Identifying the neural and behavioral correlates of cognitive reserve

Identifying the neural and behavioral correlates of cognitive

reserve

Laura Stolp, 0413321 Supervisor: Dr. Daniel Bor Co-assessor: prof. dr. Lucassen Examiner: prof. dr. Lucassen

MSc (research) Brain and Cognitive Sciences Research Project 1, 26EC

Start-date: 24/2/2020, End-date: 11/8/2020 University of Amsterdam

Abstract

Objective: Dementia is a debilitating neurodegenerative syndrome which causes a deterioration in a variety of cognitive functions, such as memory, executive functioning, language and learning capacity. Interestingly, there is not always a one-to-one relationship between the severity of the neuropathology and the degree of cognitive decline, which may be explained by a compensatory mechanism of cognitive reserve. The objective of the current study is to assess whether behavioral and neural correlates of cognitive reserve can be found

Methods: We analyzed previously acquired data from 38 elderly participants. We assessed task performance on a sustained attention task under two levels of arousal: alert and drowsy. Here, cognitive reserve was operationally defined as the degree to which task performance was affected by the

neurocognitive strain of drowsiness compared to alertness. We then related our measure of cognitive reserve with IQ, educational level, performance on various cognitive tests (e.g. MoCa), structural volumetric

differences in various hippocampal subfields, and informational complexity as established by the Lempel-Ziv algorithm.

Results: We found evidence for group level effects on performance, where task performance was consistently decreased under states of drowsiness. This was the case for all our five measures of cognitive reserve: median reaction time, mean reaction time, variability of reaction time, false alarms and misses. However, IQ, educational level or other psychological measures failed to show any consistent relationship with our cognitive reserve measures. Interestingly, we did find evidence for functional and structural neural correlates of cognitive reserve. We successfully related Lempel-Ziv for single channels (LZs) to most of our cognitive reserve measures. It turned out that LZs increased under drowsiness, and that a higher difference in LZs between alertness and drowsiness correlated with better performance under the neurocognitive strain of drowsiness. Furthermore, several hippocampal subfields correlated with various cognitive reserve measures. Higher volume in certain important hippocampal subfields, such as CA1 and CA3, was related to higher cognitive reserve as indicated by a negative correlation between volume and performance discrepancy.

Conclusion: The findings of our study show that there is a relationship between an increase of informational complexity under neurocognitive strain and task performance. It could be that the increase in informational complexity under drowsiness, as indicated by LZs, indicates some sort of compensatory mechanism, where individuals that have a more pronounced increase are better able to maintain their performance under neurocognitive strain, which is an indication for higher cognitive reserve. Furthermore, it might be that certain hippocampal subfields are involved in maintaining performance under neurocognitive strain. However, more research is needed to further examine what is happening on a neural level. An interesting, additional analysis that might provide more insight concerns network analysis, particularly the examination of small world properties. Furthermore, it is important to further investigate the functions of the various hippocampal subfields, their connectivity with other brain regions, and examine whether volumetric differences in other key regions for cognition may also be associated with cognitive reserve.

Introduction

Dementia is a debilitating neurodegenerative syndrome which causes a deterioration in a variety of cognitive functions, such as memory, executive functioning, language and learning capacity. Consequently, patients experience great difficulties in performing everyday activities. The decline in cognitive functioning often coincides with emotional regulation issues, difficulties in social interactions and other behavioral impairments. As dementia progresses, symptoms increasingly get worse and the illness eventually leads to the death of the patient (Duthey, 2013). Currently, an estimated amount of 50 million people suffer from

dementia worldwide and yearly there are around 10 million new diagnoses made. The most common form of dementia is Alzheimer disease (AD) which accounts for approximately 60% to 70% of dementia cases (WHO, 2019). Currently, there are no treatments available that can either reverse or halt the progression of the illness. Much research has been dedicated to find reliable biomarkers that can be beneficial in the assessment of risk factors, early diagnosis and keeping track of disease progression. Furthermore, a considerable amount of effort has been put into the development of pharmacological treatments that might slow or stop the progression of the illness, reverse neurological damage or prevent the occurrence of AD all together (Duthey, 2013).

Interestingly, there is not always a one-to-one relationship between the severity of the

is less severe than what would be expected from the neurodegenerative damage (Katzman et al., 1988). In these cases there seems to be a compensatory mechanism in play, also known as cognitive reserve (Stern, 2012). Gaining more insight into this trait could be of considerable clinical significance. Identifying neural correlates of cognitive reserve could help developing therapeutic interventions (e.g. pharmacological treatment) to slow down the rate of cognitive decline in dementia and improve quality of life.

A better understanding of cognitive reserve might also be helpful in diagnosing patients earlier. In recent years, the majority of AD clinical trials have failed. This failure may be explained by the fact that serious neurological damage has already been done once AD symptoms emerge and this damage may be irreversible (Cummings et al., 2014). When the disease is caught at an earlier stage, treatment with currently available medication may be more successful and prevent further damage. Moreover, since cognitive reserve often masks the seriousness of the condition it is important to be able to more adequately diagnose high cognitive reserve patients. In addition, it must be mentioned that the pattern of disease progression is markedly different in high cognitive reserve patients compared to low cognitive reserve patients. Patients with high cognitive reserve are able to maintain their cognitive functioning for a longer period of time. However, once symptoms appear, their cognitive decline accelerates and is often more rapid than the deterioration of cognitive functioning in patients with low cognitive reserve (Stern, 2012). Finally, it is of crucial importance to take cognitive reserve into account in clinical trials, since these trials often assess differences in the rate of decline in patients to determine the effectiveness of a given drug treatment compared to a placebo. If the level of cognitive reserve is not accounted for, cognitive reserve could be a confounding factor in clinical research which might lead to incorrect conclusions about the (in)effectiveness of a given treatment. Therefore, a reliable way of assessing the level of cognitive reserve in a patient would be tremendously helpful.

Previous research suggests that individuals with higher levels of education, more participation in leisure activities, social interactions, and more stimulating cognitive experiences in general have a lower risk of developing dementia, and have a higher level of cognitive reserve (Valenzuela et al., 2008; Stern et al., 1994; Scarmeas et al., 2001). Furthermore, it seems that ‘healthy’ ageing in itself is a cause for neurocognitive strain and that individuals with higher cognitive reserve are less affected by the process of normal ageing in terms of cognitive performance (Whalley et al., 2004). Educational level in particular seems to be highly related to cognitive reserve (Stern et al., 1994). Since there is a strong relation between intelligence and educational level, IQ might be a modulating factor for cognitive reserve (Stern, 2012). According to the Cattell-Horn theory of intelligence (Horn & Cattell, 1966) there are two types of intelligence. Fluid intelligence is the ability to reason and solve novel problems in a flexible way and crystallized intelligence refers to the accumulation of knowledge and the capacity to learn from past experiences. Fluid intelligence generally decreases with age, while crystallized intelligence is thought to increase throughout one’s lifetime (Horn & Cattell, 1967). Since there are differences in how fluid intelligence and crystallized intelligence are associated with age, they may also relate differently to cognitive reserve. Therefore, it is important to assess how both types of intelligence relate to cognitive reserve.

Although it might be possible to infer the level of cognitive reserve from behavioral factors such as educational level, a better strategy would be to find a way to measure cognitive reserve directly. One suitable intervention could be to manipulate levels of drowsiness and alertness and relate this to task performance. Drowsiness is a common, reversible form of neurocognitive strain that is easy to manipulate in experimental settings (Nilsson et al., 2005). Generally, people are observed to be cognitively impaired when drowsy (Durmer & Dinges, 2005) and the degree to which this cognitive impairment manifests may relate to individual

differences in cognitive reserve. Furthermore, there is a strong relation between Alzheimer symptoms and sleep disturbances (Ju et al., 2014), which makes drowsiness a particularly useful measurement of

neurocognitive strain in the context of AD research. AD patients often show significant deficits in their ability to sustain attention, which can already be observed in the early stages of the illness (Huntley et al., 2016). Therefore, a sustained attention task might be an ideal candidate when measuring cognitive reserve by assessing the effect of neurocognitive strain on task performance.

As mentioned before, there is some evidence suggesting that lifestyle factors (e.g. education, leisure activity) and IQ can provide some protective function against the neurocognitive strain of ageing or

neurological damage due to diseases such as AD (Valenzuela et al., 2008; Scarmeas et al., 2001; Whalley et al., 2004). However, the underlying neural mechanisms of this cognitive reserve are still unknown. An outstanding question concerning the neural correlates of cognitive reserve is whether it is possible to define reliable anatomical biomarkers: specific brain regions that are directly related to cognitive reserve. One candidate brain area that might be related to cognitive reserve is the hippocampus. The hippocampus is an important neural area that plays a vital part in cognition and affect (de Wael et al., 2018), and hippocampal atrophy is considered to be a major biomarker for early AD (Valenzuela et al., 2018; Zhao et al., 2019). The importance of

the hippocampus for cognition might be related to its function as a cortical hub. There are many connections between the hippocampus and anterior/posterior cortical regions, which may indicate that it is involved in a more global network activity (de Wael et al., 2018).

A longitudinal study of Valenzuela et al. (2008) found an association between significantly lower rates of hippocampal atrophy and complex cognitive stimulation (e.g. higher educational level, more complex occupations) throughout one’s lifetime. These results echo findings from population-based research that show complex mental stimulation to be associated with lower incidence of AD (Valenzuela et al., 2008). According to Valenzuela et al. (2008), complex mental activity might have neuroprotective effects, which could be related to various underlying biological mechanisms. For instance, there is an association between the upregulation of brain-derived neurotrophic factor (BDNF), which is an important neurotrophic hormone, and environmental enrichment. This effect is particularly apparent in the hippocampus (Valenzuela et al., 2007). Furthermore, there is some evidence that complex mental stimulation might have restorative effects (e.g. neurogenesis) in the hippocampus (Kempermann, 2006). The hippocampus consists of multiple cyto-architectonically distinct subfields, which vary in the anatomical connections to other brain areas (de Wael et al., 2018). Furthermore, these distinct hippocampal subfields may differ in the degree to which they are affected by AD. For instance, a study of Zhao et al. (2019) showed that volume reduction in the presubiculum, subiculum and CA1 to have the strongest relationship with memory performance decline due to disease progression, while other hippocampal subfields were much less affected (Zhao et al., 2019). In addition, there is also a volumetric effect between certain hippocampal subfields and educational level. Specifically, lower hippocampal atrophy in subfield CA3 is shown to be related to higher educational level (Jiang et al., 2019). Since the hippocampus is a key structure in Alzheimer pathology and may be related to individual differences in cognitive reserve, we aim to investigate the hippocampus and its potential association with cognitive reserve in more detail. In the study of Valenzuela et al. (2008), they looked at whole hippocampal volume. Since different subfields may relate differently to AD and also to educational level, there may also be differences in how these subfields relate to cognitive reserve. Therefore, we will investigate the various hippocampal subfields and the potential relationship with cognitive reserve.

If specific hippocampal subfields can be successfully related to cognitive reserve, a remaining question would be why such regions provides individuals with the ability to maintain cognitive performance under neurocognitive strain. A possible explanation could be that the hippocampus plays a vital part in general information processing in the brain, or is a part of a larger network that is related to information processing. An emerging line of research concerns the quantification of informational complexity in the brain and the relationship with different levels of (un)consciousness (Schartner et al., 2015). One previously-validated technique for measuring signal complexity is the Lempel-Ziv algorithm, which is often used to quantify the compressibility of various biological signals, including the EEG signal (Hu et al., 2006). Higher signal

compressibility signifies lower informational complexity and vice versa. Studies have shown that the degree of informational complexity corresponds to different levels of consciousness, where high signal complexity can be found when individuals are awake and aware and low signal complexity can, for instance, be found in

anesthetized individuals, slow-wave sleep and AD patients (Casali et al., 2013; Abasolo et al., 2006; Schartner et al., 2015).

The present study aims to examine cognitive reserve and its relation with various behavioral and neural correlates through extensive analysis of previously acquired data from elderly participants. It concerns 38 elderly participants who were previously scanned (i.e. structural MRI), previously performed various cognitive tests (e.g. IQ tests, digit span) and previously participated in an experiment in which they carried out a sustained attention task while drowsiness was induced. The level of cognitive reserve will be determined by assessing how much task performance is affected by the (reversible) neurocognitive strain due to high levels of drowsiness. We expect cognitive reserve to be positively correlated with IQ, years of education and

performance on several other cognitive tests. We also aim to assess whether the discrepancy between fluid IQ and estimated premorbid IQ is related to cognitive reserve, since a higher discrepancy might indicate a higher loss of cognitive function, including the ability to preserve cognition under strain, compared to their previous level of intelligence. Furthermore, we aim to investigate the structural and functional neural differences between participants with high cognitive reserve and low cognitive reserve. Structurally, we expect to find higher volume in various hippocampal subfields in individuals with high cognitive reserve. We also aim to investigate how volume in hippocampal subfields relates to intelligence, years of education, and performance on the Montreal Cognitive Assessment (MoCa). The MoCa is a brief screening tool to assess cognitive functioning and detect mild cognitive impairment (Nasreddine, 2005). Functionally, we expect to see a decrease in EEG signal complexity, as measured with the Lempel-Ziv algorithm, when drowsiness increases and expect this effect to be less pronounced in participants with higher cognitive reserve. We further aim to

examine how intelligence, educational level and performance on the MoCa test are related to informational complexity.

Method

Participants

The current study concerns an extensive analysis of previously-acquired data from 38 elderly participants (16 women, 22 men; age-range: 60 – 84, mean: 73.16, SD: 5.26) of whom structural MRI scans were obtained. Self-reported years of education ranged from 10 years to 39 years (mean: 18.01, SD: 4.68). The current project is a continuation of a previous project in which cognitive reserve was studied in healthy old participants by inducing drowsiness while they performed an attentional task (Swartz, 2019). All cognitive tests and experimental data were previously acquired. The study was approved by the Cambridge Psychological Research Ethics Committee of the University of Cambridge.

Cognitive tests Cattell

Fluid intelligence of the participants is determined by their performance on the Cattell Culture Fair Intelligence Test, where higher scores indicate higher fluid intelligence (scale 2A, Cattell, 1973). The Cattell test is designed to minimize the potential influence of educational or cultural biases. The focus of this test is on non-verbal intelligence as all items are visual in nature. In one section of the test, the participant is presented with a series of multiple choice questions consisting of visual patterns and is asked each time to choose which object would best complete the pattern (Cattell, 1973).

Fig. 1. An example of an item of the Cattell test. Extracted from researchgate.net. Link:

https://www.researchgate.net/figure/An-example-non-verbal-item-from-Cattells-Culture-Fair-Free-Test_fig2_249934418. The left shape presents a pattern where the dot is inside the circle, but outside of the square. Subjects have to find the same pattern amongst the five presented options. The correct answer here is the third option, since this is the only option where the dot can be put inside the circle, while similarly being outside of the square.

NART

The National Adult Reading Test (NART) is a method to estimate premorbid IQ, specifically WAIS IQ. This test is often used to estimate premorbid levels of cognitive functioning in patients with

neurodegenerative disorders, such as dementia. A discrepancy between premorbid IQ, as estimated by the NART, and current cognitive functioning, as determined by other cognitive tests, is then used as evidence of the presence of intellectual impairment in a given patient (Nelson & Willison, 1991). Since healthy ageing in itself is also a cause for neurocognitive strain (Whalley et al., 2004), a discrepancy between estimated premorbid IQ and current levels of cognitive functioning can also be expected to be found in healthy elderly populations. However, this discrepancy can be expected to be much smaller in healthy elderly populations compared to those who suffer from a neurodegenerative disorder. The NART consists of 50 words that the test-takers are instructed to read aloud. The words are all ‘irregular’ in context of general pronunciation rules. See appendix A for the complete list of words. While the words are read out loud, the number of

be approximated (Nelson & Willison, 1991). In the current study, estimated premorbid IQ (i.e. full-scale WAIS IQ) of the participants is determined by their performance on the NART. See appendix B for more details about scoring method. Here, a discrepancy between current level of cognitive functioning and premorbid cognitive functioning is operationalized as the discrepancy between the performance on the NART and the performance on the Cattell test, where the Cattell score is subtracted from the NART score.

Montreal Cognitive Assessment

The Montreal Cognitive Assessment (MoCa) is a brief 10-minute screening tool to assess cognitive functioning and detect mild cognitive impairment. The MoCa is often used by first-line clinicians in the assessment of patients with mild cognitive complaints. The test consists of several parts to assess various cognitive domains, including memory, executive functioning, attention and language (Nasreddine, 2005). The maximum score that can be achieved is 30 and a score of 26 or above is considered to be normal. See appendix C for more details about the items and scoring method. In the current study, an insufficient score on the test (26<) is interpreted as that these participants are at risk for developing Mild Cognitive Impairment (MCI), especially if an insufficient score on the MoCa test coincides with low performance on other cognitive tests. There were six participants in total who scored 25 or below on the MoCa test.

Memory Functioning Questionnaire

The Memory Functioning Questionnaire (MFQ) is developed to assess the subjective perception of everyday memory functioning. The questionnaire consists of four subscales, namely frequency and seriousness of forgetting, retrospective functioning and usage of mnemonics. The items are scored on a 1 to 7 rating scale, where 7 indicates a very good memory functioning and 1 indicates severe memory problems (Gilewski et al., 1990). See appendix D for more details about the MFQ. In the current study, only the subscales concerning frequency of forgetting and seriousness of forgetting are used.

Trail making task

The trail making test is a neuropsychological test consisting of two parts. In part A, the participant is presented with a sheet of 25 encircled numbers and is instructed to connect these in numerical order using a pencil. The participant is instructed to do this as fast as possible. Performance on this test is determined by the number of seconds it takes from start to finish. The maximum time given to the participant is 300 seconds. In part B, the participant is asked to connect 25 encircled numbers and letters in order. For instance, “1” connects to “A”, “2” connects to “B” and so forth. Performance is determined the same as in part A, but the maximum time given to the participant in part B is 500 seconds. Performance on part A of the test is related to motor speed skills and visual search, while performance on part B is mostly related to executive functioning, particularly cognitive flexibility (Bowie & Harvey, 2006). See appendix E for a more detailed description and example sheets of the test. In the current study, we are mainly interested in part B since this is related to executive functioning. Specifically, we look at performance discrepancies between part A and part B, by subtracting the score on part A from the score on part B. The reason for this is that Part B h also measures visual search and motor speed skills, in addition to executive functioning. By subtracting A from B we only look at executive functioning.

Digit span

The digit span test is a cognitive test that measures memory span and has been shown to highly correlate with intelligence. Participants are presented with a series of sequences of various lengths of randomly ordered numbers and are asked to repeat the number sequences either in the same order

(forwards) or in reversed order (backwards). Generally, repeating the numbers in reversed order is associated with more complex cognitive processes (i.e. executive functioning) than repeating the numbers in the same order as presented (Schofield & Ashman, 1986). There are more ways of scoring performance on the test. Often, performance is determined by highest length of the sequences that the participant was able to repeat correctly. However, an alternative scoring method is to keep track of the number of sequences that the participant is able to repeat correctly regardless of length (Blackburn & Benton, 1957). In the current study, the latter scoring method is used, where a higher number of sequences correctly repeated indicates a better

performance and vice versa. Only the scores on the digit span backwards task are analyzed for the purpose of the current study.

Experimental procedure

In a previous study, the participants took part in an experiment that was designed to assess individual differences in cognitive reserve. For this aim the participants performed a sustained attention task, during which levels of arousal were manipulated. The experimental procedure took place at the EEG lab of the Consciousness and Cognition research group of the University of Cambridge. Prior to taking part in the study, the participants all signed an informed consent form. On session days, the participants were asked to refrain from consuming any caffeinated beverages. The study was approved by the Cambridge Psychological Research Ethics Committee of the University of Cambridge.

Electroencephalogram (EEG) data was recorded using a 129-channel Philips EEG scanner. While their EEG was being recorded, participants performed a series of tasks. In the first phase of the experiment, participants sat in an upright position and were alert. They performed a visual memory precision task, which is an episodic memory task (30 min). The performance of the participants on the memory task is not analyzed in the current study. In the second phase of the experiment, the participants performed a sustained attention task, namely the Sustained Attention to Response Task (SART). For the aim of this study, an auditory version of the SART was used. In the auditory version of the SART participants are instructed to respond to certain auditory stimulus types and inhibit responses in reaction to other (infrequently presented) auditory stimulus types (Seli et al., 2012).

Before the start of the SART, the lights in the room were dimmed and participants sat in a reclining chair. They were offered pillows and blankets to make themselves comfortable. The following instructions were given to the participants: a) to close their eyes and keep them closed while performing the task, b) to minimize any form of movement, c) that it is no problem if they made any mistakes and in case of a mistake to just continue, d) that they might become drowsy but they do not have to worry about dozing off. However, in case that the participant actually fell asleep, signified by three consecutive non-responsive trials, there was an audio recording played to wake them up. If the audio recording did not succeed at waking the participant, the participant was woken up manually by the experimenter. The purpose of this setup was to induce states of drowsiness in the participants. The expectation was that participants would be relatively alert in the beginning of performing the SART but would become increasingly drowsy during the experiment. In this way,

performance in alert states and performance in drowsy states could be compared.

The participants first took part in a resting state block (5 min) before the onset of the SART. The auditory SART consisted of the presentation of numbers in the range from 0 to 9, in a randomized ordered. The stimuli were extracted from an online database with vocal recordings (Sayanng, 2009). Participants were instructed to respond to the stimuli by pressing a button, but were told to withhold a response when they heard the number ‘3’ (i.e. target stimulus). The randomization of the stimuli was determined by certain randomization parameters. For instance, the randomization settings made sure that the target stimuli was not presented more than three times in consecutive trials and the target stimulus was presented in 10% of the total number of trials. Before starting the task, the participants took part in one practice block consisting of 30 trials (2.5 min). The response window was fixed to 1100ms within the practice block. The duration of the response window in the task was determined by the average reaction time of the participants in the practice block (i.e. mean RT + 250 ms). The end of the response window was always signified by a ‘beep’ sound being played (SoundJay, 2019), after which there was an inter-trial time window for a randomized duration of 2 seconds to 5 seconds. See figure 2 for an overview of the trial structure. The task consisted of 700 trials in total and was split into two blocks of 350 trials each. There was a break between these two block of around 3 minutes in length. The entire experimental procedure lasted for approximately 3.5 hours.

Fig. 2 Overview of trial sequence structure.

Preprocessing of EEG data

We made use of already preprocessed EEG data. The most important steps of the preprocessing procedure are briefly discussed below.

Prior to importing the EEG data into EEGLAB, the data was down sampled to 250 Hz. Next, the open-source MATLAB toolbox Automagic was used, which is a standardized pipeline for detecting and removing artifacts from EEG data (Pedroni et al., 2019).After all the artifacts were removed from the data, the level of arousal of all the 4 second epochs was determined. For this aim, a previously-validated machine learning algorithm was used that was designed to automatically determine microstate variations in alertness and drowsiness from the EEG signal (Jagannathan et al., 2018). The algorithm is based on the Hori-scale, which is a scale that divides the sleep onset process into nine stages, from wakefulness (stage 1 and 2) to the onset of N2 sleep (stage 9) (Iber et al., 2007). An adapted Hori 10 was added based on the occurrence of K-complexes, which typically occur in the N2 stage. Furthermore, the microstate variations algorithm was specifically developed for eyes-closed experiments and is suitable to use on EEG data divided into short epochs (e.g 4s) (Jagannathan et al., 2018). Level of arousal is computed in the following way: (1) Cross-frequency coherence features and predictor frequency variance are computed, (2) These factors are used to distinguish between alert and drowsy states. Alert state is determined by the presence of Alpha wave trains (> 50%) (Hori: 1 – 2). The transition from alert state to drowsiness typically coincides with a reduction of Alpha activity (Hori: 3 = Alpha < 50%) and an increase in theta waves (Hori: 5). Drowsiness is determined by the presence of these transitions in Alpha and Theta activity, (3) EEG data classified as drowsy is further analyzed to determine whether graphoelements (i.e. vertex, spindles, K-complexes) occur, (4) Determining whether detected spindles are truly spindles, (5) EEG data is classified as ‘drowsy severe’ in the case of graphoelements, otherwise it is classified as ‘drowsy mild’ (Jagannathan et al., 2018).

In the current work, when using the microstate variations algorithm, there were generally very few trials classified as severely drowsy and some participants even had none of such trials. Therefore, all trials that were classified as severely drowsy were merged together with trials classified as mildly drowsy. As a result, the levels of arousal were divided in two categories: ‘alert’ and ‘drowsy’. Within- and between-participant effects were analyzed for these two levels of arousal.

Statistical analysis of behavioral data and psychological tests

Statistical analysis was performed using Matlab (version 2019a). To determine performance on the SART, we analyzed the median, mean and variability of reaction times and the proportion of false alarms and

Start of trial: presentation of stimulus (e.g. "6") Response window (mean RT practice block + 250 ms) "Beep" sound, which indicates

the end of the response window Inter-trial time between: 2 to 5 seconds End of trial

misses of response errors of each participant. To determine the variability of reaction times we computed the coefficient of variation (CV), which is the standard deviation divided by the mean. Proportion of false alarms was defined as the sum of all target trials with a response divided by the sum of all target trials. Proportion of misses were defined as the sum of all non-target trials without a response or a response made after the end of the response window, divided by the sum of all non-target trials. The following exclusion criteria were used: (1) Target trials that immediately proceeded 3 consecutive non-target trials without a response were excluded from further analysis. This was done to ensure that seemingly correct response omissions on target trials were not due to the participant being asleep, (2) Participants with fewer than 50 trials in either an alert or drowsy state were excluded from further analysis of reaction time data, (3) Participants with fewer than 10 target trials in either an alert or drowsy state were excluded from further analysis of response error data.

In the current study, level of cognitive reserve is operationalized as the difference between

performance in drowsy states and alert states. A larger discrepancy in performance when comparing these two levels of arousal is interpreted as an indication for lower levels of cognitive reserve and vice versa. Specifically, we computed the difference of each measure (i.e. mean, median, CV, errors) between the two levels of arousal. For the reaction time measures, we determined the performance discrepancy as: (a) mean reaction time in a drowsy state divided by mean reaction time in an alert state, (b) median reaction time in a drowsy state divided by median reaction time in an alert state, and (c) CV of reaction times in a drowsy state divided by CV of reaction times in an alert state. For the false alarms and misses, we determined the performance discrepancy as: (a) false alarms in a drowsy state subtracted by false alarms in an alert state, (b) misses in a drowsy state subtracted by misses in an alert state.

We examined group-level performance differences between alert and drowsy states using paired sample t-tests. We also examined how our measures of cognitive reserve are related to cognitive performance and educational level. For this aim, we performed Spearman's rank correlation coefficient between the performance discrepancy on the SART for each measure (median RT, mean RT, CV RT, false alarms, misses) with previously-described cognitive tests, discrepancy between current IQ and premorbid IQ, and self-reported years of education.

Lempel-Ziv complexity

Lempel-Ziv is a compression algorithm that quantifies complexity by determining the amount of distinct activity patterns in the data and is inherently a measure of signal diversity (Lempel & Ziv, 1976). For computation of Lempel-Ziv, the time series data first has to be binarized (See fig 3 for a schematic

representation). This binarization is accomplished by using a given threshold. In this case, the threshold was determined by the Hilbert transform’s instantaneous amplitude, which is taken as the absolute number of the analytic time series signal of a given EEG channel. The segment of data is then converted to a binary matrix, with columns representing the observations and rows representing the channels. To determine signal diversity, the binary matrix is first represented (per observation) as one long binary sequence. Next, the compressibility is determined by using a binary ‘dictionary’ and assessing the number of binary words that are present in the sequence. The more unique binary words that are present, the lower the compressibility and therefore, the higher the Lempel-Ziv complexity (Schartner et al., 2015). In the current work, two variants of Lempel-Ziv complexity (Schartner et al., 2017) were used to analyze 4 sec segments at 250 Hz. First, the standard Lempel-Ziv complexity (LZc) was computed, which is able to capture both temporal and spatial signal diversity. Next, a single channel version of Lempel-Ziv complexity (LZs) was used to focus on only temporal signal diversity.

Fig 3. Schematic overview of the Lempel-Ziv complexity computation process (adapted from Schartner et al. (2017). Left: the analytic signal of the ith channel is determined by the absolute value of the Hilbert transform. If the analytical signal is larger than the average a(i), b(i) is 1, otherwise b(i) is zero. This manifests as a long binary string b for each channel i. Middle: all the strings are put in an n (channels) x m (observations) matrix. Right: One long sequence s is made of the binary matrix, from which LZc is computed by determining the number of binary words present in the signal. A higher amount of binary words in a sequence signifies higher informational complexity.

We examined group-level differences in LZc and LZs between alert and drowsy states using paired sample t-tests. We expected to see a decrease in EEG signal complexity, as measured with the Lempel-Ziv algorithm, when drowsiness increases. Furthermore, we examined how LZc and LZs were related to our measures of cognitive reserve, IQ, educational level and performance on the MoCa test. For this aim, we performed Spearman's rank correlation coefficient. We expected that individuals with higher cognitive reserve, higher IQ, higher educational level and higher score on the MoCa test would have a less pronounced decrease in informational complexity in the drowsy state as compared to an alert state. Participants with fewer than 50 trials in either an alert or drowsy state were excluded from further analysis. All statistical analyses were performed using Matlab (version 2019a).

MRI preprocessing and statistical analysis

We made use of previously acquired and preprocessed structural MRI data. MRI acquisition was part of a previous collaborative studies that were carried out within three years prior to the current study. MRI data were acquired using a 3-Tesla MRI system (MAGNETOM, Prisma, Siemens). Whole-brain T1-weighted images were acquired using a 3D MPRAGE sequence with an isotropic voxel size of 1x1x1 mm. In addition, a high resolution T2-weighted image of the medial temporal lobe was acquired, with a voxel size of 0.4x0.4x0.2 mm. T1-weighted images were preprocessed using Freesurfer version 6.0.0, parcellated into various subareas, using the Desikan-Killiany atlas, and hippocampal volume was extracted. Next, the Freesurfer hippocampal subfield segmentation method was used to further parcellate the hippocampus (T2-weighted) into the various hippocampal subfields, such as CA1, CA3, subiculum and hippocampal fissure (see fig. 4 for an overview of all the hippocampal subfields), and to extract the volume of all the different subfields for each hemisphere. There is one subfield missing from the data, which is the right hippocampal-amygdaloid transition region (HATA). To control for individual differences in total intracranial volume (TIV), all subregional volumes were normalized using TIV.

ai = |hilbert(xi)| if ai(t) >avg(ai), bi(t) = 1,

else bi(t) = 0 bi: ..0111000101010111010 01

..

Bi: …10101011010… . …01000111101… . …01000111101… . …11101011010… . …01111011101… . …01101110110… Bn: …01000101011… Observations m ->

s = b1(1)...bn(m) =

0110101...

LZc = [0,1,10,101,...]

Fig. 4. Extracted from Zhao et al. (2019). Overview of all the subfields of the hippocampus

We expected to find higher volume in various hippocampal subfields in individuals with high cognitive reserve. We investigated how volume in hippocampal subfields relates to intelligence, years of education, and performance on the Montreal Cognitive Assessment (MoCa). For these two aims, we used Spearman's rank correlation coefficient. For two participants there were no scans available, therefore, these participants were not included in any of the MRI analyses. All statistical analyses were performed using Matlab (version 2019a).

Results

Participants

Table 1 provides an overview of the sociodemographic characteristics and cognitive test scores of the participants.

Table 1. Cohort characteristics participants

N = 38

(♀=

16

, ♂ =

22

)

Mean (SD) Range Age 73.16 (5.26) 60 - 84 Years of education 18.01 (4.68) 10 - 39 MoCa score 27 (1.76) 23 - 30 NART-IQ 120.63 (6.02) 105 - 129 Cattell-IQ 97.18 (10.49) 79 - 122 Difference NART – Cattell 23.48 (11.74) -10 - 42 Trails corrected 40.29 (38.54) 4 - 233 Digit span backwards 7.97 (2.56) 3 - 13 MFQ frequency 5.15 (0.58) 3.57 - 6.34 MFQ seriousness 4.36 (1.11) 2.17 – 6.56

There were four participants that had an alert trial count or drowsy trial count lower than 50 and were therefore excluded from all reaction time analyses. There were 10 participants with an alert target trial count or drowsy target trial count lower than 10 and they were therefore excluded from all error data analyses. Thus, a total number of 34 participants were included in reaction time analyses and a total number of 28 participants were included in the error data analyses.

Behavioral results

Group level differences in performance

We first tested our basic assumption that drowsiness has a significant effect on performance compared to alert states (see fig. 5). For this aim we used paired sample t-tests to examine the presence or absence of these group level effects. Our results show significantly higher median reaction times during states of drowsiness compared to alertness t(33) = -3.62, p < .001 and the same is true for mean reaction times t(33) = -4.8, p < .001. Furthermore, there was also a greater variability in reaction times during states of drowsiness compared to alertness t(33) = -5.27, p < .001. Moreover, states of drowsiness also led to a significantly higher proportion of false alarms t(27) = -2.91, p = .007 and misses t(27) = -3.79, p =<.001.

Fig 5. Group level differences in performance between alert and drowsy states. On every box, the central red line signifies the median, the top and bottom edges signify the 75th _{and 25}th _{percentiles respectively, the whiskers include all data points} that are not considered to be outliers and all outliers are indicated by the red ‘+’ sign.

We plotted histograms of the performance differences between alert state and drowsy state of all five measures (fig. 6): mean reaction time, median reaction time, variability of reaction time, number of false alarms and misses. We also computed the standard deviation for all five measures. In addition, we plotted histograms of all the cognitive tests and years of education of the entire cohort (fig. 7).

Fig. 6 Distribution of performance differences between alert and drowsy states of the mean reaction times (SD = 0.1334), median reaction times (SD = 0.0681), variability of reaction times (SD = 0.25), false alarms (SD = 0.1189) and misses (SD = 0.1091).

Fig. 7 Distribution of cognitive tests and years of education of entire cohort. First row from left to right: (1) Fluid intelligence as measured by the Cattell test, (2) Estimated premorbid WAIS IQ as measured with the NART, (3) Self-reported years of education. Second row from left to right: (4) Scores on the Montreal Cognitive Assessment (MoCa), (5) Discrepancy between NART and Cattell, where Cattell score is subtracted from NART score, (6) Memory Functioning Questionnaire (MFQ), frequency of forgetting subscale, Third row from left to right: (7) MFQ, seriousness of forgetting subscale, (8) Trails corrected score, (9) Digit span backwards score.

We further looked at correlations between performance discrepancy on the SART and the cognitive tests, and years of education (Table 2). For this aim we used Spearman’s ranking correlational test. We did not find a consistent relationship between task performance and any of the cognitive tests, or years of education. We found three correlations: (1) variability of reaction time correlated positively with Cattell IQ (Spearman ρ: 0.3628, p-value: 0.035), which means that a higher fluid intelligence showed a relationship with higher increases of variability of reaction times under drowsiness, (2) discrepancy between the NART score and Cattell score correlated negatively with variability of reaction times under drowsiness (Spearman ρ: -0.4273, p-value: 0.0117), which means that a higher difference between fluid intelligence and estimated pre-morbid IQ showed a relationship with a lower increase in variability of reaction time under drowsiness, and, (3) Years of education correlated positively with number of false alarms under drowsiness (Spearman ρ: 0.3906, p-value: 0.0399), which means that a higher educational level showed a relationship with a higher increase of false alarms under drowsiness. Finally, we plotted scatterplots to illustrate the significant correlations found (see fig. 8).

Correlational analyses

Table 2. Relationship between task performance and cognitive tests

Mean RT Median RT CV RT False Alarms Misses Cattell Spearman ρ: p-value: 0.2151 0.2217 0.2144 0.2234 0.3628 0.0350 0.2796 0.1496 0.2385 0.2215 NART Spearman ρ: p-value: -0.0530 0.7657 -0.0167 0.9253 -0.1893 0.2835 0.2678 0.1683 -0.0179 0.9281 ‘NART – Cattell’ Spearman ρ:

p-value: -0.2230 0.2050 -0.1909 0.2796 -0.4273 0.0117 -0.1263 0.5218 -0.1441 0.4644 Years of

education Spearman ρ: p-value: -0.0285 0.8728 -0.0251 0.8878 -0.0638 0.7199 0.3906 0.0399 0.1292 0.5122

MoCa Spearman ρ: p-value: 0.1449 0.4134 0.0520 0.7702 0.1129 0.5251 -0.0517 0.7939 0.2821 0.1458 MFQ frequency Spearman ρ: p-value: -0.1234 0.4870 -0.0115 0.9487 -0.1296 0.4650 0.1035 0.6002 0.0266 0.8933 MFQ

seriousness Spearman ρ: p-value: -0.0961 0.5886 0.0821 0.6445 -0.3151 0.0695 -0.1156 0.5582 -0.1793 0.3612 Trails corrected Spearman ρ:

p-value: 0.1249 0.4815 0.0810 0.6487 0.0018 0.9918 0.0649 0.7428 0.2481 0.2030 Digit span

backwards Spearman ρ: p-value: -0.2523 0.1500 -0.2975 0.0875 -0.0954 0.5916 0.0737 0.7093 -0.0922 0.6409

Fig 8. Scatterplots of the three significant correlations found. Top left: Relation between Cattell score (x-axis) and variability of reaction time difference between alert and drowsy (y-axis) (Spearman ρ: 0.3628, p-value: 0.035), Top right: Relation between the difference of NART versus Cattell (x-axis), and variability of reaction time difference between alert and drowsy (y-axis) (Spearman ρ: -0.4273, p-value: 0.0117). Bottom left: Relation between educational level (x-axis) and difference in numbers of false alarms between alert and drowsy (Spearman ρ: 0.3906, p-value: 0.0399). From the figure, it is shown that the regression line does not always follow the same relational pattern as is shown from the Spearman correlational test.

Lempel-Ziv results Group level differences

We first tested our basic assumption that drowsiness has a significant effect on informational complexity compared to alert states, as measured with Lempel-Ziv complexity (see fig. 9). For this aim we used paired sample t-tests to examine the presence or absence of these group level effects. Our results show no significant effects for LZc, comparing drowsiness to alertness t(33) = 1.4382, p = 0.1598. However, for LZs we did find a significant difference between the two conditions, t(33) = -3.1641, p = 0.0033. Interestingly, LZs was increased under drowsiness compared to alertness.

Fig 9. Group level differences for Lempel-Ziv complexity and Lempel-Ziv single channel, comparing the two levels of arousal. Only Lempel-Ziv sum shows a significant difference. On every box, the central red line signifies the median, the top and bottom edges signify the 75th _{and 25}th _{percentiles respectively, the whiskers include all data points that are not} considered to be outliers and all outliers are indicated by the red ‘+’ sign.

We further looked at correlations between performance discrepancy on the SART and the Lempel-Ziv complexity (LZc) and Lempel-Ziv for single channels (LZs) (Table 3). For this aim we used Spearman’s ranking correlational test. We found various interesting results. Specifically, LZc showed a significant correlation with: (1) mean reaction time (Spearman ρ: -0.5074, p-value: <0.01), which means that a higher difference in LZc showed a relationship with higher increases of mean reaction times under drowsiness, (2) median reaction time (Spearman ρ: -0.5468, p-value: <0.01), which means that a higher difference in LZc showed a relationship with higher increases of median reaction times under drowsiness, and, (3) misses (Spearman ρ: -0.6366, p-value: <0.01), which means that a higher difference in LZc showed a relationship with higher increase of misses under drowsiness. The results for Lempel-Ziv for single channels (LZs) showed a significant correlation with: : (1) mean reaction time (Spearman ρ: -0.3696, p-value: 0.0321), which means that a higher difference in LZs showed a relationship with higher increases of mean reaction times under drowsiness, (2) median reaction time (Spearman ρ: -0.3436, p-value: 0.0472), which means that a higher difference in LZs showed a

relationship with higher increases of median reaction times under drowsiness, and, (3) false alarm (Spearman ρ: -0.3821, p-value: 0.0448), which means that a higher difference in LZs showed a relationship with higher increase of false alarms under drowsiness. Furthermore, a non-significant trend was found for misses (Spearman ρ: -0.3383, p-value: 0.0788). Finally, we plotted scatterplots to illustrate the relation between LZc, LZs and reaction time data (fig. 10) and error data (fig. 11).

Table 3. Relationship between task performance and Lempel-Ziv complexity

Lempel-Ziv Complexity Lempel-Ziv single channels Cattell Spearman ρ:

p-value: -0.0281 0.8748 -0.1592 0.3686

NART Spearman ρ:

p-value: -0.0452 0.7995 -0.2819 0.1062 ‘NART – Cattell’ Spearman ρ:

p-value: 0.0024 0.9890 0.0367 0.8367 Years of

education Spearman ρ: p-value: -0.0074 0.9669 -0.2562 0.1436

MoCa Spearman ρ: p-value: 0.0711 0.6893 0.0518 0.7709 Mean RT Spearman ρ: p-value: -0.5074 <0.01 -0.3696 0.0321 Median RT Spearman ρ: p-value: -0.5468 <0.01 -0.3436 0.0472 CV RT Spearman ρ: p-value: -0.2733 0.1177 -0.1429 0.4187 Misses Spearman ρ: p-value: -0.6366 <0.01 -0.3383 0.0788 False Alarms Spearman ρ:

p-value: -0.2365 0.2257 -0.3821 0.0448

Fig 10. Scatterplots of significant correlations between LZc, LZs and the reaction time data. Top left: Relation between Lempel-Ziv complexity (x-axis) and mean reaction time difference between alert and drowsy (y-axis) (Spearman ρ: 0.3628, p-value: 0.035), Top right: Relation between Lempel-Ziv single channels (x-axis), and mean reaction time difference between alert and drowsy (y-axis) (Spearman ρ: -0.4273, p-value: 0.0117). Bottom left: Relation between Lempel-Ziv complexity (x-axis) and median reaction time difference between alert and drowsy (y-axis) (Spearman ρ: 0.3906, p-value: 0.0399). Bottom right: Relation between Lempel-Ziv single channels (x-axis) and median reaction time difference between alert and drowsy (y-axis) (Spearman ρ: 0.3906, p-value: 0.0399).

Fig 11. Scatterplots of significant correlations between LZc, LZs and the error data. Top left: Relation between Lempel-Ziv complexity (x-axis) and misses difference between alert and drowsy (y-axis) (Spearman ρ: 0.3628, p-value: 0.035), Top right: Relation between Lempel-Ziv single channels (x-axis), and misses difference between alert and drowsy (y-axis) (Spearman ρ: -0.4273, p-value: 0.0117). Bottom left: Relation between Lempel-Ziv complexity (x-axis) and false alarm difference between alert and drowsy (y-axis) (Spearman ρ: 0.3906, p-value: 0.0399).

MRI results: hippocampal subfields

We further looked at correlations between performance discrepancy on the SART and volumetric differences in the hippocampal subfields. Table 4 shows an overview of all the significant correlations, which are highlighted in red and non-significant trends, which are highlighted in orange. As is shown in the table, there are several significant results between performance discrepancy on the SART and volume in various hippocampal subfields. Furthermore, there are also a few non-significant trends. All correlations found are negative, which means that a higher volume in these subfields corresponds to a lower performance

discrepancy between alert and drowsy, and vice versa. In addition, there are some negative correlations found between volume in some hippocampal subfields and MoCa score, and Cattell score. Which means that a lower volume in these subfields corresponds with higher IQ scores and MoCa scores. Also, for the difference between NART and Cattell and two of the hippocampal subfields, there is one positive and one negative significant correlation. In the first case, this means that higher volume corresponds with higher IQ difference, and in the second case, this means that lower volume corresponds with higher IQ differences. Finally, it must be mentioned that these results have to be regarded as provisional, since a test for multiple comparisons (e.g. Bonferroni) has not been performed yet.

Table 4. Relationship between task performance and volume of hippocampal subfields Mean

RT Median RT CV RT False Alarms Misses Cattell NART-Cattell MoCa Left-HATA Spearman ρ: p-value: _0.067 -0.329 -0.453 _0.01 -0.366 _0.057 Left hippocampal tail Spearman ρ: p-value: -0.392 _0.027 Left Hippocampal fissure Spearman ρ: p-value: -0.425 _0.015 0.483 _0.005

Left fimbria Spearman ρ:

p-value: -0.421 _0.0167 Right-CA1 Spearman ρ: p-value: -0.355 _0.047 -0.35 _0.068 -0.414 _0.018 Right-CA3 Spearman ρ: p-value: -0.37 _0.038 -0.529 _0.004 -0.402 _0.023 Right-CA4 Spearman ρ: p-value: - 0.411 _0.020 -0.494 _0.0083 Right molecular layer HP Spearman ρ: p-value: -0.37 _0.04 -0.365 _0.0568 Right GC-ML-DG Spearman ρ: p-value: -0.451 _0.0169 -0.449 _0.01 Right hippocampal tail Spearman ρ: p-value: -0.342 _0.0758 Right presubiculum Spearman ρ: p-value: -0.35 _0.0687

Discussion

In the current study, we aimed to relate individual differences in cognitive reserve to various neural correlates of cognitive functioning as established by performance on various cognitive tests, subjective rating of memory performance and educational level. Cognitive reserve was defined as the performance discrepancy (i.e. reaction time, errors) between drowsy and alert states on a sustained attention task. We found evidence for group level effects on performance, where task performance was consistently decreased under states of drowsiness. This was the case for all our five measures of cognitive reserve: median reaction time, mean reaction time, variability of reaction time, false alarms and misses. However, IQ, educational level or other psychological measures failed to show any consistent relationship with our cognitive reserve measures. Interestingly, we did find evidence for functional and structural neural correlates of cognitive reserve. We successfully related Lempel-Ziv to most of our cognitive reserve measures. Furthermore, several hippocampal subfields correlated with various cognitive reserve measures. Higher volume in certain important hippocampal subfields, such as CA1 and CA3, was related to higher cognitive reserve as indicated by a negative correlation between volume and performance discrepancy. Strikingly, the volume of certain hippocampal subfields was correlated with IQ and MoCa score in the opposite direction as expected, where higher intelligence and higher

MoCa scores coincided with lower volume in these areas. A more detailed interpretation of the most important findings will be discussed below.

We found evidence for group level effects on the Lempel-Ziv algorithm for single channels, when comparing drowsiness to alertness. However, no such effects were found for the Lempel-Ziv complexity measure. Lempel-Ziv complexity (LZc) captures both spatial and temporal signal diversity as it is a measure that determines signal diversity across all observations and channels. In contrast, single channel Lempel-Ziv (LZs) only captures temporal signal diversity (Schartner et al., 2017). In our case, the group level effects of

informational complexity differences between the two levels of arousal could be established when only taking temporal signal diversity (LZs) into account. When spatial signal diversity (LZc) was added, no significant differences in informational complexity between the two levels of arousal could be established. In previous studies, where LZc was compared to LZs to assess the increase in informational complexity in the psychedelic state compared to normal wakefulness, the effect for LZs was more pronounced compared to LZc (Schartner et al., 2017). LZc has been successfully used to distinguish between various levels of (un)consciousness (e.g. wakefulness, anesthesia, sleep, coma) (Casali et al., 2013). However, it could be that LZs is a more sensitive measure than LZc when assessing more subtle variations in consciousness, such as drowsiness versus alertness or the psychedelic state versus normal wakefulness.

Surprisingly, the group level effects for the single channel Lempel-Ziv (LZs) that we found were in the opposite direction to what we expected. Our findings showed that LZs significantly increased under drowsiness as compared to alertness, instead of the expected decrease. Importantly, the difference of LZs between drowsy and alert was also correlated with three of our cognitive reserve measures, namely mean reaction time, median reaction time and the number of false alarms. Furthermore, there was a nonsignificant trend between LZs and the number of misses. Specifically, a higher difference in LZs between alert and drowsy correlated with a lower task performance discrepancy. Since it was shown that LZs increases under drowsiness, these results together suggest that a higher increase in LZs under drowsiness is associated with higher

cognitive reserve as indicated by a negative correlation between LZs difference and task performance discrepancy. The same effects were found for LZc, where a higher difference in signal diversity between the two states of arousal correlated with lower performance discrepancy in the case of mean reaction time, median reaction time and the number of misses. However, since LZc failed to show significant group differences, we cannot use these findings as evidence.

It could be that the increase in informational complexity under drowsiness, as indicated by LZs, indicates some sort of compensatory mechanism, where individuals that have a more pronounced increase are better able to maintain their performance under neurocognitive strain, which is an indication for higher cognitive reserve. Therefore, the increase in informational complexity under neurocognitive strain might be interpreted as a functional neural correlate for cognitive reserve. Interestingly, in a study of Fernandez et al. (2010) they found an altered relationship between Lempel-Ziv complexity and ageing in individuals with mild cognitive impairment (MCI), as compared to individuals with Alzheimer and healthy controls. Individuals with AD and healthy controls showed a reduction in informational complexity as a function of age. However, in MCI individuals, the relationship was in the opposite direction, and they exhibited higher overall informational complexity compared to the other two groups. Fernandez et al. (2010) argued that this might be an indication of the presence of a compensatory mechanism in MCI individuals, before progressing to full-blown dementia. Taken together, it could be that an increase in informational complexity under neurocognitive strain is a general marker for a compensatory mechanism for various different types of neurocognitive strain, whether it is due to pathology (e.g. MCI) or due to temporary factors (e.g. drowsiness). If this is the case, then it would be expected that individuals with higher cognitive reserve show the strongest increase in informational

complexity, as this compensatory mechanism is more effective in high cognitive reserve individuals. To further investigate this intriguing finding, using additional measures to analyze informational complexity under neurocognitive strain could provide more insight into what is happening on a neural level. There are many different techniques that are able to capture different facets of informational complexity and information transfer in the brain (Arsiwalla & Verschure, 2018). Lempel-Ziv is very useful to capture neural signal diversity and relate this to various conscious states but it may be less suitable to analyze the effectivity of information transfer at a global neural level. An interesting, additional analysis that might be useful in the context of our study concerns network analysis, particularly the examination of small world properties. A small-world network is a type of network that consists of many short-range connections and a few long-range connections, and is believed to be an optimal type of network organization for effective information transfer (Watts & Strogatz, 1998). There is experimental evidence that the brain network exhibits small-world properties, both on a functional (e.g. alpha network efficiency) and a structural level (Bassett & Bullmore, 2006), and the degree of small-worldness is often related to various pathologies, where a diseased state

generally is associated with a disruption of the small-world brain network (Chennu et al., 2014). In addition, neurocognitive strain, such as mental fatigue, has also been shown to affect small-world properties in an adverse way (Sun et al., 2014). Intriguingly, there is a relationship between maintaining performance under the neurocognitive strain under mental fatigue and less or no disruption of small-world properties on a functional level (Sun et al., 2014). In the context of our study, it could be very interesting to assess small-world properties (e.g. in the alpha range) and relate this to the LZs increase and maintaining task performance under neurocognitive strain. It would be expected that the participants with a higher increase in LZs and better maintenance of task performance would show no or less disruption in small-world properties, and vice versa. If this is the case, this would provide more evidence for informational complexity as a neural correlate for cognitive reserve.

Another key finding of the current study, concerns the association between better task performance under neurocognitive strain and higher volumes in several hippocampal subfields. Specifically, the findings were: (1) A lower increase in the number of misses under drowsiness significantly correlated with higher volumes in the left HATA, right CA3, right CA4, right molecular layer HP and the right GC-ML-DG. Furthermore, non-significant trends were found for the right CA1, the right hippocampal tail and the right presubiculum, (2) A lower increase in the variability of reaction times under drowsiness significantly correlated with higher volumes in the right CA1, right CA3, right CA4 and the right molecular layer HP, (3) A lower increase in the median reaction times under drowsiness was significantly correlated with higher volumes in the left HATA, and, (4) A lower increase in mean reaction times under drowsiness showed a non-significant trend with higher volumes in the left-HATA. There were no associations found between false alarms and hippocampal volume.

These findings are not easy to interpret. The hippocampus is a heterogeneous structure, consisting of several different subfields that are all histologically distinct and the cyto-architecture of the different subfields may be related to subtle differences in functionality. Although the function of the hippocampus is mainly associated with memory (Van Petten, 2004), there is evidence that several subfields are also related to different cognitive functions. For instance, the CA1 has been implicated in baseline attention (Foo et al., 2017) and Muzzio et al. (2009) suggested that the CA1 might be an attentional gate, since its neuroanatomical position allows it to compare information within the tri-synaptic loop, to new information coming in from the cortex. Another example concerns the HATA, which is the hippocampal-amygdaloid transition region and has been implicated in visuospatial processing and object recognition (Foo et al., 2017). Furthermore, it has been suggested that volume reductions in the HATA, CA3, CA4, GC-ML-DG, or parasubiculum could lead to reduced connectivity between the hippocampus and amygdala and might be regarded as early biomarkers of cognitive decline (Foo et al., 2017).

Interestingly, our results show that higher volumes of the CA1, HATA, CA3, CA4 and GC-ML-DG were all related to better task performance under neurocognitive strain. However, at this point it cannot be fully ruled out that some of these associations were task specific. For instance, if the CA1 indeed functions as an attentional gate, it could be that the CA1 effects we found were task specific, since the SART is a sustained attention task. Nonetheless, it must be said that our findings show several subfields to be more or less consistently associated with task performance under neurocognitive strain. Furthermore, the hippocampus is thought to have neuroprotective effects (Valenzuela et al., 2008), is important for cognitive functioning (de Wael et al., 2018), and is implicated in cognitive decline in AD and other neurological illnesses (Zhao et al., 2019; Foo et al., 2017). Taken together, the idea that the hippocampus is involved in some sort of reserve mechanism is plausible, but alternative explanations cannot be confidently ruled out at this stage. To gain a more in-depth understanding of what our findings mean in the context of cognitive reserve, a more detailed assessment of the functionality of the different hippocampal subfields and an examination of structural and functional connectivity of these subfield with other brain regions is needed. The hippocampus may be part of a larger cognitive reserve network in the brain. Therefore, it is important to additionally assess the potential relation between volume reduction and task performance in other brain regions that may be relevant for cognitive reserve, such as prefrontal areas due to their importance for executive functioning (Collette, et al., 2006).

A final point that must be made about the hippocampal results and task performance concerns the difference between brain reserve and cognitive reserve. Brain reserve is related to structural properties of the brain, for instance, higher volume in important brain areas. Because of this higher structural density, more structural loss can be tolerated before cognitive functioning is adversely affected by neurodegenerative damage due to illnesses such as AD (Stern, 2012). Cognitive reserve is more related to the dynamic functioning of the brain and its ability to cope under neurocognitive strain (Stern et al., 2019). However, these two types of reserve are not mutually exclusive and it has yet to be discovered how cognitive reserve and brain reserve relate to one another (Stern, 2012). Furthermore, it is important to note that brain reserve is not a completely

static factor. Several behavioral interventions, such as meditation training (Luders et al., 2009) and memory training (Valenzuela et al., 2003) are known to improve functional and structural changes in the form of increased volume and activity in various brain areas (e.g. prefrontal areas, hippocampus). In the case of our findings, it could be that the results found in the hippocampus are mostly related to brain reserve while the dynamic changes in informational complexity relate more to cognitive reserve.

Another striking finding was the association between IQ, MoCa and volume of several hippocampal subfields that were contrarily to the expected direction. Specifically, our findings were: (1) A higher fluid intelligence IQ, as measured with the Cattell test, correlated with lower volumes in the left hippocampal tail and the left hippocampal fissure, (2) A higher difference between estimated premorbid IQ (NART) and fluid intelligence correlated with a higher volume in the left hippocampal fissure but a lower volume in the left fimbria, (3) A higher score on the MoCa test correlated with lower volumes in the right CA1, right CA3 and the right GC-ML-DG. In case of the MoCa this may be explained in the following way. Most of the participants had a high score on the MoCa test, and the average score in our cohort was 27. Only six participants scored below 25, with scores of 24 or 23. Of these six participants, two were excluded from the analyses due to a too low number of alert or drowsy (target) trials. Since most of the participants scored similarly on the MoCa test and exhibited good cognitive functioning, it could be that the correlations found with the Spearman’s ranking correlation test were false positives. Furthermore, MoCa score was not related to the informational complexity measures nor was it related to task performance. Also, it is important to mention that with older participants the difference between a good day and a bad day (e.g. due to lack of sleep) is more apparent, since aging is already a form of cognitive strain, which could have influenced some of the results regarding the various psychological tests. Similarly to MoCa, IQ was not related to informational complexity and failed to relate consistently to task performance. The only effect that was found concerns a correlation between IQ and variability in reaction times, where a higher IQ was associated with increased variability in reaction times under drowsiness. However, IQ did not relate to an increase in reaction time under drowsiness, nor was it associated with an increase in errors.

A potential issue with the current study that has to be mentioned concerns the drowsiness classification algorithm (Jagannathan et al., 2018) that was used in the study. This algorithm has only been tested in young participants and the distinction between and alert state and a drowsy state is heavily dependent on transitions in the Alpha and Theta frequency bands (Hori 1-5), where Alpha waves decrease when individuals get drowsy while Theta waves increase. However, it is more challenging to detect these transitions in the Alpha frequency band in individuals who do not exhibit prominent Alpha waves (Ogilvie, 2001) and Alpha activity is known to decrease with age (Klimesch, 1999). Since our study used participants in the age-range of 60-84 years old, it is possible that some of our participants lacked prominent Alpha waves. This might also explain the fact that many participants had to be excluded from further analysis, because they did not have enough alert (target) trials. Moreover, even in participants that were included in the analyses the drowsiness classification algorithm might not have worked optimally. Therefore, it is important to mention that false negatives and false positives cannot be ruled out, especially concerning the psychological data. Almost none of the proxy measures of cognitive reserve (e.g. IQ, educational level) could be successfully related to task performance, and the ones that did correlate (e.g. false alarms and educational level, IQ and coefficient of variation) failed to do so consistently. In addition, some of the effects that did show a more consistent pattern of statistical significance (e.g. LZs), which makes it more plausible to assume that these are real effects, might have had larger effect sizes with a more suitable method to classify states of alertness and drowsiness in elderly participants.

One way to tackle this problem is to use a different method to establish individual differences in cognitive reserve, which is not dependent on drowsiness. For example, it could be interesting to use the concept of mental fatigue instead of drowsiness. As mentioned before, a study of Sun et al. (2014) investigated the effect of mental fatigue on task performance, and related individual differences in performance

maintenance in young subjects directly to functional network properties in the brain. The most interesting finding was that individual differences in how well participants performed under the neurocognitive strain of mental fatigue was directly related to small-world properties in the functional Alpha frequency brain, which was less affected as compared to baseline in participants that were able to maintain their performance level. Mental fatigue was induced by the performance of a cognitively and attentionally demanding task, namely the psychomotor vigilance task (PVT) (Dinges et al., 1997). The advantage of using this task over the SART (Seli et al., 2012) is that the PVT is a more challenging task which may lead to more clear differences in task

performance discrepancy between the participants (i.e. mental fatigue versus baseline), and therefore, might be a more sensitive measure for cognitive reserve. Another important advantage of the PVT is that clear