without encoding of temporal order information would require that responses are based on familiarity alone, resulting in a large number of false alarms (positive “yes” responses to familiar 3B lure trials). This was not the case as indicated by a high level of correct responses to lure trials.

Here, we demonstrate that engagement of a core network including the IFG, HC, and caudate nucleus during encoding is linked to subsequently lower levels of PI. Recurrent encoding-related activity fluctuations may account for trial-to-trial variability in participants' encoding performances, which is a fundamental source of variability in PI at retrieval. Current results provide insight to the processes underlying control of PI in WM and suggests that encoding of temporal context details support subsequent interference control.

Corresponding author: Jonas Persson, Aging Research Center, Karolinska Institute and Stockholm University, Tomtebodavägen 18 A, 171 65 Solna, Sweden, or via e-mail: [email protected].

Data and code can be made available upon request.

George Samrani: Data curation; Formal analysis; Investigation; Software; Writing—Review & editing. Jonas Persson: Conceptualization; Funding acquisition; Illustrations; Project administration; Supervision; Writing—Original draft; Writing—Review & editing.

The project was supported by the Swedish Research Council (https://dx.doi.org/10.13039/501100004359), grant number: 2018-01609 to J. P.

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.