N-Methyl-D-Aspartate Receptor-Antibody Encephalitis Impairs Maintenance of Attention to Items in Working Memory

Abstract

NMDA receptors (NMDARs) may be crucial to working memory (WM). Computational models predict that they sustain neural firing and produce associative memory, which may underpin maintaining and binding information, respectively. We test this in patients with antibodies to NMDAR (n = 10, female) and compare them with healthy control participants (n = 55, 20 male, 35 female). Patients were tested after recovery with a task that separates two aspects of WM: sustaining attention and feature binding. Participants had to remember two colored arrows. Then attention was directed to one of them. After a variable delay, they reported the direction of either the same arrow (congruent cue) or of the other arrow (incongruent cue). We asked how congruency affected recall precision and measured types of error. Patients had difficulty in both sustaining attention to an item over time and feature binding. Controls were less precise after longer delays and incongruent cues. In contrast, patients did not benefit from congruent cues at longer delays [group × congruency (long condition); p = 0.041], indicating they could not sustain attention. Additionally, patients reported the wrong item (misbinding errors) more than controls after congruent cues [group × delay (congruent condition), main effect of group; p ≤ 0.001]. Our results suggest NMDARs are critical for both maintaining attention and feature binding.

Significance Statement

Computational theories suggest NMDA receptors (NMDARs) are critical for actively maintaining information, while other theories propose they allow us to associate or “bind” objects features together. This is the first causal test in humans of the role of NMDARs in actively maintaining attention in working memory and feature binding. We found patients have difficulty with both these processes in support of computational models. Notably, we demonstrate that patients with NMDA receptor-antibody encephalitis are an ideal model condition to study roles of receptors in human cognition. Secondly, few studies follow these patients long after treatment. Our findings demonstrate a specific long-term neuropsychological deficit, previously unreported to our knowledge, that highlights the need for greater focus on neurocognitive rehabilitation with these patients.

Introduction

Working memory (WM) is a limited capacity system for temporary storage of information used for processing (Baddeley, 2003). The neural mechanism of WM storage and the role of attention remains unclear (Oberauer, 2019) but animal and human studies suggest N-methyl-D-aspartate receptors (NMDARs) may be essential for encoding and storage (M. Wang et al., 2013; Koychev et al., 2017).

In WM, an item in the focus of attention is maintained in a prioritized state such that it is more readily available and more accurately maintained (Zokaei et al., 2014a). Prioritization can be studied using a retrospective cue (“retrocue”) task. Here, participants see an array of items to remember. A cue is presented during the retention interval, indicating for example the location or color of one of the items, which guides the focus of attention to that item within WM. Subsequently, they are asked to recall a feature of one item. Recall of cued items is typically more accurate and faster than uncued items (Souza and Oberauer, 2016).

There may be different reasons why retrocues improve performance. Most commonly, cues indicate the probability an item will be probed (Griffin and Nobre, 2003). However, benefits are seen even if the cue causes an item to be merely accessed during the delay (Zokaei et al., 2014b). This is termed an “incidental” retrocue, as it provides no information about which item will be tested. This allows us to measure the benefit of shifting the focus of attention. What drives improvements? To answer this, we can ask what kinds of error people make on this task. Participants can be asked to reconstruct a feature from memory using a continuous scale, allowing response errors to be modeled (Bays et al., 2009; Pertzov et al., 2017) and inform how representations in WM are corrupted.

One proposed mechanism of WM retention involves NMDARs (Murray et al., 2014). Computational models of WM using artificial spiking neural networks demonstrated that without NMDARs, activity representing a stimulus dies away within 200 ms, but rapid transient increases in synaptic strength could support sustained activation during a memory delay (Compte et al., 2000; M. Wang et al., 2013). Simulations with and without NMDARs, with a decay time of 80 ms, demonstrate that NMDARs permit sustained activity. Similarly, models that test the ability of networks to bind features together into objects demonstrate that in rate-coded neurons, removing rapid Hebbian plasticity prevents the network from remembering associations but preserves feature memory (Manohar et al., 2019). Spiking versions of these networks demonstrate NMDARs could underlie this computational property (Fiebig and Lansner, 2017; Fiebig et al., 2020). In summary, computational models have demonstrated both feature binding and active maintenance over time may be crucially underpinned by NMDARs.

In patients with NMDAR-antibody encephalitis (NMDAR-Ab-E), autoantibodies found in patient CSF target the NR1 subunit of the NMDAR (Dalmau et al., 2008). It largely affects young women but is a disease of all ages and both sexes (Irani et al., 2010). Patients present with prominent mixed psychiatric symptoms, amnesia, and a movement disorder, among other features (Dalmau et al., 2008; Irani et al., 2010; Al-Diwani et al., 2019). Longer-term sequelae such as reduced executive function and memory impairments are likely influenced by time to diagnosis and treatment aggression but can persist years posttreatment (McKeon et al., 2018). NMDAR IgG antibodies may induce receptor endocytosis (Hughes et al., 2010) and selectively reduce postsynaptic available NMDARs (Jézéquel et al., 2017; Galovic et al., 2023). Accordingly, reduced NMDARs have been measured both in postmortem brain tissue and living patients with lower open NMDAR channel density on PET imaging, even long into recovery (Galovic et al., 2023). Hence, patients with NMDAR-Ab-E provide an ideal human model to ask whether selective NMDAR modulation impacts WM and attention (Finke et al., 2012; McKeon et al., 2018; Stein et al., 2020). In this study, we explored shifts of attention within WM using a task with an incidental cue and a variable delay. We measure effects of cueing and delay on memory precision and binding.

Materials and Methods

Participants

We recruited 58 healthy volunteers (HC, healthy controls) and 10 patients with NMDAR-Ab-E who provided written consent under ethics approved by the National Research Ethics Service (18/LO/2152). The mean number of years from acute presentation and diagnosis for patients was 4.02. They had all been successfully treated at initial presentation with first line immunotherapy (steroids, intravenous immunoglobulin, and plasma exchange). Three patients had required second-line immunotherapy much later (rituximab and/or cyclophosphamide), two of whom necessitated this for a relapse. Data on specific treatment was unavailable for two patients. Healthy people were recruited via an online sign-up system or a participant database, and patients were recruited from the Oxford Autoimmune Neurology Clinic. We tested 32 HCs and five patients in-person, and following COVID-19 restrictions, 26 HCs and five patients were tested remotely via an adapted online platform participants could access on their own devices. The in-person testing took place individually in a quiet room. Stimuli were displayed on a 20 inch computer monitor at an approximate viewing distance of 70 cm. To screen for mood, behavioral, and major cognitive confounders, each participant answered the HADS (Hospital Anxiety and Depression Scale; Zigmond and Snaith, 1983) and BIS-11 (Barratt Impulsiveness Scale; Patton et al., 1995) questionnaires and performed a digit span test. Both online and in-person sessions lasted ∼1.5 h, and participants were compensated for their time (£10/h).

Experimental protocol

In-person experiments were programmed using MATLAB R2018b and the Psychophysics Toolbox extension (Brainard, 1997; Pelli, 1997). Online testing was programmed with PsychoPy hosted by Pavlovia (Open Science Tools; Peirce et al., 2019). Online surveys and consent forms were generated using Qualtrics software version August 2021 of Qualtrics (Copyright © 2023 Qualtrics, Qualtrics and all other Qualtrics product or service names are registered trademarks or trademarks of Qualtrics; https://www.qualtrics.com).

Trials commenced with a display of a central fixation cross for 500 ms (Fig. 1A). After this, participants saw a memory array (1,000 ms) of two arrows (each 56 pixels long, 250 pixels either side of a white fixation cross on a black background), randomly oriented and differently colored. The colors were selected randomly from eight distinct colors (red, green, blue, yellow, cyan, magenta, orange, and white). After a 500 ms delay, participants saw the name of one of these colors (incidental retrocue, Ariel, 32 pt., text color matched name) and clicked the left or right mouse key to report which side of the screen this arrow appeared on. After a delay of 1,000 or 3,000 ms (50% probability), they were probed with either the cued arrow (congruent condition) or the uncued arrow (incongruent condition), again with 50% probability. Therefore, the retrocue was uninformative about which arrow would be probed (target arrow). The probe stimulus, an arrow, could not be the same orientation as the target arrow and always differed by >20° from the true orientation, so that adjustment was always needed. Participants responded by using the mouse to orient the probe to match the target arrow.

Figure 1.

A, Schematic illustration of experimental design. Participants encoded a memory array with two arrows of different color, orientation, and location. After the interval a retrocue was displayed, following which participants had to click the left or right mouse key depending on the remembered location of the colored arrow. In the congruent condition, the cued arrow was probed, whereas in the incongruent condition, the uncued arrow was probed. The probe stimulus appeared after a black postcue interval of delay 1,000 or 3,000 ms occurring with equal probability. B, Schematic of error sources. There are three error sources in this model demonstrated by the red and yellow colored areas and the width of the distributions centered on the target and nontarget arrow: (1) the SD (standard deviation) of a von Mises distribution of responses centered on the orientation of both the target and nontarget arrows (SD is the same for both) represents the imprecision or variability in recall of arrow orientation; (2) a uniform distribution, representing random guess responses unrelated to either of the two arrows from the array (yellow); and (3) a von Mises distribution centered on the nontarget arrow orientation represents misbinding errors (red).

Participants performed four slowed practice trials, where they received feedback about their precision, and then three blocks of 84 trials of normal delay schedule as described above (63 trials per condition). The first three controls tested in person performed four blocks of 100 trials. In the online format, participants completed four blocks of 80 trials (80 trials per condition).

Analysis

We employed a probabilistic three-factor mixture model (Bays et al., 2009; Fougnie et al., 2012) to characterize the source of errors made by participants. In this model, errors can arise from the following: imprecision (variability in memory of arrow orientation, standard deviation of the von Mises distribution), misbinding (misreporting errors where responses are centered on the nontarget arrow's orientation), or random guessing (a fixed probability; Fig. 1B). This can be mathematically described as Pertzov et al. (2017):

$$p(\hat{\theta })=\alpha {\varphi }_{\kappa }(\hat{\theta }-\theta )+\beta {\varphi }_{\kappa }(\hat{\theta }-\phi )+\gamma {\displaystyle \frac{1}{2\pi },}$$

here, $p(\hat{\theta })$ is the probability of reporting the orientation $\hat{\theta }$, $\varphi$ is the circular Gaussian (von Mises) probability density, and κ is the “concentration” parameter which represents the precision of recall around the target arrow. To obtain a value commensurable with angular error, κ was converted to a 1/κ that maps onto standard deviation in degrees, using the von Mises distribution. 1/κ has units of degrees. It is a type of error that is restricted to the item being tested; i.e., it is more selective than overall absolute error, as this would not factor in other errors such as misbinding or guessing. The target orientation is $\theta$ and the nontarget is $\phi$. The free parameters α, β, and γ, represent the probability of reporting the orientation of the correct target, the nontarget (misbinding), and responding randomly (guessing), respectively, with $\alpha +\beta +\gamma =1$. The maximum likelihood estimates of all these values were obtained from the estimated probability density function, separately for each subject and condition.

To analyze results, we used 2 × 2 × 2 mixed repeated-measures ANOVAs (patients vs HC). Where there were significant main effects or interactions of group, 2 × 2 repeated-measures ANOVAs were conducted. Statistical analyses were calculated with SPSS (v.25). Normality of data was analyzed with the Shapiro–Wilk (SW) test and homogeneity of variance with Levene's test to assess the assumptions of the ANOVA. Misbinding, guess, and target data were arcsine transformed for statistical analyses to meet the assumptions of normality for ANOVA (please see Table 1 for Shapiro–Wilk results). Data and scripts are available at OSF (https://osf.io/rcb3m/).

Table 1.

Results of the Shapiro–Wilk normality tests of model parameters for HC and patient groups

	HCs	Patients
Misbinding	W(55) = 0.670; p < 0.001	W(10) = 0.848; p = 0.058
Guess	W(55) = 0.793; p < 0.001	W(10) = 0.901; p = 0.231
Target	W(55) = 0.791; p < 0.001	W(10) = 0.915; p = 0.330
Imprecision	W(55) = 0.965; p = 0.114	W(10) = 0.867; p = 0.094

Significant results in bold indicate this group of data is not normally distributed.

Results

Trials with an incorrect retrocue response were excluded as well as trials where the time taken to initially move the mouse was >5 s, the time taken to position the arrow after first moving the mouse was <0.2 s, or total reaction time was >10 s (data removed: HC in-person = 3.74%, HC online 4.71%; NMDA in-person = 3.02%, NMDA online = 1.44%); the proportion of trials removed per group did not differ (Table 2). Participants were excluded entirely if they had retrocue accuracy lower than chance (N = 1 in-person in HC) or if the estimated precision was >2SD greater than the group mean (N = 2 online HC). One patient misunderstood instructions and repeated the study; only their second session was analyzed. Online and in-person mean data was not significantly different (Table 3) therefore, we combined the data in our analyses.

Table 2.

Independent t tests comparing proportion of removed data

Two-tailed independent samples t tests	df	t value	Sig.
HC online vs HC in-person	53	0.540	0.591
NMDA online vs in-person	8	0.639	0.551
HC combined vs NMDA combined	63	0.570	0.571

There was no significant difference in the proportion of removed data between patients and HCs.

Table 3.

Independent samples t tests comparing in-person and online data

HC online vs HC in-person
Raw error	53	0.518	0.607
Target	53	0.026	0.979
Misbinds	53	1.29	0.204
Guess	53	0.547	0.587
Imprecision	53	1.75	0.086
NMDA online vs in-person
Raw error	8	0.352	0.734
Target	8	0.116	0.911
Misbinds	8	0.035	0.973
Guess	8	0.187	0.856
Imprecision	8	1.30	0.245

There is no significant difference between data collected online versus in-person for both patients and HCs.

For overall error (raw error), defined as the absolute angular difference in degrees between the reported orientation and the actual orientation of the target arrow, patients were worse overall with no interactions with delay or congruency (Table 4). Reaction times, measured from the start of probe presentation on screen to the start of mouse movement, were correspondingly longer in patients (main effect of group, p = 0.004; Fig. 2, Table 5). Congruence sped up reaction times (main effect of congruency p < 0.001), but patients benefited less from congruency specifically when the delay was long (congruency × delay × group; p = 0.016).

Table 4.

Results of raw error analyses

Raw error	Variable	df1	df2	F	Sig.	ηp 2
Group × congruency × delay	Congruency	1	63	14.9	<0.001	0.191
	Congruency * group	1	63	0.273	0.603	0.004
	Delay	1	63	9.12	0.004	0.126
	Delay * group	1	63	0.726	0.397	0.011
	Congruency * delay	1	63	3.25	0.076	0.049
	Congruency * delay * group	1	63	0.106	0.746	0.002
	Group	1	63	5.04	0.028	0.074

Raw error is defined as the absolute angular difference in degrees between the reported orientation and the actual orientation of the target arrow in the task. Significant results in bold. Patients had worse raw error overall with no interactions with delay or congruency.

Figure 2.

Congruency reduces reaction time for both patients and HCs but patients benefit less from congruence particularly with longer delays. Reaction time is defined as the time between presentation of the probe on screen and participant initiation of response in milliseconds. There is a beneficial congruency effect for both patients and HCs but patients have longer reaction times overall and benefit less from congruency especially with longer delays. HC, healthy controls; NMDA, NMDAR-antibody encephalitis patients. Error bars represent SEM, *p < 0.05. C, congruence; D, delay; G, group.

Table 5.

Results of reaction time analyses

Reaction time	Variable	df1	df2	F	Sig.	ηp 2
Group × congruency × delay	Congruency	1	63	22.8	<0.001	0.266
	Congruency * group	1	63	0.044	0.835	0.001
	Delay	1	63	2.46	0.122	0.038
	Delay * group	1	63	0.922	0.341	0.014
	Congruency * delay	1	63	12.9	<0.001	0.170
	Congruency * delay * group	1	63	6.10	0.016	0.088
	Group	1	63	9.15	0.004	0.127
Group × delay (incongruent condition)	Delay	1	63	8.81	0.004	0.123
	Delay * group	1	63	0.044	0.835	0.001
	Group	1	63	8.19	0.006	0.115
Group × delay (congruent condition)	Delay	1	63	0.024	0.878	0.000
	Delay * group	1	63	3.78	0.056	0.057
	Group	1	63	9.94	0.002	0.136
Group × congruency (short condition)	Congruency	1	63	33.8	<0.001	0.349
	Congruency * group	1	63	1.03	0.313	0.016
	Group	1	63	9.27	<0.003	0.128
Group × congruency (long condition)	Congruency	1	63	5.91	0.018	0.086
	Congruency * group	1	63	2.05	0.157	0.031
	Group	1	63	8.71	0.004	0.122

Reaction time is defined as the time between presentation of the probe on screen and participant initiation of response in milliseconds. Significant results in bold. Reaction times were longer in overall. Congruence reduced reaction times but patients benefited less from congruency specifically when the delay was long.

We applied a probabilistic mixture model to the orientation recall responses (Fig. 1B) to analyze the contribution of different error sources to responses (Bays et al., 2009). The model yields parameters for three different kinds of error: (1) imprecision indicates the overall degradation in the quality of a memory (standard deviation calculated as 1/κ of the von Mises distribution, in degrees, around the target); (2) misbinding is the probability of reporting the orientation of the wrong item; (3) guessing is the probability of recall being completely at chance, i.e., a uniformly distributed response.

Data and scripts are available at OSF (https://osf.io/rcb3m/). Demographics and questionnaire data are presented in Table 6.

Table 6.

Participant demographics and raw questionnaire scores

Measure	HC-ip	HC-on	NMDA-ip	NMDA-on
Mean age	22.9 (3.89)	26.6 (6.93)	28.0 (6.44)	30.4 (9.07)
No. participants	31	24	5	5
Gender (M:F)	9:22	11:13	0:5	0:5
Years since diagnosis	N/A	N/A	5.09 (4.66)	3.45 (1.98)
Digit span—max forwards	7.29 (1.13)	7.13 (1.60)	6.80 (0.837)	6 (0.71)
BIS-11 (total)	60.9 (10.51)	72.6 (14.70)	64 (12.73)	64.8 (13.86)
HADS (total)	9.48 (5.55)	9.86 (6.07)	14.2 (6.30)	11.8 (8.73)

Standard deviation is in brackets. Two HC-on patients did not answer the HADS questionnaire. Independent t tests of combined HC versus combined NMDA for HADS, BIS-11, and DSF were not significant. HC-ip, healthy control in-person; HC-on, healthy control online; NMDA-ip, patient in-person; NMDA-on, patient online. Mean values and standard deviation in brackets. Years since diagnosis is based on date of clinical diagnosis with antibody testing, where this data was unavailable, patient reported date of diagnosis was used. One in-person patient was clinically diagnosed during a relapse—for them, the patient reported time of first episode was used. There were no significant differences in the performance of male and female healthy controls within both in-person and online testing groups (independent, two-tailed t tests; p > 0.05).

Recovered NMDAR-antibody encephalitis patients show paradoxical congruency effect with long delays

A 2 × 2 × 2 ANOVA was performed on imprecision (1/κ), comparing HCs and patients (Fig. 3A). There was a three-way group × delay × congruency interaction (p = 0.016; Table 7) which showed that while HCs showed the expected benefit for congruent cues and short delays, patients demonstrated a different congruency effect at long delays. To understand this three-way interaction, we examined the congruency effect for long and short delay trials separately (2 × 2 ANOVAs of group × congruency; Table 7). At short delays, there was the expected congruency benefit [main effect of congruency (short condition); p = 0.003] with no group differences. At long delays, however, the congruency benefit was greater in HCs than in the patient groups, who may even show a congruency cost [group × congruency (long condition); p = 0.041]. In other words, patients were unable to sustain the attentional benefit at longer delays. To quantify the degree to which these results were driven by individual patients, we plotted absolute error by trial type for each patient (Fig. 3B). More than half the patients displayed the paradoxical improved precision with incongruency in the long condition compared with 33% HCs. Figure 3C illustrates the negative congruency effect for patients in the long condition compared with the short condition (cluster in the right lower quadrant) while HCs have a positive congruency effect for both delays (for F statistics and effect sizes, see Table 7).

Figure 3.

Inverted congruency effect on imprecision at longer memory delays in recovered NMDAR-antibody encephalitis patients. A, Healthy people are less precise for incongruent trials and longer delays. Patients show similar congruency benefit at short delays, but an opposite direction effect (i.e., a cost when they have to report the same item twice) at longer delays. B, This graph demonstrates the result is not driven by a few patients as more than half of the patients show the paradoxical improved precision with incongruency in the long condition indicating the results are not driven by only a few patients. The blue line is the mean precision in HCs. C, The differences between congruent and incongruent conditions are shown for short delays (x-axis) and long delays (y-axis) for each participant. Patients demonstrate a negative congruency effect for long delay trials evidence by the cluster in the right lower quadrant. Blue, HCs; orange, patients. HC, healthy controls; NMDA, NMDAR-antibody encephalitis patients. Error bars represent SEM, *p < 0.05. C, congruence; D, delay; G, group.

Table 7.

Imprecision analyses results

Imprecision (1/κ)	Predictor	df1	df2	F	Sig.	ηp 2
Group × congruency × delay	Congruency	1	63	2.54	0.116	0.039
	Congruency * group	1	63	1.37	0.246	0.021
	Delay	1	63	3.61	0.062	0.054
	Delay * group	1	63	0.001	0.982	0.000
	Congruency * delay	1	63	2.38	0.128	0.036
	Congruency * delay * group	1	63	6.19	0.016	0.089
	Group	1	63	2.45	0.123	0.037
Group × congruency (short condition)	Congruency	1	63	9.61	0.003	0.132
	Congruency * group	1	63	1.29	0.259	0.020
	Group	1	63	2.74	0.103	0.042
Group × congruency (long condition)	Congruency	1	63	0.012	0.912	0.000
	Congruency * group	1	63	4.34	0.041	0.064
	Group	1	63	1.57	0.215	0.024

Patients and HCs have significantly different precision with different congruencies and delays. In the long condition, patients do not benefit from congruency, whereas HCs have better precision as would be expected. Significant results in bold.

Patients misbind more with congruency

Next, we asked what proportion of errors were due to reporting the wrong arrow (swap errors or “misbinding”). While HCs again show a standard congruency benefit, with less misbinding with congruent cues and with longer delay, unusually patients misbind more in the congruent condition rather than the incongruent condition [Fig. 4A; (group × congruency × delay ANOVA), congruency × group; p = 0.014]. We asked whether this was because patients misbound more than controls on congruent trials, or less on incongruent trials, by performing separate group × delay ANOVAs for the congruent and incongruent trials. Patients misbound more than controls on congruent trials, regardless of delay [group × delay (congruent condition); main effect of group, p ≤ 0.001], demonstrating again a congruency cost. On incongruent trials, there was no difference in misbinding between patients and HCs (Table 8).

Figure 4.

Patients misbind more overall and with congruency in the short condition. They also guess more overall than HCs. A, Healthy people with incongruency and delay. NMDA patients show greater misbinding with congruency rather than incongruency in the short condition. B, Patients and healthy people have the same pattern of guessing (more guessing with incongruency and delay) but patients guess more overall. C, This graph demonstrates misbinding data per individual NMDA patient. The blue line is the mean precision in HCs. D, This graph shows guessing data per individual NMDA patient. The blue line is the mean precision in HCs. HC, healthy controls; NMDA, anti-NMDA-R encephalitis patients. Error bars represent SEM, *p < 0.05. C, congruence; G, group.

Table 8.

Results of the misbinding analyses

Misbinding	Variable	df1	df2	F value	Sig.	ηp 2
Group × congruency × delay	Congruency	1	63	0.271	0.605	0.004
	Congruency * group	1	63	6.32	0.014	0.091
	Delay	1	63	0.719	0.400	0.011
	Delay * group	1	63	1.50	0.225	0.023
	Congruency * delay	1	63	1.22	0.274	0.019
	Congruency * delay * group	1	63	0.051	0.823	0.001
	Group	1	63	3.97	0.051	0.059
Group × delay (incongruent condition)	Delay	1	63	1.09	0.300	0.017
	Delay * group	1	63	0.535	0.467	0.008
	Group	1	63	0.004	0.948	0.000
Group × delay (congruent condition)	Delay	1	63	0.330	0.568	0.005
	Delay * group	1	63	1.69	0.198	0.026
	Group	1	63	16.8	<0.001	0.211

Significant results in bold. Patients misbind more with congruency.

Patients guess more than controls and have fewer target responses

Guessing errors occur when participants select responses randomly from a uniform distribution. Patients guessed more overall than HCs (group main effect, p = 0.031), and both groups guessed more in the incongruent condition (congruency main effect, p ≤ 0.001). There was also a congruency × delay interaction (p = 0.040) but no group interactions (Tables 9, 10).

Table 9.

Results of the guess analyses

Guessing	Variable	df1	df2	F value	Sig.	ηp 2
Group × congruency × delay	Congruency	1	63	18.9	<0.001	0.231
	Congruency * group	1	63	0.195	0.66	0.003
	Delay	1	63	0.290	0.592	0.005
	Delay * group	1	63	0.397	0.531	0.006
	Congruency * delay	1	63	4.39	0.040	0.065
	Congruency * delay * group	1	63	1.84	0.179	0.028
	Group	1	63	4.86	0.031	0.072

Significant results in bold. Patients guess more overall than HCs.

Table 10.

Results of the target analyses

Target	Variable	df1	df2	F Value	Sig.	ηp2
Group × congruency × delay	Congruency	1	63	13.3	<0.001	0.174
	Congruency * group	1	63	0.863	0.356	0.014
	Delay	1	63	1.64	0.204	0.025
	Delay * group	1	63	1.40	0.241	0.022
	Congruency * delay	1	63	8.22	0.006	0.115
	Congruency * delay * group	1	63	1.55	0.218	0.024
	Group	1	63	3.96	0.051	0.059

Significant results in bold. Patients have fewer target responses.

Discussion

NMDA receptors are important in WM storage (Lisman et al., 1998), but few causal studies in humans have asked what component mechanisms they support. Patients with NMDAR-antibody encephalitis have a selective reduction in postsynaptic NMDAR availability, including long into recovery (Galovic et al., 2023). We asked whether sustaining attention to items in WM and feature binding are affected by NMDAR hypofunction using a task that allowed modeling of error types. A striking finding was that in the long-delay condition, patients lacked the typical congruency benefit. In other words, patients had difficulty sustaining attention to prioritized items. Furthermore, patients make more misbinding errors than controls in congruent trials, suggesting they may struggle to remember associations, especially for prioritized information.

In a recent study of one-item WM, patients lacked interference from previous trials, which was explained by modeling in terms of diminished short-term potentiation caused by NMDAR hypofunction (Stein et al., 2020). Previous studies illustrated that incorporating prefrontal synaptic disinhibition (NMDAR hypofunction decreasing activation of GABA interneurons) into spatial WM modeling leads to reduced precision and greater distraction from near distractors (Murray et al., 2014). This may be one explanation for patients’ misbinding errors, as similar representations may be more difficult to differentiate. NMDAR antibodies lower receptor-mediated potentiation and currents (Hughes et al., 2010; X. Wang et al., 2019) which is evident in different cortical areas including the hippocampus (X. Wang et al., 2019) a region of high NMDAR density and a region often linked to memory (McHugh et al., 2007) particularly associative feature binding (Zokaei et al., 2019). It is possible that patients misbind more in the congruent condition as aberrant NMDAR-mediated synaptic function involved in paying attention to an item could make the representation more difficult to differentiate from the other item.

An important point to note is the possibility of potential hippocampal damage in patients which is difficult to completely distinguish from the direct effects of antibodies on NMDARs. There is a predominantly high expression of NMDARs in the hippocampus (Monaghan and Cotman, 1985) and, while we are interested in a receptor level deficit, it is likely that the deficit is partially localized to the hippocampus. MRI evidence suggests only a minority of patients may have long-term microstructural damage to the hippocampus (Finke et al., 2016: imaging mean 26.6 months after disease onset). Consistent with this, in our patient group the most recent MRI imaging of 8/10 patients posttreatment (average 26.1 months postdiagnosis) found the following: 6/10 had largely normal hippocampi; 2/10 had unreported scans—on assessment by the authors; one showed no hippocampal atrophy; and the other demonstrated mild hippocampal atrophy. Two out of 10 had no MRI imaging available. This indicates most of our patients had no substantial residual structural hippocampal damage following the acute phase of the disease.

Hebbian models of WM implicate NMDARs in binding features together to form neural “conjunctions,” potentially in the prefrontal cortex (Manohar et al., 2019). Sustained activity of neurons encoding these conjunctions may constitute attention to a specific item and is also dependent on NMDARs. Within such models, switching attention away from an item disrupts this persistent firing, but the synaptic weight changes remain. This allows unattended information to be reactivated, shifting attention back to that item. In our study, patients lost the congruency effect at long delays, with even decreased precision for the attended item. This indicates patients struggle to sustain attention on prioritized items over time. In the model this would translate to faster decay of persistent firing in conjunctive neurons. It is interesting that patients do not seem to show a decay in precision over time in the incongruent condition. One possibility is that their attention is more likely to switch between items during the delay. Alternatively, decaying persistent firing may weaken feature associations, such that the established synaptic weights are not strong enough to maintain the features of a representation (the features are bound too homogenously). As such, over time the attended representation is less precisely maintained, but the unattended representation is less affected. This may also account for patient misbinding errors in the congruent condition. If the attended item has both faster decay of persistent firing and weaker synaptic weights over time compared with the unattended item, this could lead to more overlap and interference with other features, rendering pattern separation more difficult to achieve and misbinding more likely to occur. A rate-coded model simulating this task has been previously described and demonstrates the need for Hebbian plasticity for both sustained activity and binding (Manohar et al., 2019). To directly simulate the role of NMDARs, the model would need to simulate spiking neurons. Spiking simulations have indeed shown the importance of NMDARs but have not simulated this task specifically (Fiebig et al., 2020). Future simulation work could make more quantitative predictions about how receptor density affects neural activity, memory recall, the effect of delay, and reaction times.

NMDARs are ubiquitous within the brain and this is reflected in the complex neuropsychiatric presentation of the disease, with affective, psychotic, motor and cognitive components. Similarly, WM models require NMDA plasticity in multiple locations and interaction across multiple brain areas. Neural models suggest that NMDAR-related plasticity may be required in both the cortical areas that maintain features (e.g., parietal cortex) and also cortical areas that hold the binding “pointers” (e.g., hippocampus or prefrontal cortex) (Manohar et al., 2019). This makes NMDAR-Ab-E a particularly useful model as it offers a view of specificity in mechanism at the level of cellular computations, even for cognitive functions that, like WM, appear to have little specificity to brain area. This study therefore complements previous work in patients with spatially localized lesions (Aggleton et al., 1986; Pertzov et al., 2013; Yonelinas, 2013; Allen et al., 2014).

The cohort of 10 patients we tested were mostly young women (mean age, 29.2), who had recovered from the acute illness and with good functional outcomes including return to work or full-time education. Some did report residual cognitive difficulty including anxiety, poor sleep, word finding difficulties, and “slowed reading and writing.” Potential confounding factors include the use of psychotropic and anticonvulsant medications. Two patients had been diagnosed with comorbid psychological illnesses which were being treated; two patients reported taking levetiracetam, one sertraline and one benzodiazepine. Medications or untreated symptoms may have interfered with our results and affected performance. Furthermore, some HCs were taking psychotropic medications, two of whom confirmed they were taking antidepressants. Four patients had suffered two clinical episodes of depression, one of whom was diagnosed during an encephalitis relapse. The main limitations of our study are the small sample size and limited diversity of our cohort. This is because NMDAR-E is a rare disease which more commonly occurs in young women though it is a disease of all ages and both sexes. It would therefore be beneficial to replicate the results in a larger and more diverse group of patients. Finally, a few of our patients were not clinically diagnosed with NMDAR-Ab-E until months after symptoms initially started, making the exact time of disease onset difficult to determine. Here, years from diagnosis was based on date of antibody diagnosis where the data was available, and where unavailable, based on patient reported date of diagnosis.

For 7/10 of our patients, serum NMDAR-IgG titers were available [3/7 at diagnosis (1:320, 1:320, 1:160); 4/7 from later timepoints (average of 625 d from diagnosis—1:80 d424; 1:20 d275; 1:40 d1736; 1:500 d74)]. While the presence of current antibodies could theoretically impact function, PET imaging evidence demonstrates that receptor hypofunction continues posttreatment (Galovic et al., 2023) and serum titers do not correlate perfectly with disease course (Gresa-Arribas et al., 2014). Furthermore, the effect we were particularly interested in examining here was not specifically based on the presence of antibodies in these patients but on the NMDA receptor hypofunction induced by the antibodies and the effects of this receptor deficit on WM. Therefore, antibody status/titer or their treatment-related improvements are not directly relevant to our specific objectives.

It should be highlighted this is a disease where treatment largely focuses on acute control with immunomodulation, while longer-term neurological rehabilitation to compensate for NMDAR abnormalities is hugely underexplored. These findings suggest a specific long-term neuropsychological deficit in this patient group which has not previously been reported in any other studies to our knowledge. Our results indicate patients experience WM deficits years postdiagnosis (4.02 mean years since diagnosis) highlighting an important focus for future research. Targeting NMDAR after initial recovery may provide benefit for longer-term neuropsychological sequelae. Agents could include positive allosteric modulators such as D-cycloserine which has previously been trialed with schizophrenia though effects have been poor thus far (Kuppili et al., 2021). Targeting associated proteins such as EphB2 receptors could also prove beneficial. These receptors are known to modulate NMDARs, and evidence has demonstrated that autoantibody-mediated NMDAR loss is prevented by EphB2 receptor activation (Mikasova et al., 2012).

Furthermore, poor attention and WM are important subjective findings but can be difficult to quantify especially in the context of a patient attending a medical clinic. Current clinical tools for assessing NMDAR function are minimal. There may be a clinical utility in this retrocue task, easily completed on a computer, as a neuropsychological assay to help quantify recovery posttreatment in both a clinical setting and outcome studies particularly when assessing symptoms >12 months posttreatment where specific neuropsychological tasks may prove more sensitive (Guasp et al., 2022).

Conclusion

Our data indicate that patients with anti-NMDA-R encephalitis not only have difficulty sustaining attention on prioritized items in WM but also struggle with feature binding of attended items. This implicates a role for NMDAR in maintaining attention on information in WM and in feature binding, in line with some computational models of WM. Moreover, these results highlight an important long-term neuropsychological deficit in this group of patients and suggest greater treatment focus should be directed toward neuropsychological rehabilitation.