Abstract
Associative binding is key to normal memory function and is transiently disrupted during periods of post-traumatic amnesia (PTA) following traumatic brain injury (TBI). Electrophysiological abnormalities, including low-frequency activity, are common following TBI. Here, we investigate associative memory binding during PTA and test the hypothesis that misbinding is caused by pathological slowing of brain activity disrupting cortical communication. Thirty acute moderate to severe TBI patients (25 males; 5 females) and 26 healthy controls (20 males; 6 females) were tested with a precision working memory paradigm requiring the association of object and location information. Electrophysiological effects of TBI were assessed using resting-state EEG in a subsample of 17 patients and 21 controls. PTA patients showed abnormalities in working memory function and made significantly more misbinding errors than patients who were not in PTA and controls. The distribution of localization responses was abnormally biased by the locations of nontarget items for patients in PTA, suggesting a specific impairment of object and location binding. Slow-wave activity was increased following TBI. Increases in the δ-α ratio indicative of an increase in low-frequency power specifically correlated with binding impairment in working memory. Connectivity changes in TBI did not correlate with binding impairment. Working memory and electrophysiological abnormalities normalized at 6 month follow-up. These results show that patients in PTA show high rates of misbinding that are associated with a pathological shift toward lower-frequency oscillations.
SIGNIFICANCE STATEMENT How do we remember what was where? The mechanism by which information (e.g., object and location) is integrated in working memory is a central question for cognitive neuroscience. Following significant head injury, many patients will experience a period of post-traumatic amnesia (PTA) during which this associative binding is disrupted. This may be because of electrophysiological changes in the brain. Using a precision working memory test and resting-state EEG, we show that PTA patients demonstrate impaired binding ability, and this is associated with a shift toward slower-frequency activity on EEG. Abnormal EEG connectivity was observed but was not specific to PTA or binding ability. These findings contribute to both our mechanistic understanding of working memory binding and PTA pathophysiology.
Introduction
Post-traumatic amnesia (PTA) is a transient state of cognitive impairment following traumatic brain injury (TBI), characterized by disorientation and mnemonic deficits. PTA is usually short-lived, but duration is highly variable and may persist throughout inpatient care (Marshman et al., 2013). PTA duration predicts TBI severity, with longer periods associated with poorer functional outcomes (Friedland and Swash, 2016; Ponsford et al., 2016). Reduced connectivity between parahippocampal gyrus and posterior cingulate cortex is associated with working memory impairments, which resolve on PTA emergence (De Simoni et al., 2016). This suggests that working memory impairments during PTA result from disruption to cortical communication between brain regions involved in encoding information held in working memory.
A range of memory impairments are present in PTA (Hennessy et al., 2021). In visual working memory, information about an object's identity and location must be encoded and bound together (Schneegans and Bays, 2019). Binding is a sensitive measure of working memory impairment in various disease states (Parra et al., 2010, 2015; Liang et al., 2016). A precision working memory paradigm allows recognition memory for object identification, free recall of spatial location, and the source of errors to be independently quantified (Pertzov et al., 2012). This allows detailed investigation of working memory impairment in PTA, which we use to investigate whether binding failures are associated with electrophysiological abnormalities that potentially disrupt cortical communication.
Neural oscillations are central to working memory function (Luo and Guan, 2018). Long-range θ phase synchronization supports communication between PFC and temporal lobe during encoding, retrieval, and working memory maintenance (Fell and Axmacher, 2011). Frontal θ phase also modulates temporal-parietal γ amplitude in a process known as phase-amplitude coupling (PAC). θ-γ PAC is important in encoding, working memory, and associative binding and is mechanistically supported by long-range θ synchronization (Lega et al., 2016; Daume et al., 2017; Köster et al., 2018). θ rhythms, and interactions with other frequencies, therefore have a primary role in associative memory that likely underpins memory processes at the large-scale network level (Jann et al., 2010; Yuan et al., 2012; Hacker et al., 2017). The precise synchronization of neural oscillations necessary for working memory may be sensitive to disruption by large electrophysiological abnormalities that often occur transiently after head injury.
Increased low-frequency oscillations are seen in many disease states, including TBI (Modarres et al., 2017; Jafari et al., 2020). Increased slow-wave activity is associated with poorer neurologic outcome, personality change, and cognitive impairment following TBI (Huang et al., 2014; Robb Swan et al., 2015). Relative changes can be described using δ-α ratio (DAR) reflecting the difference in contribution between δ and α. Reducing DAR using transcranial electrical stimulation improves cognition after TBI (Ulam et al., 2015). Synchronization disturbances are also present following TBI, including decreased γ connectivity (Wang et al., 2017), reduced connectivity across θ, α and β associated with working memory impairment (Kumar et al., 2009; Bailey et al., 2017), and improved cognition with reduced δ connectivity (Castellanos et al., 2010). Cross-frequency coupling is also disrupted following TBI (Antonakakis et al., 2016). Together, this suggests that a shift to pathological low-frequency power after TBI may be mechanistically important for working memory disturbance by disrupting the neural oscillations that support the maintenance of information in working memory.
Here for the first time, we investigate working memory binding in an acute cohort of moderate to severe TBI patients with and without PTA. Electrophysiological abnormalities are quantified using EEG. We test the following hypotheses: (1) PTA patients will show increased misbinding errors; (2) PTA will be associated with increased low-frequency oscillations; (3) binding impairment will correlate with increased low-frequency oscillations reflected in increased DAR; (4) PTA will be associated with abnormal long-range θ phase synchronization and θ-γ PAC; and (5) emergence from PTA will be associated with a reduction in misbinding and a normalization of EEG measures.
Materials and Methods
Participant demographics and clinical details
TBI group
Thirty patients (25 males, 5 females, mean age 40.73 years, range 17-73 years) admitted with a recent history of TBI were recruited from the Major Trauma Ward, St Mary's Hospital, London. All patients had moderate to severe injuries according to the Mayo Classification system for TBI severity (Malec et al., 2007). Injuries were secondary to road traffic accidents (40.0%), falls (33.3%), assault (20.0%), and sports injury (6.67%). Patients were included if they were between the ages of 16 and 80 and clinically stable. Exclusion criteria were as follows: premorbid psychiatric or neurologic illness; history of other significant TBI; current or previous drug or alcohol abuse; pregnancy or breastfeeding; significant language or visuospatial impairments. For the EEG part of the study, neurosurgical intervention (or other contraindication to scalp EEG) was also an exclusion criterion. Detailed clinical characteristics of all TBI patients are available in Extended Data Table 1-1.
Clinical demographics for all participantsa
Table 1-1
. Download Table 1-1, DOCX file.
Written informed consent was obtained for patients judged to have capacity according to the Declaration of Helsinki. Patients in PTA who were judged not to have capacity were deemed unable to give informed consent for participation in the study. In this case, written assent was obtained as well as informed written assent by a caregiver on the patient's behalf. Informed consent was gained retrospectively once patients emerged from PTA. No patients withdrew consent. The study was approved by the West London Research Ethics Committee (09/HO707/82).
Control group
Twenty-six healthy controls (20 males, 6 females, mean age 28.96 years, range 18-70 years) were recruited friends/relatives of TBI patients or through word of mouth. Participants had no history of psychiatric or neurologic illness, previous TBI, or alcohol or substance misuse. All participants gave written informed consent.
Protocol
At baseline, all participants underwent neuropsychological assessment, precision working memory task, and resting-state EEG. Additionally, patients underwent PTA assessment according to the Westmead Post Traumatic Amnesia Scale (WPTAS) (Shores et al., 1986). The WPTAS is a 12-item scale with seven items that assess orientation and five items that assess memory. Patients must score full marks (12 of 12) for 3 consecutive days before they are deemed to no longer be in PTA. PTA duration is calculated as the time between injury and the first of these successful consecutive assessments. Patients were on average within 10 d of injury (range 1-32 d) and were divided into two groups (PTA+ and PTA–). PTA status was confirmed using clinical notes and WPTAS scores on the days surrounding the assessments. Patients were thus classified as PTA+ or PTA– according to standard clinical criteria as detailed above. Patients were invited to attend follow-up assessment within 6 months of hospital discharge, at which the baseline protocol was repeated. Controls were assessed at one time point.
Where possible, participants took part in all aspects of the study. However, because of the nature of recruitment within an acute clinical setting, this was not always possible or appropriate. A subsample of 17 patients and 21 controls underwent resting-state EEG. Of these, there was 1 patient and 1 control who did not complete the precision working memory task.
Neuropsychological assessment
All participants completed a detailed neuropsychology assessment with a focus on episodic and working memory. Immediate and delayed verbal recall was assessed using the Logical Memory I and II subtests of the Wechsler Memory Scale, Ed 3 (Wechsler, 1997). Immediate and delayed visuospatial memory was assessed using the Brief Visuospatial Memory Test–Revised (Benedict et al., 1996). Delayed recall elements of the neuropsychology tasks were administered 25 min after the initial assessment. A battery of computerized tests based on classical paradigms from the cognitive psychology literature was delivered on a tablet device using a custom-programmed application. Details of each task have been previously reported (Hampshire et al., 2012). In brief: Visuospatial Working Memory, based on a task from the nonhuman primate literature was used to assess visuospatial working memory (Inoue and Matsuzawa, 2007); Paired Associates, based on a paradigm commonly used to assess memory impairments in aging clinical populations, was used to assess object-location association memory (Gould et al., 2005); Spatial span, based on the Corsi Block Tapping Task, was used to assess spatial short-term memory capacity (Corsi, 1973); Self-Ordered Search was used to measure strategy during search behavior, based on an existing task (Collins et al., 1998); and Feature Match, based on the classical feature search task, was used to measure attentional processing (Treisman and Gelade, 1980).
Statistical analysis of neuropsychological data
One-way ANOVAs were used to identify group effects at baseline. Post hoc independent-samples pairwise t tests, using false discovery rate (FDR) multiple comparisons corrections were performed to determine which pairwise comparisons were driving any significant main effects. Linear mixed-effects models were used to assess longitudinal changes between baseline and follow-up in which group and time point were defined as fixed effects and subject was defined random intercepts included for each subject. Post hoc paired samples t tests were used to investigate any significant main effects or interactions. Unless otherwise stated all statistical analysis was performed using R (version 1.3.1056).
Experimental task paradigm
Classic span tasks are sensitive to general impairments of working memory but do not allow the binding of information to be specifically studied. Recently, precision working memory tasks have been developed that allow recognition memory for object identity and spatial location to be measured separately from the binding of this information (e.g., providing separate tests of whether an object was remembered) and whether the spatial location of that object was remembered. Participants completed a precision recall working memory task (Fig. 1A) based on Pertzov et al. (2012).
Precision working memory task. A, Experimental task paradigm. One or two fractal objects are shown (Encoding Array) before a delay of 2 s (Maintenance). One of the original objects (Target) is then displayed alongside a novel object (Nontarget), which had not appeared in the encoding array. Participants are required to select the object they recall (Object Identification) and move it to its remembered location (Object Localization). B, Transformation of data from screen space to transformed space in which the target lies at the origin, the nontarget at 1,0 (x, y), and the object placement for each trial is plotted in relation to these constants. C, The placements across all trials are visualized using the distribution of responses across the x and y axes in the normalized space. Responses clustered around the target indicate correct responses, and- second cluster around the nontarget indicated misbinding errors.
The stimuli consisted of 60 fractal images (Sprott, 1996). Fractals were used as complex visual objects that unlike images of everyday objects cannot be readily verbalized. A maximum width and height of 120 pixels was used. Stimuli were presented on an interactive touch-sensitive screen with a 1920 × 1080 pixel matrix (Dell). The experiment consisted of 80 trials (20 1 item; 60 2 item). Object location was determined by a MATLAB script (The MathWorks) with (x,y) coordinates randomly drawn from across the horizontal and vertical pixel dimensions of the screen, respectively, with the following exceptions: objects were never located within 600 pixels of each other within a single trial. They were positioned within a minimum distance of 200 pixels from the screen edge. The threshold for the distance at which the response matched the target (or distractor) was set to 200 pixels.
Task procedure
Each trial began with a central fixation point followed by the memory array consisting of 1 or 2 distinct fractals presented for 2 s. A blank screen was displayed for 2 s (maintenance period) after which the object identification stage began in which two fractals were displayed alongside one another in the center of the screen. One fractal had been present in the memory array of that trial, the other was a foil item that was not present in the memory array. Participants were required to touch the item that they remembered from the memory array and then drag it to the remembered location. Localization performance was only analyzed in trials in which the object identity was correctly remembered.
Task analysis
Data from the precision working memory task were analyzed using a thresholded approach in which a correct localization was considered to be within 200 pixels of the target (or a misbinding error within 200 pixels of the nontarget) to calculate a proportion of misbinding errors. A novel, threshold-free approach to study the distribution of responses across a transformed space, defined by the relative locations of target versus nontarget items was used to visualize the distribution of responses across trials (Fig. 1B). In this transformed space, the target position is located at the origin and the nontarget at 1,0 (x,y). Therefore, responses deviating from the origin along the x axis indicate a spatial bias by the nontarget item location, as happens when the target identity is misbound with the nontarget location. In contrast, responses deviating along the y axis do not indicate a spatial bias by the nontarget location but are in keeping with imprecision in remembering the correct location, or with random error (Fig. 1C). The responses in this transformed space are binned across 10 uniformly spaced bins. This approach allows the influence of the nontarget on the remembered location to be visualized without dichotomizing responses based on arbitrarily defined thresholds. The design of this paradigm thus allowed us to measure the distinct processes of object-location associative memory independently to quantify the types of error made (i.e., failure to remember the location or failure to bind).
EEG data acquisition
EEG data were acquired using a 32-channel active electrode standard actiCAP positioned with Cz at the midline with electrodes grounded to FPz and referenced to Fz. Signals were amplified using the actiCHamp system and data recorded using the BrainVision Recorder software (Brain Products) at a sampling rate of 1000 Hz.
EEG preprocessing
Raw EEG data were exported from BrainVision into MATLAB via EEGLAB. Data were preprocessed, cleaned, and quality assessed using the Harvard Automated Processing Pipeline for EEG (HAPPE), an automated preprocessing pipeline specifically developed for high artifact data as would be expected in acute TBI patients. HAPPE has been shown to be superior in optimizing signal-to-noise ratio compared with manual editing in clinical data (Gabard-Durnam et al., 2018). The following preprocessing steps were taken using the HAPPE pipeline: Data were high-pass filtered at 1 Hz, low-pass filtered at 100 Hz, and bandpass filtered at 1-249 Hz as recommended for HAPPE processing. Electrical line noise at 50 Hz was removed using the CleanLine multitaper approach (Mullen T, 2012). Bad channels (i.e., those with poor signal quality) were rejected using joint probability evaluation, and those exceeding 3 SDs from the mean were excluded from further analyses (to be later interpolated). Wavelet-enhanced independent component analysis was performed to correct for EEG artifact while retaining the whole dataset to improve independent component analysis decomposition. Independent component analysis was performed on the corrected data, and components were rejected using the multiple artifact rejection algorithm (Winkler et al., 2011). HAPPE rejects components with artifact probabilities > 0.5. Data were segmented into 5 s segments and subjected to amplitude-based (±40 µV) and joint probability (<3 SDs relative to the activity of other segments) rejection to remove segments with remaining artifact. Previously rejected channels were repopulated using spherical interpolation. Data were rereferenced to average.
EEG analysis
Channels were grouped into four regions: frontal (Fp1, Fp2, F3, F4, F7, F8), temporal (T7, T8, TP9, TP10, FT9, FT10), parietal (Pz, P3, P4, P7, P8), and occipital (O1, O2, Oz). Frequency bands were defined as follows: δ (0-4Hz); θ (4-8Hz); α (8-13Hz); β (13-30Hz) and γ (30-40Hz).
Normalized power
Power in each channel was calculated for each frequency band using multitaper frequency transformation normalized to total power across all five bands. Global power was calculated by averaging across all channels.
Statistical comparison of global normalized power was conducted using one-way independent-measures ANOVA followed by post hoc t tests. All p values were corrected using FDR method for multiple comparisons. Group-level statistical analysis of normalized power was also performed using a cluster-based permutation approach in the frequency/channel domain on the whole montage (Maris and Oostenveld, 2007). Power was compared between groups at each channel using two-sided independent-samples t tests and results clustered according to spatial adjacency at p < 0.05 using the maximum size criterion. Permutation distributions were generated using the Monte-Carlo method and 5000 random iterations, and corrected p values were then obtained through comparison of observed data to the random distributions. Follow-up data were assessed as above. No cluster-based permutation analysis was performed on follow-up data.
Phase synchronization
Phase synchronization was quantified using the debiased weighed phase lag index (dwPLI), a measure of phase-based functional connectivity. Phase lag index (PLI) calculates to what extent the phase of one signal is consistently lagging or leading relative to another signal, regardless of the magnitude of the phase leads and lags (Stam et al., 2007). The weighted PLI is weighted by the imaginary component of the cross-spectrum to overcome issues with spuriously related connectivity, which can arise because of volume conduction (Vinck et al., 2011). A further debiasing term to correct for inflation because of small sample size was additionally implemented within the ft_connectivity_wpli.m function in FieldTrip (Oostenveld et al., 2011). dwPLI is calculated across all frequencies within the band of interest for each channel pair.
To statistically compare connectivity at the group level, we constructed 31 × 31 whole-brain channel-wise connectivity matrices for each subject by averaging dwPLI values for each channel pair across the frequency band. The network-based statistic (NBS) implemented in MATLAB was used to compare connectivity across groups using an independent-samples t test design (Zalesky et al., 2010). The NBS is a nonparametric statistical method in which values at every node are tested against the null hypothesis and those surviving the primary threshold are entered for Monte-Carlo simulation permutation testing at every channel pair. The primary threshold was set to z = 3.1, and 10,000 random permutations were conducted with a threshold of p < 0.05.
PAC
To quantify the intensity of PAC between θ phase and γ amplitude, we computed the modulation index (MI) (Tort et al., 2010). We estimated the phase of frequencies between 4 and 8 Hz (in steps of 1 Hz) in the frontal channel group and the amplitudes of frequencies between 30 and 40 Hz (in steps of 2 Hz) in parietal and temporal channel groups individually. The MI was calculated separately for each electrode within the respective channel groups.
To test for group differences in PAC, MI values were averaged across frontal-parietal and separately frontal-temporal channel groups for each participant and compared using independent-samples Welch t tests. In patients who returned for follow-up, within-subjects t tests were used to assess differences in the dwPLI and MI across time.
Relationship between EEG and behavioral measures
We assessed the relationship between EEG measures and clinical information, such as duration of PTA or behavioral performance, using Pearson correlations. To assess whether the relationship between variables differed between groups, the correlation coefficients were compared using Fisher's z statistic implemented in the 'cocor' package in R.
Data availability
The data that support the findings of this study are available from the authors on reasonable request.
Results
Clinical demographics
Thirty moderate to severe TBI patients were recruited during their inpatient stay (age range 17-73 years; Table 1). Seventeen patients were in PTA at the time of enrollment (PTA+, mean WPTAS 9.18, SD = 1.38). Thirteen patients were not in PTA (PTA–, WPTAS 12 for all). The PTA+ group had a longer PTA duration and longer hospital admission, but groups did not differ in time since injury (Table 1). PTA+ and PTA– groups were well matched for age (p = 0.95) and gender (Table 1).
Twenty-six control participants were recruited (age range 18-70 years; Table 1). ANOVA showed an effect of age across the groups driven by controls being younger than both PTA+ and PTA– patients (Table 1). Additionally, there were group differences in years of education because of the PTA+ group having fewer years of education than controls and the PTA– group (Table 1).
Neuropsychological performance
TBI patients generally had significant cognitive abnormalities (Fig. 2; Table 2), with impairments in the following relative to controls: immediate and delayed verbal memory, immediate and delayed visuospatial memory, associative working memory, spatial short-term memory, search strategy, and attentional processing. There were no group differences in retention rates between immediate and delayed recall for either verbal memory or visuospatial memory.
Neuropsychology at baselinea
Neuropsychological performance of PTA+ and PTA– TBI patients and healthy controls at baseline. A, Visuospatial memory performance as measured using the Brief Visuospatial Memory Test (top row) and verbal memory performance as measured using the Logical Memory test (bottom row) for immediate recall, delayed recall, and retention. B, Performance on the computerized battery of tests on the paired associates learning, visuospatial working memory, attentional processing, search strategy, and spatial short-term memory tasks. ***p < 0.001. **p < 0.01. *p < 0.05. Error bars indicate standard error of the mean. CON, Healthy controls.
Patients in PTA also showed impairments compared with both PTA– patients and controls in the following: immediate verbal recall, delayed visuospatial memory recall, associative working memory, and search strategy. PTA patients also showed impairments compared with controls in delayed verbal memory recall, immediate visuospatial memory recall, visuospatial working memory, short-term spatial memory, and attentional processing (Fig. 2; Table 2).
Impaired working memory binding in PTA
A precision working memory task was used to assess the binding of object and location information in working memory. All three groups identified the target with >90% accuracy, indicating that subjects understood the task regardless of whether they were in PTA (Fig. 3A). As expected, accuracy decreased as working memory load increased, with two item trials less accurate than one item trials (F(1,51) = 55.32, p < 0.001). There was also a group effect of object identification (F(2,51) = 9.98, p < 0.001): PTA+ patients showed more errors in identifying targets than PTA– patients (t(105) = 3.49, p = 0.001) and controls (t(105) = −5.46, p < 0.001). A group × load (1 or 2 item trial) interaction was present of borderline significance (F(2,51) = 2.89, p = 0.065).
Performance on the object-location association precision working memory task in healthy controls and PTA+ and PTA– TBI patients at baseline and follow-up. A, Object identification accuracy at baseline for 1 and 2 item trials. B, Percentage of correctly identified trials that were moved to the correct location in 1 and 2 item trials at baseline. C, Scatter-histogram represents the distribution of location responses correctly identified trials in a normalized space in reference to the target and nontarget locations at baseline. D, Percentage of correctly identified trials that resulted in a misbinding error at baseline. E, Scatter-histogram represents the distribution of location responses for PTA+ and PTA– TBI patients that returned for follow-up. F, Proportion of misbinding errors made in correctly identified trials by PTA+ and PTA– patients at baseline and follow-up. ***p < 0.001. **p < 0.01. *p < 0.05. Error bars indicate standard error of the mean. CON, Healthy controls.
Similar results were observed when considering the spatial accuracy of target placement (Fig. 3B). A group × load interaction was present (F(2,51) = 7.32, p = 0.002) because of PTA+ patients showing less accurate placements compared with PTA– patients (t(105) = 4.10, p < 0.001) and controls (t(105) = −6.31, p < 0.001) in 2 item trials. Figure 3C shows the distribution of responses in relation to the target and nontarget items in the normalized space.
Next, we quantified the numbers of misbinding errors, defined as a response being placed within 200 pixels of the nontarget location. There was a group effect on misbinding errors (F(2,51) = 25.62, p < 0.001), the result of PTA+ patients making significantly more misbinding errors than PTA– (t(51) = −5.60, p < 0.001) and controls (t(51) = 6.65, p < 0.001; Fig. 3D). There was no relationship between years of education and misbinding errors in TBI patients (R = −0.43, p = 0.10).
Impaired binding ability is transient and specific to a period of PTA
Eighteen patients returned for follow-up (average 182 d after baseline). Cognitive function generally improved at follow-up (Fig. 4; Table 3), and memory binding in PTA+ patients normalized (Fig. 3E). There was a group × visit interaction (F(1,15) = 20.99, p < 0.001). This was driven by a reduction in misbinding errors in PTA+ patients between visits but no longitudinal change in PTA– patients (Fig. 3F; PTA+ v1 > PTA+ v2 (t(7) = 4.60, p = 0.0025); PTA – v1 > PTA– v2 (t(8) = −0.180, p = 0.862)).
Neuropsychology at follow-upa
Neuropsychological performance of PTA+ and PTA– TBI patients who returned for follow-up at visit 1 and visit 2. A, Visuospatial memory performance as measured using the Brief Visuospatial Memory Test (top row) and verbal memory performance as measured using the Logical Memory test (bottom row) for immediate recall, delayed recall, and retention. B, Performance on the computerized battery of tests on the paired associates learning, visuospatial working memory, attentional processing, search strategy, and spatial short-term memory tasks. Error bars indicate standard error of the mean.
Patients in PTA show increased low-frequency power
Resting EEG showed changes in power across a range of frequency bands in patients with PTA (δ: F(2,35) = 5.97, p = 0.009; α: F(2,35) = 9.05, p = 0.003; γ: F(2,35) = 5.08, p = 0.014; FDR-corrected; Fig. 5A). In the δ band, these effects were driven by PTA+ patients exhibiting increased power compared with controls (PTA+ > CON: t(35) = 3.43, p = 0.005) and a trend toward an increase compared with PTA– (PTA+ > PTA–: t(35) = −2.16, p = 0.057). In the α band, PTA+ showed reduced power compared with controls (PTA+ < CON: t(35) = −4.25, p < 0.001) and a trend toward a decrease compared with PTA– (PTA+ < PTA–: t(35) = 2.09, p = 0.065). PTA+ patients also showed reductions in β power (PTA+ < CON: t(35) = −3.01, p = 0.014; PTA– < CON: t(35) = −2.66, p = 0.018). In the γ band, the PTA– group showed reduced power compared with controls (PTA– < CON: t(35) = −2.97, p = 0.016), with no differences in the PTA+ group.
Global normalized power in healthy controls, PTA+, and PTA– TBI patients at baseline. A, Mean global normalized power for all groups across all frequency bands. B, Global δ to α ratio across all groups. C, Topoplots represent mean distribution of power for (top to bottom) healthy controls (CON), PTA+, and PTA– groups in (left to right) δ, θ, α, β, and γ frequency bands. Blue represents low power; red represents higher power. ***p < 0.001. **p < 0.01. *p< 0.05. Error bars indicate standard error of the mean.
Visual inspection of topoplots displaying group means (Fig. 5C) indicated abnormal patterns of power across multiple frequency bands in both TBI groups. Absolute differences between groups are displayed as raw contrasts in Figure 6A. Cluster-based statistics were used to quantify the differences between groups (Fig. 6B). In the δ band, direct comparison between patient groups and controls showed increased power for both PTA+ (p = 0.002) and PTA– patients (p = 0.002) in frontal, parietal, and temporal channels, with PTA+ showing a more widespread cluster extending into occipital channels. There were no significant clusters arising from a contrast between PTA+ and PTA– patients. Conversely, in the α band, PTA+ showed reduced (p = 0.001) occipital and right parietal power compared with controls, and widespread reductions compared with PTA– in frontal, temporal, and parietal channels (p = 0.007). There were no significant differences between PTA– and controls. In the θ band, PTA+ patients showed reduced θ power (p = 0.049) compared with PTA– in a right parietal-occipital cluster of channels, while PTA– showed increased θ power (p = 0.003) compared with controls in temporal and parietal channels. Changes in the β band were of a similar pattern to θ. The PTA– group showed increased power in temporal and parietal channels compared with controls (p = 0.023) and increased parietal power compared with PTA+ (p = 0.004). There were no clusters found between any of the groups in the γ frequency band.
Contrasts in normalized power between healthy controls, PTA+, and PTA– TBI patients at baseline. A, Absolute differences in power across each channel displayed as raw contrasts for (left to right) δ, θ, α, β, and γ frequency bands between (top to bottom) PTA+ and PTA–, PTA+ and healthy controls (CON), and PTA– and healthy controls. B, Cluster-based statistical comparisons of power in (left to right) δ, θ, α and β frequency bands between (top to bottom) PTA+ and PTA–, PTA+ and healthy controls, and PTA– and healthy controls (CON). Areas of red represent significant clusters (p <0.05) where power is increased. Areas of blue represent significant clusters where power is decreased.
EEG abnormalities in the PTA+ group were transient, and power had normalized at follow-up (Fig. 7A). There was a significant effect of visit in δ and α (δ: F(1,8) = 13.65, p = 0.006; α: F(1,8) = 8.37, p = 0.020). This was the result of increases in α and decreases in δ in the PTA+ group between baseline and follow-up (α: PTA+ v1 < PTA+ v2 (t(4) = −4.0726, p = 0.030); δ: PTA+ v1 > PTA+ v2 (t(4) = 4.1234, p = 0.029)). There was also a significant group × time interaction in β and γ, but no longitudinal effects were observed in θ (β: F(1,8) = 13.23, p = 0.007; γ: F(1,8) = 6.30, p = 0.036). Figure 7C depicts the spatial distribution in channel space of the change across time (follow-up minus baseline) for each frequency band for PTA+ and PTA– patients.
Global normalized power in PTA+ patients and PTA– TBI patients at follow-up. A, Mean global normalized power for all patients with baseline and follow-up across all frequency bands. B, Global δ to α ratio at baseline and follow-up in PTA+ and PTA– TBI patients. C, Topoplots represent change between baseline and follow-up (follow-up minus baseline raw contrast) in PTA+ and PTA– TBI patients in power across all frequency bands. Blue represents absolute reduction in power at follow-up from baseline; white represents no change; red represents absolute increase in power at follow-up from baseline. **p < 0.01. *p < 0.05. Error bars indicate standard error of the mean.
EEG differences between PTA+ and PTA– were particularly marked in δ and α bands, so the DAR was calculated. DAR has been used as a sensitive marker of cerebral dysfunction (Claassen et al., 2004; Schleiger et al., 2014; Finnigan et al., 2016). Group differences in global DAR were present (Fig. 5B; F(2,35) = 9.12, p < 0.001), the result of significantly higher in PTA+ compared with both PTA– (t(35) = −2.51, p = 0.025) and controls (t(35) = 4.26, p < 0.001). Abnormalities in DAR normalized at follow-up, as demonstrated by a significant group × time interaction (F(1,8) = 5.48, p = 0.047), the result of a reduction in DAR in PTA+ with no change in PTA– group (PTA+ v1 > PTA+ v2: t(4) = 3.008, p = 0.040; Fig. 7B).
Individual case studies
To better describe the transient binding impairment and how this might relate to the transient shift toward slow-wave power, we considered these changes at the single-patient level. Figure 8 illustrates four individual case studies to highlight that the EEG changes reported here are more sensitive to abnormalities occurring during a period of PTA than conventional routine clinical imaging. Case studies 1-3 show TBI patients during a period of PTA at baseline and at follow-up once they were no longer in PTA. Case study 4 shows a TBI patient who was not in PTA. We include these case studies to highlight that the group-level findings are also observable in individual patients and therefore have the potential to be clinically meaningful.
Individual case studies. CT head scans at admission are shown on the far left of each case study panel. The distribution of responses in correctly identified trials in the precision spatial working memory task are shown for baseline and follow-up in the middle panels. Topoplots (right) represent changes in δ (top) and α (bottom) between baseline and follow-up.
Case study 1 is a 45-year-old male with a moderate to severe TBI acquired through a fall from standing. Clinical imaging reported presence of subdural hemorrhage, subarachnoid hemorrhage, bifrontal contusions, and midline shift. On the day of assessment (day 8 after injury, WPTAS score 8), he was clinically deemed to be in PTA and had a total PTA duration of 12 d. At baseline, he showed a bias toward the nontarget on the precision working memory task with 34% rate of misbinding errors. He showed a dramatic improvement at follow-up to 12.7% misbinding errors. At baseline, he had a global DAR of 6.46, which reduced to 0.99 at follow-up.
Case study 2 is a 26-year-old male with a moderate to severe TBI acquired through a cycling collision. There was no intracranial hemorrhage or space-occupying lesions on the initial clinical imaging. Further imaging with MRI revealed evidence of diffuse axonal injury. He was assessed on day 24 after injury, scoring 7 on the WPTAS and thus still in a period of PTA. He had a total PTA duration of 38 d. Working memory binding performance at baseline was poor with misbinding of 22.6%, which reduced to 5.6% at follow-up. At baseline, his global DAR was 6.52 and reduced to 0.82 at follow-up.
Case study 3, a 67-year-old male, acquired a moderate to severe TBI acquired while cycling. He had a total PTA duration of 28 d. Clinical imaging showed a left subdural hemorrhage, left temporal contusions, (probable) right extradural hemorrhage, skull base fractures, and pneumocephalus. On assessment, he was clinically deemed to be in PTA (WPTAS score 8; day 5 after injury). At baseline, he made 20% misbinding errors, which improved to 5.6% at follow-up. His global DAR was 4.21 at baseline and reduced to 0.75 at follow-up.
Case study 4 is a 33-year-old male with a moderate to severe TBI acquired through a road accident as a pedestrian. Clinical imaging reported left-sided extra-axial hematoma with associated comminuted fracture involving the left frontal bone. Evidence of diffuse axonal injury was present on MRI. He had a PTA duration of 0 d and was thus not in PTA at the time of assessment (WPTAS score 12; day 5 after injury). At baseline, he demonstrated extremely low rates of misbinding (1.7%), which increased slightly at follow-up to 3.4%. At baseline, he had a global DAR of 2.04, which decreased to 0.48 at follow-up.
Power abnormalities in acute TBI are associated with working memory performance
To quantify the relationship between EEG measures and working memory performance following TBI, we grouped PTA+ and PTA– patients to examine the relationship between a shift toward low-frequency power and cognitive performance (Fig. 9). We first assessed δ and α individually to understand what the relative contributions of each of these frequencies were to any relationships between DAR and cognition.
Relationship between global normalized power and cognitive impairment. A, Global δ power was increased in the TBI group compared with healthy controls and correlated with proportion of misbinding errors and performance on the paired associates learning task. B, Global α was significantly decreased in TBI patients compared with healthy controls, and significantly correlated with the performance on the paired associates learning task. C, Global DAR was significantly increased in TBI patients compared with healthy controls and correlated with the proportion of misbinding errors, and performance on the paired associates learning task. **p < 0.01. *p < 0.05.
Global δ power was associated with working memory performance in TBI patients but not controls (Fig. 9A; Misbinding errors (TBI: R = 0.52, p = 0.038; Controls: R = −0.07, p = 0.780); paired associates learning (TBI: R = −0.56, p = 0.018; Controls: R = 0.24, p = 0.38)). The correlation coefficients differed significantly between patients and controls for the relationship between global δ and performance on the paired associates learning task (z = 2.51, p = 0.01) but did not differ for the relationship between global δ and misbinding errors (z = −1.75, p = 0.08).
Global α power was positively associated with performance on the paired associates learning task in TBI patients but showed a negative relationship to performance in controls (TBI: R = 0.74, p < 0.001; Controls: R = −0.58, p = 0.025). The relationship between global α and performance on the PAL was significantly different between patients and controls (z = −4.27, p < 0.001). Global α showed a trend toward a relationship with misbinding errors in patients but showed no relationship in controls (Fig. 9B; TBI: R = −0.49, p = 0.06; Controls: R = 0.15, p =0.52). There was a trend toward a difference in the correlation coefficients between patients and controls (z = 1.87, p = 0.06).
The global DAR significantly correlated with working memory performance in TBI patients but not healthy controls (Fig. 9C; Misbinding errors (TBI: R = 0.58, p = 0.02; Controls: R = −0.23, p = 0.33); paired associates learning (TBI: R = −0.67, p = 0.01; Controls: R = 0.44, p = 0.10)). The relationship between DAR and working memory performance differed significantly between TBI patients and controls (Misbinding errors (Fisher's z = −2.45, p = 0.01); paired associates learning (Fisher's z = 3.21, p = 0.001)).
Attentional processing was not associated with power in either δ or α frequency bands nor the DAR in either TBI patients or controls (global δ: TBI (R = −0.15, p = 0.56); Controls (R = 0.11, p = 0.71); global α: TBI (R = 0.057, p = 0.83); Controls (R = 0.22, p = 0.42); DAR: TBI (R = −0.21, p = 0.42); Controls (R = −0.039, p = 0.89). The total duration of PTA, a proxy of injury severity, was not associated with power in either δ or α frequency bands nor in the DAR (global δ (R = 0.23, p = 0.39); global α (R = −0.27, p = 0.32); DAR (R = 0.46, p = 0.08)).
The degree to which DAR was increased at baseline did not predict misbinding errors at follow-up in TBI patients (R = 0.31, p = 0.38). Rather, there was a relationship between normalization of EEG measures and resolution of working memory impairment across time. When looking at changes between baseline and follow-up in TBI patients, there was a strong positive correlation between change in global DAR and change in misbinding errors (follow-up minus baseline; R = 0.87, p = 0.001).
TBI patients demonstrate θ hyperconnectivity and α hypoconnectivity
One mechanism through which a shift toward lower-frequency power could be disrupting cortical communication is through altered connectivity across large-scale networks. In order to test connectivity within the δ, θ, and α bands, phase synchronization between electrodes was quantified using the dwPLI. Whole-brain connectivity matrices were constructed on a channel-wise basis for controls and patients in each band separately (Fig. 10A). Network-based statistics revealed that patients show θ hyperconnectivity compared with controls across one robust network consisting of 19 edges and 15 nodes (p = 0.005; Fig. 10B). Connectivity within the θ network was not related to binding ability in TBI patients or Controls (TBI: R = 0.19, p = 0.49; Controls: R = 0.09, p = 0.70; Fig. 10D). There was no difference in the mean connectivity within the θ network between PTA+ and PTA– patients (t(12.36) = 0.66, p = 0.52). There was a significant relationship between the mean dwPLI across the θ network and the total duration of PTA (R = 0.57, p = 0.016; Fig. 10D).
Phase synchronization. A, Whole-brain connectivity matrices in δ, θ, and α frequency bands for healthy controls (top row), all TBI patients at baseline (middle row), and all TBI patients returning for follow-up (bottom row). B, Network-based statistics revealed one robust network of hyperconnectivity in TBI patients at baseline compared with controls in the θ band and (C) one robust network of hypoconnectivity in the α band. D, Mean dwPLI in the resulting θ network was positively associated with total PTA duration but showed no relationship to misbinding errors in patients or controls. E, Mean dwpLI in the θ network in patients at baseline and follow-up. F, Mean dwpLI in the resulting α network showed no relationship between misbinding errors in patients or controls or to PTA duration. G, Mean dwpLI in the α network in patients at baseline and follow-up.
In the α band, there was a single robust network of hypoconnectivity in TBI patients compared with controls consisting of 6 edges and 6 nodes (p = 0.020; Fig. 10C). Connectivity within the α network was not related to binding ability in TBI patients or Controls (TBI: R = −0.23, p = 0.39; Controls: R = −0.08, p = 0.72; Fig. 10F). There was no difference in the mean connectivity within the α network between PTA+ and PTA– patients (t(11.98) = 0.38, p = 0.71). The mean dwPLI in the α network did not show a relationship with PTA duration (R = −0.08, p = 0.760; Fig. 10F). At follow-up, whole-brain connectivity in the δ and θ bands decreased and increased in the α band (Fig. 10A), and there was a general trend toward decreases in connectivity across the θ network and increases in connectivity across the α network, although these were not statistically significant (θ: Fig. 10E, t(9) = 1.81, p = 0.103; α: Fig. 10G, t(9) = −1.90, p = 0.09).
Mean connectivity within the α network showed a negative correlation to the DAR in controls (R = −0.48, p = 0.028) but not in patients (R = −0.13, p = 0.62). There was no relationship in either group between connectivity in the θ network and DAR (Patients: R = 0.24, p = 0.36; Controls: R = −0.41, p = 0.067).
θ-γ cross-frequency coupling is not altered in acute TBI
PAC, specifically frontal θ phase to parietal and temporal γ amplitude, was quantified using the MI. Contrary to our hypothesis, patients did not show increased PAC for either frontal to parietal (t(35.97) = −1.413, p = 0.1662) or frontal to temporal (t(34.42) = −0.634, p = 0.5303) modulation (Fig. 11A). There were no significant longitudinal changes in patients returning for follow-up in the mean MI for either frontal-parietal (t(9) = 1.567, p = 0.1516) or frontal-temporal (t(9) = 0.276, p = 0.7891) channels (Fig. 11B). Separating the groups by PTA status did not produce an effect of group: there was no significant group effect on frontal-parietal θ-γ PAC (F(2,35) = 1.83, p = 0.175) or frontal-temporal θ-γ PAC (F(2,35) = 2.46, p = 0.100) when comparing PTA+, PTA–, and healthy controls.
Frontal θ phase to parietal and temporal γ amplitude coupling. A, Average MI between frontal and parietal channels (left) and frontal and temporal channels (right) in all TBI patients (PAT) and controls at baseline. B, Longitudinal changes in PAC between frontal-parietal (left) and frontal-temporal (right) channels in patients at follow-up. Error bars indicate standard error of the mean.
Discussion
PTA is a common consequence of TBI characterized by profound but transient cognitive disturbances. Here, we show that PTA is associated with marked impairment in binding of information in working memory that resolves on emergence from PTA. This behavioral abnormality is associated with a shift to low-frequency oscillations on EEG, quantified by increased DAR of EEG power, which is specifically seen in PTA and not TBI patients without PTA. Increased DAR correlated with misbinding failures, potentially indicating a causal role in the disruption of working memory function. Connectivity is nonspecifically increased in the θ and decreased in the α following TBI. The results suggest that abnormal low-frequency oscillations, seen in many disease states, may disrupt the precise synchronization of neural oscillations necessary for working memory function.
We used a precision spatial working memory task to assess binding of object-location information. Participants were required to accurately encode both the identity of an object and its spatial location. Separate identification and location phases probed identity and spatial information. Importantly, PTA patients had an accurate recognition memory for object identity in most trials. However, there were high levels of misbinding (i.e., the correct object was chosen but moved to the incorrect location). This error was not random: PTA patients systematically chose the location of the nontarget object from a free-recall space, indicating that both spatial information and object identity had been independently encoded, but that the binding of this information was impaired.
Assessing misbinding relies on an arbitrarily defined threshold to decide whether a response is within the boundaries of a specific location and does not fully describe the response distribution across trials. It is informative to display responses in a transformed space as this can illustrate the presence of misbinding errors and the absence of random responses. We show clear clusters around target and nontarget items demonstrating that spatial errors in PTA are nonrandom. This shows subjects are engaged, encode object identity and location, but show binding impairments. This approach is also informative when applied to individual cases. The case studies highlight that transient increases in DAR and misbinding observed at the group level in PTA+ patients may also be informative at the single-subject level. A precision working memory task of this type could therefore be used to assess PTA, with misbinding measures offering a sensitive measure of how a PTA state changes over time.
While a key strength of the task design is the precision aspect of the spatial response, it is also possible that misbinding likelihood may be influenced by the proximity between the target and nontarget items; for example, when positioned closer, a misbinding error may become more likely. To mitigate this risk, we kept a minimum distance of 600 pixels between target and nontarget items. There was no correlation between a subject's misbinding rate and the mean distance between target and nontarget (R = 0.0352, p = 0.82). Furthermore, there was no difference in the mean distance between target and nontarget across groups (F(2,24) = 0.05, p = 0.96). Hence, the group-level results we present could not be explained by any differences in proximity between target and nontarget items.
Our PTA patients have widespread cognitive deficits compared with healthy controls. Some of these are produced by having an acute TBI, and are not specific to a period of PTA. Other cognitive deficits are more pronounced in PTA+ compared with PTA–. Tasks requiring working memory integration, such as paired associates learning and object-search strategy, demonstrated differences between patient groups, but those without the binding component (e.g., spatial-span) were less discriminatory. PTA+ also showed impaired attentional processing that may underpin some of their mnemonic deficits, although this was not related to misbinding (R = 0.09, p = 0.75). Attentional deficits would have made “random” responses in their precision recall of location more likely, but instead these were clustered around target and nontarget locations. It is not therefore the case that PTA+ are globally more impaired than PTA–, but rather there are transient and specific deficits, including binding ability, associated with PTA.
Resting-state EEG was used to identify electrophysiological abnormalities in acute TBI patients. In TBI, increased lower-frequency power has previously been associated with long-term neuropsychological and functional outcomes (Leon-Carrion et al., 2009; Robb Swan et al., 2015). EEG slowing is often viewed as a nonspecific sign of cerebral dysfunction. We show that PTA patients show decreased α power, increased δ power, and significantly higher DAR than TBI patients without PTA and controls. This abnormality correlated with misbinding and normalized at follow-up. Increased DAR may therefore be a sensitive electrophysiological marker of PTA that is closely related to the key cognitive impairments seen in this state.
Low-frequency oscillations are observed in other disease states (Howells et al., 2018; Cassidy et al., 2020; Jafari et al., 2020). Alzheimer's disease patients show impaired binding (Della Sala et al., 2012) and a shift toward increased low-frequency oscillations characterized by δ synchronization and α desynchronization (Benwell et al., 2020). Additionally, increased low-frequency oscillations are associated with reduced awareness (Howells et al., 2018). We previously used voltage imaging in rodents to show that anesthesia is associated with low-frequency hypersynchrony and reduced cortical communication (Scott et al., 2014; Fagerholm et al., 2016). Increased low-frequency oscillations are therefore common to states of dementia, disordered consciousness, and PTA, and may provide a common electrophysiological mechanism by which the cortical dynamics necessary for higher cognitive functions, including working memory and awareness, are disturbed. PTA patients provide a unique way to study transient changes in the relationship between electrophysiological abnormalities and working memory binding. Here, we show that DAR normalizes in conjunction with emergence from PTA and normalization of working memory binding impairments. Together, our results suggest that the presence of increased low-frequency oscillations may disrupt electrophysiological interactions necessary for successful integration of object-location memory, and this mechanism may be relevant across a range of disease states. This study is observational; and future interventional work, such as animal modeling or brain stimulation studies, would be required to establish a causal relationship between a shift toward more dominant lower frequencies and working memory failures.
To investigate how low-frequency oscillations might disrupt working memory, we explored the effect of TBI on cortical connectivity. Coordinated activity across the brain is important for communication. We calculated the phase-lag index (dwPLI) to quantify channel-wise correlations. Acute TBI patients showed θ hyperconnectivity and α hypoconnectivity, but these changes were not specific to PTA. Additionally, although θ connectivity between frontal and temporal-parietal regions has been shown to increase with working memory load and executive control (Fell and Axmacher, 2011), it did not correlate with binding performance following TBI. However, θ connectivity increased with PTA duration, suggesting it might relate to injury severity (McMillan, 2015). We additionally explored PAC in the θ/γ bands, as this is associated with working memory binding (Daume et al., 2017; Köster et al., 2018). This was not abnormal following TBI and did not relate to binding performance. It is possible that we were simply not statistically powered to detect an effect. An alternative explanation for this is that changes in PAC might only be seen during working memory performance. As we investigated resting-state EEG, we cannot explore the relationship between neural oscillatory dynamics and distinct working memory stages. Future work could usefully explore the relationship between EEG abnormalities and working memory impairment during event-related analyses of task performance.
Our patients were heterogeneous in terms of TBI pathophysiology. The pattern of TBI seen on neuroimaging varied but did not explain the presence or absence of PTA. Importantly, all patients in both PTA+ and PTA– groups sustained a moderate to severe TBI, and it is not the case that the PTA+ group were merely more severely injured. Nevertheless, future work in a larger sample might investigate the relationship between distinct types of brain injury and the EEG/behavioral abnormalities we have observed. Additionally, the PTA group were older than controls. This is unlikely to explain our results. During healthy aging, there is a linear decrease of slow frequency resting-state activity (Vlahou et al., 2014). This suggests that, if age were influencing the results, PTA+ patients would be expected to show lower δ power than controls. Indeed, the opposite was observed, and PTA+ showed significantly greater δ power. A difference in age between the groups should therefore not alter the interpretation of a shift toward lower-frequency oscillations during PTA. Finally, our PTA+ patients had fewer education years than PTA– patients. However, there was no relationship between education years and proportion of misbinding errors in TBI patients, and we do not therefore think that this difference can explain the behavioral results observed.
In conclusion, acute TBI patients experiencing PTA show a marked impairment of associative working memory. We quantified this using misbinding errors and showed how this relates to electrophysiological changes on EEG. Our results demonstrate a clear relationship between a shift to pathological oscillatory slowing and transient working memory impairment in PTA, which is informative in understanding the profound but transient effects of TBI on higher cognitive functions.
Footnotes
This work was supported by an Academy of Medical Sciences Starter Grant for Clinical Lecturers (awarded to NG) and by an equipment grant from the Department of Medicine, Imperial College London. E.-J.M. was supported by the UK Dementia Research Institute Care Research and Technology Center. G.S. was supported by National Institute for Health and Care Research (NIHR). D.J.S. was supported by UK Dementia Research Institute Care Research and Technology Center, NIHR Professorship NIHR-RP-011-048, NIHR Clinical Research Facility, and Biomedical Research Center at Imperial College Healthcare NHS Trust & the Royal British Legion Center for Blast Injury Studies. We thank all the participants who took part in this study; the staff at the Major Trauma Ward and the Neurosurgery, Emergency and Trauma Research Team, St Mary's Hospital, and Imperial NHS Healthcare Trust, London for assistance in patient screening.
The authors declare no competing financial interests.
- Correspondence should be addressed to Gregory Scott at gregory.scott99{at}imperial.ac.uk
