Spectral power ratio as a measure of EEG changes in mild cognitive impairment due to Alzheimer’s disease: a case-control study

Abstract

Adopting preventive strategies in individuals with subclinical Alzheimer’s disease (AD) has the potential to delay dementia onset and reduce health care costs. Thus, it is extremely important to identify inexpensive, scalable, sensitive, and specific markers to track disease progression. The electroencephalography spectral power ratio (SPR: the fast to slow spectral power ratio), a measure of the shift in power distribution from higher to lower frequencies, holds potential for aiding clinical practice. The SPR is altered in patients with AD, correlates with cognitive functions, and can be easily implemented in clinical settings. However, whether the SPR is sensitive to pathophysiological changes in the prodromal stage of AD is unclear. We explored the SPR of individuals diagnosed with amyloid-positive amnestic mild cognitive impairment (Aβ+aMCI) and its association with both cognitive function and amyloid load. The SPR was lower in Aβ+aMCI than in the cognitively unimpaired (CU) individuals and correlated with executive function scores but not with amyloid load. Hypothesis generating analyses suggested that aMCI participants with a lower SPR had an increased probability of a positive amyloid PET. Future research may explore the potential of this measure to classify aMCI individuals according to their AD biomarker status.

1. Introduction

In 2018, 50 million people worldwide were diagnosed with dementia, generating an estimated cost of 1 trillion US dollars for public health systems (Alzheimer’s Disease International, 2018). Dementia is projected to affect almost 152 million people by 2050 (Alzheimer’s Disease International, 2018). The main cause of dementia is Alzheimer’s disease (AD), a neurodegenerative disease characterized by the accumulation of amyloid plaques and neurofibrillary tangles (Knopman et al., 2019). Amnestic mild cognitive impairment (aMCI) is a condition that involves cognitive deficits, predominantly in the memory domain, while preserving the personal capacity to execute daily activities; and is most commonly the precursor of dementia due to AD.

Ongoing research suggests that adopting lifestyle-based and pharmacological preventive strategies for AD in the early stages of the disease may preserve cognitive functions better than later interventions (Ngandu et al., 2015; Rosenberg et al., 2020). However, the need for prospective trials involving individuals at the highest risk of developing dementia due to AD remains. Therefore, finding scalable, sensitive and specific early markers of cognitive decline is of paramount importance. Current biomarkers of dementia due to AD rely on the quantification of proteins including amyloid-beta (Aβ) and tau, in samples of cerebrospinal fluid (CSF) or measured by positron emission tomography (PET), as well as measures of atrophy via magnetic resonance imaging (Frisoni et al., 2017). However, such techniques are relatively invasive, costly, may be inaccessible outside of subspecialty medical centers, and vary in the degree to which they correlate with the severity of clinical and cognitive symptoms. The International Federation of Clinical Neurophysiology (IFCN)-sponsored panel of experts has proposed the electroencephalogram (EEG) as an alternative for tracking cognitive decline at prodromal stages of the spectrum from healthy aging to dementia due to AD, in combination with other markers (Rossini et al., 2020). EEG is well-suited to track disease progression since it is sensitive to synaptic dysfunction, a pathophysiological feature with a pre-symptomatic onset, and it is a low-cost, non-invasive technique that is currently implemented in healthcare systems worldwide (Dennis J. Selkoe, 2002; Palop and Mucke, 2016).

EEG records the rhythmic changes in voltage generated by the synchronized activity of populations of neurons, through electrodes placed on the scalp (Cohen, 2017). This synchronization is key for neural computation and depends on neural circuit physiology (Wang, 2010). Alterations in EEG patterns are a consistent feature of a number of neurocognitive disorders, such as AD (Babiloni et al., 2004; Horvath et al., 2018; Moretti et al., 2004). The most prominent EEG hallmark in such disease is a general shift of the power spectrum from faster to slower frequencies, including a widespread power density increase in the delta (1–4 Hz) and theta (4–8 Hz) frequency bands, a posterior decrease in the frequencies within the alpha (8–13 Hz) and beta (13–30 Hz) bands, and a slowing of the individual alpha peak frequency (Babiloni et al., 2004; Horvath et al., 2018; Moretti et al., 2004).

A previous study by our research group in healthy older participants, individuals with type-2 diabetes mellitus (T2DM), and patients with mild-to-moderate AD; quantified these overall shift in spectral power with a single measure, the spectral power ratio (SPR), calculated as the ratio of power in alpha (α) and beta (β) bands over power in delta (δ) and theta (θ): (α+β)/( δ+θ) (Benwell et al., 2020). The results of this study revealed a lower SPR in AD relative to both T2DM and healthy elderly, as well as an association of lower SPR with worse cognitive function (most prominently with Executive Function, but also for Learning and Memory, and global Dementia Severity scores). In contrast, the posterior dominant frequency (PDFr), an adaptation of the individual alpha frequency sensitive to shifts in a larger spectrum than the alpha band (5–15 Hz), varied across diagnoses, but showed no correlation with cognitive function. These results suggest the SPR may be a measure of cerebral slowing that more closely tracks cognitive decline than individual alpha frequency as measured as the PDFr.

The present study investigates whether the SPR detects pathophysiological changes in individuals with amnestic mild cognitive impairment (aMCI). We hypothesized that the disruption of brain oscillations previously reported in AD would be present before the onset of dementia. Therefore, we expected a lower whole brain SPR in patients with aMCI relative to age-matched cognitively unimpaired (CU) individuals. Also, we expected that lower SPR values in individuals with aMCI would correlate with worse Executive Function scores. For completeness and comparison with the results of Benwell et al., 2020, we also examined changes in the PDFr and power at PDFr in the present cohort.

Soluble Aβ 42 oligomers are thought to mediate synaptic and dendritic damage from pre-symptomatic stages of the AD (Shankar et al., 2007). In later stages of the disease progression, amyloid plaques are thought to serve as reservoirs of amyloid oligomers generating a pool of toxic proteins that damage neighboring neurites and synapses (Forloni and Balducci, 2018; Haass and Selkoe, 2007). In line with this premise, we tested whether the EEG-based metrics of the shift in spectral power—SPR, PDFr and power at PDFr—are associated with amyloid plaque burden, as measured through positron emission tomography (PET).

2. Methods

2.1. Participants

In this cross-sectional study, we recruited 37 participants aged between 53 and 87 years old, composed of 23 aMCI individuals (10 females), and 14 cognitively unimpaired controls (CU) (7 females) (Fig 1). Participants’ demographics are described in Table 1. All procedures took place between 2016 and 2019 at the Berenson-Allen Center for Noninvasive Brain Stimulation (BA-CNBS) at the Beth Israel Deaconess Medical Center (BIDMC). The study was approved by the Institutional Review Board at BIDMC and all participants provided written informed consent before enrollment in the study, in concordance with the Declaration of Helsinki. Participants were recruited via flyers as well as through a repository of individuals who had participated in previous studies at our Center. None of the cognitively unimpaired individuals included in the present study were spouses or relatives of the MCI group, nor did any of them express any subjective cognitive complaints.

Figure 1. Consort diagram of study screening, enrollment and analysis:

a total of 17 Aβ+aMCI and 14 CU were used for the analysis. aMCI: amnestic mild cognitive impairment; Aβ+aMCI: amyloid positive individuals with mild cognitive impairment; Aβ−aMCI: amyloid negative individuals with mild cognitive impairment; CU: cognitively unimpaired individuals

Table 1.

Participant’s demographics

	Prospective cohort			Retrospective cohort
	CU	Aβ+aMCI	Aβ−aMCI	CUretro	AD
Sample size (n)	14	17	6	17	17
Number of females	7	8	2	9	11
Years of education
Mean	16.20	16.53	18.33	16.06	16.53
Std. Deviation	2.49	2.62	3.01	2.54	3.64
Shapiro-Wilk test	0.89	0.95	0.97	0.91	0.91
P-value Shapiro-Wilk test	0.1524	0.4585	0.9009	0.1108	0.1061
Age
Mean	66.71	70.71	71.67	67.65	70.18
Std. Deviation	9.24	8.72	8.78	6.69	6.97
Shapiro-Wilk test	0.97	0.86	0.87	0.93	0.92
P-value Shapiro-Wilk test	0.8925	0.0157*	0.2133	0.2156	0.1434
Premorbid IQ
Mean	117.10	116.76	116.67	117.06	109.47
Std. Deviation	12.38	10.15	11.62	8.41	10.26
Shapiro-Wilk test	0.74	0.79	0.76	0.84	0.95
P-value Shapiro-Wilk test	0.0028*	0.0016*	0.0260*	0.0086*	0.4628
MMSE
Mean	29.50	25.76	27.17	29.59	21.53
Std. Deviation	0.73	2.82	1.34	0.77	2.23
Shapiro-Wilk test	0.68	0.93	0.96	0.58	0.85
P-value Shapiro-Wilk test	0.0003*	0.2496	0.8043	6.7137e -6*	0.0095*
GDS
Mean	1.30	2.12	1.83	0.65	2.71
Std. Deviation	2.87	2.03	1.72	1.00	2.49
Handedness
Right	12	15	5	17	14
Left	1	2	1	0	3
Ambidextrous	1	0	0	0	0
Individuals on cholinesterase inhibitors	0	4	2	0	6
Individuals on memantine and cholinesterase inhibitors	0	1	0	0	9

Std.: standard deviation; CU: cognitively unimpaired controls; Aβ+aMCI: amyloid positive mild cognitive impairment; Aβ−aMCI: amyloid negative mild cognitive impairment; CU_retro: cognitively unimpaired controls from retrospective cohort; AD: Alzheimer’s disease; GDS: Geriatric Depression Scale.

^*:significant deviation from a normal distirbution.

The exclusion criteria for enrolling in the study included unstable medical conditions, history of neurological or psychiatric illness including depression and anxiety (Geriatric Depression Scale), history of diabetes, moderate or severe neurovascular disease, or premorbid IQ below 80 as measured by the age-adjusted Wechsler Test of Adult Reading (Bright and van der Linde, 2020). Inclusion criteria for the aMCI group criteria consisted of a diagnosis of aMCI confirmed by a board-certified neurologist (DSM-V and Key Symposium criteria (Sachs-Ericsson and Blazer, 2015; Winblad et al., 2004); as well as a Clinical Dementia Rating (CDR) global score of 0.5 and Mini-Mental Status Examination (MMSE) score between 21 and 26 (Folstein et al., 1975). The amyloid status of all aMCI (n=23) participants was assessed, and 17 individuals were amyloid positive. Amyloid status was evaluated based on [18F]-Florbetapir PET in 16 individuals, and on a lumbar puncture-based assessment of cerebrospinal fluid in one individual. The data from 17 amyloid positive participants (8 females) (Aβ+aMCI) was considered for hypothesis testing analyses whereas the data from 6 amyloid negative individuals (Aβ−aMCI) (2 females) was used only for a separate set of post-hoc hypothesis generating analyses, including a logistic regression and the ROC curve, due to the strongly limited sample size of this group. The former was implemented to assess whether the SPR could aid the classification of aMCI individuals according to their amyloid status, whereas the latter aimed to find the threshold value that could optimize classification.

Participants’ medications are shown in Table 1. The CU inclusion criteria comprised unimpaired cognition (MMSE ≥ 27) and no history of diabetes since prior work from our group has shown abnormal spectral SPR in participants with T2DM (Benwell et al., 2020).

The dataset composed of CU and Aβ+aMCI was named the ‘prospective cohort’. Previous results inspecting the SPR change in AD relative to healthy controls reported an effect size (Cohen’s d) of 1.20 (Benwell et al., 2020). Under the hypothesis that the SPR shift occurs from earlier stages of the disease, the minimum sample size per group to detect the difference between the CU and Aβ+aMCI groups in a two-tailed t-test with an α=0.05 and a statistical power of 80% is 12, as verified with the G*power software (Cohen, 1988; Faul et al., 2007). In consequence, the sample sizes of the CU (14) and Aβ+aMCI (17) groups are sufficient for the present analysis.

Resting-state EEG recordings from the study by Benwell and collaborators (2020) were used to contextualize the Aβ+aMCI group. This data, named as the ‘retrospective cohort’, was acquired from 17 cognitively unimpaired controls (CU_retro) (9 females) and 17 adults diagnosed with dementia due to AD (11 females) who participated in research at the BA-CUBS at the BIDMC between 2012 and 2015. The inclusion criteria of the CU_retro were the same as of the CU of the prospective cohort. The AD inclusion criteria included a diagnosis of mild-to-moderate dementia due to probable AD according to DSM-V/NINCDS-ADRDA criteria (McKhann et al., 2012) with an MMSE score between 18 and 24 and a CDR of 1.0. Six patients were medicated with cholinesterase inhibitors, nine were on cholinesterase inhibitors and/or memantine, whereas two were not taking dementia-specific medications. See Benwell et al., 2020 for full details on this cohort.

All participants across cohorts underwent a standardized neurological exam, medical history review, neuropsychological screening and EEG recordings. Demographic data are shown in Table 1. A Shapiro-Wilk test revealed participant education years were normally distributed so we compared it among groups using an ANOVA. By comparison, age, premorbid IQ and MMSE scores deviated from the normal distribution, so these variables were compared among groups using a non-parametric Kruskal Wallis test. Gender proportions among groups were tested using a Fisher’s exact test.

2.2. Neuropsychological tests

A comprehensive cognitive evaluation was performed on a separate visit from the EEG recordings by a psychometrist under the supervision of a senior neuropsychologist (MOC). Tests and inventories from the National Alzheimer’s Coordination Center’s Uniform Data Set version 3.0 (Weintraub et al., 2018) were used, including the Trail Making Test Part A (TMT-A, time in seconds); TMT Part B (TMT-B, time in seconds); Craft 21 Story Recall: Immediate verbatim (Story Recall Immediate, 44-item); Craft 21 Story Recall: Delayed verbatim (Story Recall Delayed, 44-item); Number Span Test Backwards (NST-B, longest span); Semantic Fluency (# Animals named in 1 minute). The Digit Symbol Substitution Test (DSST, # correct in 1.5 minutes; (Jaeger, 2018) and the Rey Auditory Verbal Learning Test (RAVLT) were also applied. RAVLT sub-scores were obtained at 20-minute delayed recall (RAVLT Recall, % correct), and 20-minute delayed recognition (RAVLT Recognition, % correct; (Gale et al., 2007). Additionally, participants were assessed with a 23-item Activities of Daily Living inventory (ADLs) and the Alzheimer’s disease Assessment Scale-cognitive subscale (ADAS-Cog Total, 70 items). ADAS-Cog sub-scores were obtained for the word list immediate recall test (ADAS-Cog Recall, 10-item), and the delayed recognition test (ADAS-Cog Recognition, 12-item; (Graham and Massman, 2004). Raw scores are shown in table S1 from Additional file.

To facilitate the direct comparison between tests and the statistical analysis, raw scores for each neuropsychological measure were transformed into z-scores using published normative values reported for healthy controls around the mean age across groups (69 years old) (Amariglio et al., 2012; Gale et al., 2007; Goldberg et al., 2010; Graham and Massman, 2004; Weintraub et al., 2018). Additionally, the scores on the TMT_B-A, ADAS-Cog Total, ADAS-Cog Recall, and ADAS-Cog Recognition were inverted so that high scores reflected better performance across all measures.

Following an approach from the Alzheimer’s Disease Neuroimaging Initiative (Crane et al., 2012; Gibbons et al., 2012) and used in prior investigations by our group of the associations in early-AD between cognition and EEG spectral power (Benwell et al., 2020), cortical atrophy (Buss et al., 2018), cortical plasticity (Buss et al., 2020), and cortical hyperexcitability (Zadey et al., 2021), we averaged z-scores of tests assessing similar cognitive processes. We used these composite indices rather than specific cognitive measures to examine the association between the EEG measures and broad domains of cognition. We implemented a single score per cognitive domain to prevent the loss of statistical power that would result from testing all items measuring the same cognitive domain and to reduce the measurement error arising from single items in the neuropsychological battery. Three categories were explored: Dementia Severity (ADAS-Cog and ADLs) measuring general cognitive functioning and functional independence; Executive Function/Strategic Thinking (NST-B, TMT_B-A, semantic fluency and DSST) measuring working memory, attention, cognitive flexibility, strategic thinking, and processing speed; and Learning and Memory (RAVLT sub-scores of immediate recall, delayed recall and delayed recognition; and the Craft Story) measuring verbal short-term memory with and without context. For a detailed description of the calculations please refer to the Additional file 1. As z-scores have a normal distribution, the variance in the composite scores between groups was tested with a Pillai’s trace MANOVA as it is a more robust test to account for the deviations of the data from homogeneity of covariance (Box’s M-test = 0.0047) or multivariate normality (Shapiro-Wilk = 0.0049). The effect size was calculated as f² with the O’Brian and Shieh approach (Faul et al., 2007). Post-hoc independent sample t-tests were used to evaluate the change in each composite score across groups and Cohen’s d was calculated to evaluate their effect sizes. Additionally, the correlations between each composite score and the EEG measures were calculated using Pearson correlations.

2.3. Amyloid quantification

The amyloid status of aMCI individuals was assessed using CSF in one participant and PET in the remaining 22 individuals (Palmqvist et al., 2015). The CSF sample was obtained by lumbar puncture and confirmed as amyloid positive in concordance with the clinical cut-off of Aβ42 < 600 (Niemantsverdriet et al., 2017). Amyloid PET scans were obtained on BIDMC’s Siemens Biograph 64 mct multi-detector helical PET-CT scanner (Siemens Healthcare). Each scan was acquired during a 10-minute emission with a 128 x 128 matrix (zoom x 2), 50 minutes after intravenous injection of 10 mCi (370MBq) of [18F] Florbetapir (Doraiswamy et al., 2012). A board-certified nuclear medicine specialist blinded to the EEG results assessed the total amyloid burden on PET scans using MIMneuro™ software, version 6.8.2 (MIM Software Inc., Cleveland, OH). Default affine registration of images was used. Based on a database of normative values (Clark et al., 2012), and considering the cerebellar uptake as the standard, the amyloid z-scores for the precuneus, lateral temporal lobe, inferior medial frontal gyrus, anterior cingulate gyrus, superior parietal lobule, and posterior cingulate gyrus were calculated. A global amyloid burden score was calculated for each aMCI participant by averaging the z-scores of all six regions across the left and right hemispheres. The correlations between the global amyloid burden score and the EEG metrics (SPR, PDFr, or the power at the PDFr) of the Aβ+aMCI group were calculated. Then, the variations of the SPR across the CU, Aβ−aMCI and Aβ+aMCI groups were subjected to an exploratory hypothesis-generating analysis.

2.4. EEG recording and analysis

MCI and CU individuals within the prospective cohort underwent two identical EEG recordings in separate visits within a period of 6 weeks. The recordings from the first visit of all participants were used for SPR and PDFr comparisons, whereas the ones from the second visit were used to inspect the test-retest reliability of the EEG measures, as assessed by Pearson correlation. To verify whether the variation in inter-visit interval across subjects had an effect on the result, we implemented a partial Pearson correlation. The latter was chosen in concordance with the assumption that EEG patterns shift with disease progression and consequently, variations in inter-visit intervals could modify the results of the Pearson correlation. In each visit, 5-minute resting-state EEG was recorded using an extended version of the international 10–20 system montage with the ground and reference electrodes placed on the forehead. The EEG recordings of 20 individuals (10 CU and 10 Aβ+aMCI) from the prospective cohort were acquired using a 60-channel (eXimia EEG, version 3.2, Nexstim Ltd, Finland) system with a sampling rate of 1450 Hz. The EEG recordings of the remaining 11 individuals (4 CU and 7 Aβ+aMCI) were recorded with a 62-channel (BrainVision, BrainProducts, GmbH, Germany) EEG system with a sampling rate of 1000 Hz. No significant differences were observed between EEG montages in any of the explored measures (Table S2 from Additional file). Only the 52 electrodes shared by both montages were used for the main analyses (F1, F2, F5, F6, Fp1, Fpz, Fp2, Fz, C1, Cz, C2, C3, C4, C5, C6, CP1, CPz, CP2, CP3, CP4, CP5, CP6, FC1, FC2, FC3, FC4, FC5, FC6, F7, F8, FT7, FT8, TP10, TP7, TP8, TP9, T3, T4, O1, Oz, O2, P1, Pz, P2, P3, P4, P7, P8, POz, PO3, PO4 and Iz), and the recordings acquired were down-sampled to 1000 Hz before the power analysis.

Eyes-closed resting-state EEG was recorded while subjects sat in a semi-reclined armchair. Participants were instructed to remain quiet with their face muscles relaxed. Every ~2 minutes, individuals were asked to blink their eyes a few times to reduce drowsiness. EEG data preprocessing was performed offline using a combination of the EEGLab toolbox (Delorme & Makeig, 2004) and custom scripts in MATLAB 2017a (Mathworks, USA). As the data was collected in America, the data was filtered for line noise (around 60 Hz) using a 55–65 Hz notch filter. For this particular study, this is above the frequencies of interest, but we followed a standard processing stream (Choi & Kim, 2018) to ensure that the data can be more easily used in any potential further analysis. Then, low-pass (100 Hz) and high-pass (1 Hz) zero-phase, second-order Butterworth filters were applied. The recordings were divided into 3 s epochs for visualization and excessively faulty channels were removed (average (± SD) of 3.7 (± 2.2) per participant, range = 0–9). The remaining data were re-referenced to the average of all channels. After re-referencing, noisy epochs were identified semi-automatically and those containing excessive artifacts were rejected through visual inspection (average (±SD) of 31.3 per participant (± 13.7); in a range of 5–55), resulting in an average (± SD) of 70.0 (± 14.5) remaining epochs per participant (range = 48–99), equivalent to 210 s (± 43.5 s; range:144–297 s). Independent component analysis, using the eeglab FastICA function (Hyvarinen, 1997) was implemented to separate the data into 50 components. Components reflecting artifacts, including eye movement, muscular activity, or electrode noise were removed using a semi-automated approach. After the rejection of components, the previously rejected channels were interpolated using a spherical spline interpolation.

The EEG recordings from 17 CU_retro and 17 AD individuals from the study of Benwell et al. (2020) were used for post-hoc comparisons of the SPR and PDFr with the findings in the prospective cohort. These data consisted of 5-minute resting-state recordings acquired with the 64-channel-Nexstim system with a sampling rate of 1450 Hz and cleaned with the same methodology and similar criteria as used for the prospective cohort. To equalize the data across EEG systems, only the electrodes shared by both montages were used for the analysis (52 channels), and the recordings acquired with the Nexstim system were down-sampled to 1000 Hz before the power analysis.

The mean absolute power spectral density across epochs was calculated for each frequency band within the spectrum from 1 to 40 Hz at all electrodes using the spectopo EEGlab function (window-size = 1000 samples, 50% window-overlap, 0.5 Hz resolution after zero-padding). The absolute power within classic EEG frequency bands was obtained by summing the power estimates across all frequencies within each of the following bands: delta (1–4 Hz), theta (4–8 Hz), alpha (8–13 Hz) and beta (13–30 Hz). We calculated the SPR, PDFr and power at the PDFr of each participant in the prospective cohort and calculated the test-retest reliability of all EEG measures using a Pearson correlation. We also accounted for potential effects of variations in the inter-visit interval with a partial correlation. No significant difference was observed between EEG systems (Table S2 from Additional file).

2.4.1. Posterior dominant frequency

The PDFr was identified and the power at the PDFr was calculated for each participant. In congruency with the study by Benwell and colleagues (Benwell et al., 2020), the PDFr was calculated as the frequency within the 5–15 Hz range with the highest power density across the posterior ROI electrodes (O1, Oz, O2, P1, Pz, P2, P3, P4, P7, P8, POz, PO3, PO4, Iz). The power at the PDFr was calculated by obtaining the power over the peak frequency ±2.5 Hz relative to the posterior ROI full spectrum. In the prospective cohort, the between-group differences in the PDFr and power at the PDFr were tested using a two-tailed Welch test and an independent-samples two-tailed t-test respectively because PDFr deviated from normality (Table 3 from Additional file). Cohen’s d was calculated to assess the effect size of the differences. The differences in PDFr, and the power at PDFr of Aβ+aMCI and AD, as well as CU and CU_retro, were compared to verify whether these groups are two samples from the same underlying population.

2.4.2. Spectral power ratio

The SPR was calculated as the ratio of (α + β)/(δ + θ). Each frequency band and the SPR were calculated across all electrodes (whole brain) and groups of electrodes at defined regions of interest (ROIs), as defined in the study by Benwell and colleagues (2020): frontal (F1, F2, F5, F6, Fp1, Fpz, Fp2, Fz), central (C1, Cz, C2, C3, C4, C5, C6, CP1, CPz, CP2, CP3, CP4, CP5, CP6, FC1, FC2, FC3, FC4, FC5, FC6), temporal (F7, F8, FT7, FT8, TP10, TP7, TP8, TP9, T3, T4 ) and posterior (O1, Oz, O2, P1, Pz, P2, P3, P4, P7, P8, POz, PO3, PO4, Iz). The base 10 logarithm (log₁₀) of the SPR for the whole brain and each ROI were calculated to normalize the data for further analysis.

In the prospective cohort, the between-group difference in whole-brain SPR was log₁₀-transformed to obtain a normal distribution and then tested using an independent-samples two-tailed t-test. Cohens’s d was implemented to evaluate its effect size. As a post-hoc analysis, the differences in SPR of Aβ+aMCI and AD were compared, as well as of CU and CU_retro, to verify that these two groups are two samples from the same underlying population. Additionally, the distribution of the SPR was examined using a mixed ANOVA with the Group as a between-subjects factor and the ROI as a within-subjects factor. Partial eta squared was calculated as a measure of effect size.

2.4.3. EEG measures association with amyloid

The SPR and the PDFr were correlated with cognitive function and amyloid load and a series of hypothesis-generating analyses were performed to inspect the association between amyloid load and the SPR. A Kruskal-Wallis test was used to assess if the SPR varied across the CU, Aβ+aMCI and Aβ−aMCI groups to account for different sample sizes and the small sample size of the Aβ−aMCI group. The effect size was evaluated based on epsilon squared.

A logistic regression was implemented to test how well the SPR could be used to distinguish the amyloid status of the aMCI individuals; pseudo r² was used as a measure of the effect size. The ROC curve was calculated to determine the optimal threshold value for classification. The logistic regression and ROC analyses were treated as exploratory as a means of generating hypotheses for future studies.

3. Results

3.1. Participants

The prospective cohort (CU, Aβ+aMCI and Aβ−aMCI) showed no between-group difference in education (F (2) = 1.3420, p = 0.2766), age (H (2) = 2.4599, p = 0.2923), premorbid IQ (H (2) = 2.2689, p = 0.3216) or gender proportions (X² (2, N = 37) = 0.4854, p = 0.7845). By design, MMSE differed significantly across groups (H (2) = 18.3115, p = 0.0001). Dunn’s Post Hoc Comparisons with Holm-Bonferroni correction showed no significant difference between Aβ+aMCI and Aβ−aMCI groups (p = 0.3068) but a significant difference between CU and Aβ+aMCI (p = 4.6811e−5) and CU and Aβ−aMCI (p = 0.0096). As reported in a previous study by our research group (Benwell et al., 2020), the retrospective cohort (CU_retro and AD) demographics followed a similar trend by showing no significant difference in age or gender proportions and a significant variation of MMSE across groups (W = 289, p = 3.4999e−7).

To ensure that the prospective and retrospective cohorts were comparable, we also inspected whether all groups across cohorts (CU, Aβ+aMCI, Aβ−aMCI, CU_retro and AD) differed in education, age, premorbid IQ, gender proportions and MMSE scores. All groups were equivalent in education (F (4) = 0.7177, p = 0.5831), age (H (4) = 4.8576, p = 0.3022), premorbid IQ (H (4) = 8.0741, p = 0.0889) and gender proportions (X² (4, N = 71) = 2.1319, p = 0.7115). By design, MMSE differed significantly across groups (H (4) = 52.4204, p = 1.1266e−10). Dunn’s Post Hoc Comparisons with Holm-Bonferroni correction showed a significant difference between all groups (p < 0.05) except between CU_retro and CU (p = 0.6336) and between Aβ+aMCI and Aβ−aMCI (p = 0.6336).

3.2. Test-retest reliability

We calculated the Pearson correlation between each EEG measure at the first and second visits (total N=25). We found that the correlations were within the moderate to excellent range (0.5957–0.9850). Particularly, strong correlations between the SPR in the first and second visits were found for both the CU and Aβ+aMCI groups (r_CU = 0.9786, p_CU = 1.7509e-07; r_Aβ+aMCI = 0.8629, p_Aβ+aMCI = 0.0001), suggesting that this measure is highly replicable in both groups (Table 2).

Table 2.

Pearson correlations between EEG measures in visit 1 and visit2

	ALL		CU		Aβ+aMCI
	Pearson’s r	p-value	Pearson’s r	p-value	Pearson’s r	p-value
Log10(SPR)	0.91	1.1887e-9	0.98	1.7509e-7	0.86	0.0001
Relative delta	0.93	8.9377e-11	0.92	6.4330e-5	0.93	4.6891e-6
Relative theta	0.90	1.8955e-9	0.97	9.8366e-7	0.78	0.0016
Relative alpha	0.95	1.5311e-12	0.99	3.590e-8	0.93	4.1077e-6
Relative beta	0.95	5.5169e-13	0.95	7.5216e-6	0.96	2.1665e-7
PDFr	0.70	0.0002	0.94	1.6619e-5	0.60	0.0317
PDFr power	0.91	5.4191e-10	0.95	6.1747e-6	0.90	2.5123e-5

SPR: spectral power ratio; PDFr: posterior dominant frequency; CU: cognitively unimpaired group; Aβ+aMCI: amyloid positive group.

To verify if potential changes in inter-visit interval across subjects could affect the correlation we also calculated the partial correlation between each EEG measure using the method described by Bailey (Bailey, 1995). Partial correlations were within the moderate to excellent range (0.6160–0.9878). An excellent partial correlation was also found between the SPR in the first and second visits (r_CU = 0.9787, p_CU = 8.72e-07; r_Aβ+aMCI = 0.9491, p_Aβ+aMCI = 2.47e-06), further supporting the reproducibility of the SPR (Table S3 from Additional file).

3.3. Spectral power ratio

Descriptive statistics of all EEG measures are presented in Table S4 from the Additional file 2. All measures were tested for normality with the Shapiro-Wilk test. A two-tailed independent samples t-test revealed a significantly higher mean SPR in the CU group relative to the Aβ+aMCI group (t (29) = 2.3277, p = 0.0271, Cohen’s d = 0.8401), suggesting a large effect (Fig 2.A). When including the retrospective cohort, the Aβ+aMCI and AD SPR distributions overlapped, suggesting that the SPR could reflect pathophysiological changes that precede the onset of dementia (Figure S1.A). Therefore, the difference between Aβ+aMCI and dementia due to AD was explored as a post-hoc analysis. No significant differences between Aβ+aMCI and AD (t (32) = −0.4950, p = 0.6240, Cohen’s d = −0.1698) or CU_retro and CU (t (29) = 1.0971, p = 0.2816, Cohen’s d = 0.3960) were observed in two-tailed independent samples t-tests.

Figure 2. EEG measures of the power shift in Aβ+ aMCI.

A. Whole-brain SPR per group; B. Spectral power per ROI; Error bars represent the 95% confidence interval; C. PDFr per group. Notice that a significantly lower SPR and PDFr were observed in the Aβ+aMCI group relative to the CU group. The SPR presented an approximately equivalent regional distribution with the posterior ROI showing the highest values in both groups. However, the highest difference between the SPR of both groups was observed in the electrode P04. CU: cognitively unimpaired controls; Aβ+aMCI: amyloid positive mild cognitive impairment.

As the SPR is a whole-brain measure, it may reflect the alteration in oscillations at distinct frequency bands which are underlined by oscillators with diverse and potentially overlapping distributions (Figure S2). We thus tested the distribution of the SPR using a mixed ANOVA with the group (CU and Aβ+aMCI) as a between-subject factor and the ROI as a within-subjects factor. The mixed ANOVA (Greenhouse-Geisser sphericity corrected) showed a significant main effect of the factor ROI on the SPR (F (2.4019) = 21.0342, p = 1.0631e−8, η_p² = 0.4204), suggesting a large effect size. Similarly, there was a significant main effect of the factor group with a large effect size (F (1) = 5.3377, p = 0.0282, η_p² = 0.1554). Nevertheless, no significant interaction was observed between ROI and Group (F (2.4019) = 0.8335, p = 0.4576, η_p² = 0.0279). The posterior ROI showed the highest SPR values in both groups (Fig 2.B). These results indicate that there is regional variation in the SPR, but that this distribution is equivalent between the groups.

3.4. Posterior dominant frequency

The PDFr was lower in the Aβ+aMCI group than in the CU group further supporting the slowing of the power spectrum. The independent samples Welch’s t-test showed that the PDFr was significantly lower in the Aβ+aMCI relative to the CU group (Welch (28.0045) = 2.3582, p = 0.0256, Hedges’ g = 0.8136), suggesting a large effect (Fig 2.D). When including the retrospective cohort, the distribution of PDFr in Aβ+aMCI lay between the CU and AD distributions (Figure S1.B). The CU and AD distributions violated the assumption of normality, thus Mann-Whitney tests were used to make a comparison between the CU_retro and CU medians, as well as between Aβ+aMCI and AD. Two-tailed Post-hoc Mann-Whitney tests showed no significant difference between CU (Median = 9.5) and CU_retro (Median = 9.25), U = 139.0000, p = 0.4324, Rank-Biserial correlation = 0.1681; and no significant difference between Aβ+aMCI (Median = 7.5) and AD (Median = 7), U = 103.0000, p = 0.1525, Rank-Biserial correlation = −0.2872.

The average relative power at the peak frequency of Aβ+aMCI and CU did not differ significantly (t (29) = 0.5146, p = 0.6107, Cohen’s d = 0.1819). The distribution of the relative power at the PDFr of Aβ+aMCI was also located between CU and AD (Figure S1.C). As the relative power at the PDFr violated the assumption of the equality of variances in the Aβ+aMCI and AD comparisons, a Welch’s t-test was implemented. The Welch test showed a significant difference between AD and Aβ+aMCI that did not however survive Holm-Bonferroni correction for multiple comparisons (Welch (26.7176) = −2.3036, p = 0.0293, p_holm = 0.0580, Hedges’ g = −0.7714). No significant difference was observed between CU_retro and CU (t (29) = −0.5473, p = 0.5883, Cohen’s d = −0.1975).

3.5. Spectral features and cognitive function

A MANOVA demonstrated that the composite cognitive scores varied significantly between the CU and Aβ+aMCI individuals (Pillai’s Trace = 0.7527, F (3, 23) = 23.3341, p = 3.6532e−7, effect size f² = 3.0486) with a large effect size (Fig 3.A). A post-hoc analysis showed a significant difference for Learning and Memory (t (25) = 8.1611, p = 1.6305e−8, Cohen’s d = 3.2524), Executive Function (Welch (24.9821) = 6.7783, p = 4.2063e−7; Hedges’ g = 2.5270), and Dementia Severity (Welch (24.4470) =7.7151, p=5.2605e−8; Hedges’ g = 2.7330). The SPR was correlated with Executive Function only within the Aβ+aMCI group (r = 0.5634, p = 0.0185) (Fig 3.B), while the PDFr and the power at the PDFr were not significantly correlated with any composite score (Table 3). We inspected the distribution of the electrodes at which SPR was correlated with the Executive Function compound score and found that a significant correlation was present only in frontal and mid-posterior electrodes after FDR correction for multiple comparisons (Fig 3.C).

Figure 3. SPR and cognitive function.

A. Compound cognitive scores per group. B. Association between executive function and the SPR within the Aβ+aMCI group. C. Executive function Pearson’s correlation with SPR topography within the Aβ+aMCI group. Electrodes with a significant correlation are shown in black. Notice that the composite cognitive scores varied significantly between the CU and Aβ+aMCI individuals. The SPR was correlated with executive function within the Aβ+aMCI group. Only the SPR of frontal and mid-posterior electrodes presented a significant correlation with the executive function compound score after FDR correction for multiple comparisons.

Table 3:

Correlation of composite scores and EEG features

		Log10 (SPR)		PDFr		Power at PDFr
		CU (n=14)	Aβ+aMCI (n=17)	CU (n=14)	Aβ+aMCI (n=17)	CU (n=14)	Aβ+aMCI (n=17)
Dementia Severity	Pearson’s r	0.58	0.13	−0.12	−0.04	0.45	0.29
	p-value	0.0810	0.6205	0.7488	0.8792	0.1874	0.2598
	R2	0.33	0.02	0.01	0.00	0.21	0.08
Learning and Memory	Pearson’s r	0.26	0.10	−0.04	−0.16	0.22	0.30
	p-value	0.4646	0.692	0.9124	0.5379	0.5446	0.2451
	R2	0.07	0.01	1.60E-03	0.03	0.05	0.09
Executive Function	Pearson’s r	0.36	0.56 *	−0.23	0.43	0.28	0.32
	p-value	0.3105	0.0185	0.5259	0.0886	0.4300	0.2159
	R2	0.13	0.32	0.05	0.18	0.08	0.10

SPR: spectral power ratio; PDFr: posterior dominant frequency; CU: cognitively unimpaired group; Aβ+aMCI: amyloid positive group;

^*p < 0.05.

3.6. Amyloid load and EEG metrics

The correlation between the global amyloid burden score and the EEG metrics of the Aβ+aMCI was calculated. No significant correlations were found for the SPR, the PDFr, or the power at the PDFr at neither the whole-brain level nor the regions of interest.

A Kruskal-Wallis test showed that the SPR varied significantly across the CU, Aβ−aMCI and Aβ+aMCI groups (H (2) = 6.6101, p = 0.0367, ε² = 0.1836), with a small effect size. Interestingly, the CU and Aβ−aMCI SPR distributions strongly overlapped and presented a higher mean SPR relative to the Aβ+aMCI group. A Dunn tests with a Holm-Bonferroni correction suggested that while the mean SPR in the Aβ+aMCI group was significantly lower relative to the CU group (p_holm = 0.0449), the Aβ−aMCI and CU mean SPR did not differ significantly (p_holm = 0.2729) (Figure S3.A).

A logistic regression was performed as a post-hoc exploratory analysis to ascertain the effects of SPR on the likelihood that a participant in the MCI group has a positive amyloid test. The logistic regression model was statistically significant (Χ² (20) = 8.6126, p = 0.0033, pseudo r² = 0.3341) (Figure S3.B). The odds ratio of the model was 0.0047 suggesting that lower SPR values increase the likelihood of a positive amyloid test. The ROC curve analysis was also performed to find the cutoff value that could classify the individuals in Aβ+aMCI and Aβ−aMCI with the greatest sensitivity and specificity. A Log₁₀(SPR) ≥ 0.0166, corresponding to an SPR = 1.0390, was predictive of negative amyloid status with a sensitivity and specificity of 83% and 88%, respectively with 86% of participants correctly classified (AUROC = 0.8958) (Figure S3.C). Due to the limited number of subjects in the Aβ−aMCI group, the further exploration of this effect in future studies is highly encouraged.

4. Discussion

AD is a progressive neurodegenerative disease consisting of a cascade of pathophysiological processes that eventually lead to dementia onset. One prominent neurophysiological change associated with dementia due to AD is a shift in oscillatory EEG power and dominant rhythms from higher to lower frequencies (‘cerebral slowing’). The present study explored whether two measures that capture these changes, SPR and PDFr, are present in aMCI, a common precursor of dementia due to AD. In line with the results from Benwell et al. (2020), both the SPR and PDFr differed significantly between CU and Aβ+aMCI, with a large effect size for both.

We found no significant difference between the SPR of the Aβ+aMCI and AD groups when comparing the prospective cohort with a retrospective cohort. Despite the differences in disease severity between the Aβ+aMCI and AD groups, the similarity between their SPR distributions is highly suggestive that the SPR is sensitive to neurophysiological abnormalities that precede dementia due to AD, possibly driven by synaptic and dendritic changes due to amyloid species toxicity. If this is true, the shift in SPR could be sensitive to pathophysiological changes occurring at the preclinical stages of the AD, and the SPR would provide a tool for identifying individuals at high risk of AD, promoting their inclusion in clinical trials exploring new interventions as well as the earlier implementation of preventive strategies for dementia (Shankar et al., 2007). We ran a series of hypothesis-generating analyses to inspect SPR changes across individuals with different amyloid status. We compared the SPR of the Aβ−aMCI group with the CU and Aβ+aMCI groups and found no significant difference between the CU and Aβ−aMCI groups and a significantly higher mean SPR in Aβ−aMCI relative to the Aβ+aMCI group. We also investigated the correlation between amyloid burden and the EEG metrics (SPR and PDFr) but found no correlation. Due to the limited sample size these results should be validated in future studies. However, our preliminary analysis suggests a shift in SPR according to amyloid status. If this proves true in future studies, this effect could be driven by synapsis toxic amyloid species, such as amyloid oligomers, that lack affinity for the [18F]-Florbetapir tracer. In such case scenario, the SPR measure could enable the earlier detection of AD relative to Amyloid PET scans as the accumulation of amyloid oligomers is thought to precede senile plaques formation (Forloni and Balducci, 2018). Future longitudinal studies should explore if the shift in SPR antedates amyloid-positive PET status. Another possibility is that the shift in power spectrum may not be associated with amyloid but with other AD pathophysiological changes also present in Aβ+aMCI individuals, such as the presence of tau or neurodegeneration.

In this study, we employed logistic regression to evaluate whether the SPR could aid the classification of aMCI individuals according to their amyloid status because we considered that this analysis could have potential clinical value, if validated in a bigger sample size. Through the utilization of logistic regression, we were able to perform a receiver operating characteristic (ROC) analysis and establish a cutoff value for differentiating between amyloid-positive and amyloid-negative individuals. Although this cut-off value requires further validation in an independent dataset, it serves as a preliminary step towards generating testable hypotheses for future research and may guide the identification of novel markers for clinical use.

and the area under the ROC curve was analyzed to find the threshold value that could optimize classification. The model was significant and potentially indicates that an SPR lower than 1.039 may increase the chance of a positive amyloid test. This result suggested that the cutoff value for distinguishing the amyloid status of the individuals occurs when the power in the fast frequency bands (alpha and beta) and the slow frequency bands (delta and theta) is approximately equal.

Taken together, our results suggested that the SPR changes are present at the earliest clinically apparent stage of AD and could potentially reflect synaptic dysfunction due to early neuropathological changes occurring in prodromal AD. Thus, future studies should validate its capacity to predict the amyloid status of individuals with aMCI by using equal and larger sample sizes, as well as validate the sensitivity and specificity of the proposed cutoff value for classifying participants into the Aβ+aMCI and Aβ−aMCI groups. Additionally, future research should focus on assessing whether the shift in SPR is associated with other species of amyloid as well as other AD pathophysiological changes such as atrophy or Tau/p-Tau load.

4.1. SPR distribution

In the present study, we confirmed that the SPR differs by ROI and by Group but found no ROI*Group interaction. These findings are consistent with our prior comparisons of AD vs CU_retro and suggest that alterations in the SPR associated with early AD-type amyloid pathology occur relatively uniformly across the brain. The SPR as a whole-brain measure may reflect the alteration in oscillations at distinct frequency bands which are underlined by oscillators with diverse and potentially overlapping distributions. Nevertheless, future studies might determine whether network-specific changes in oscillatory activity, perhaps most prominently in the networks implicated in AD pathophysiology, underlie the observed whole-brain SPR changes.

4.2. High test-retest reliability

The SPR had an excellent correlation (Pearson’s r_Aβ+aMCI = 0.8629 and Pearson’s r_CU = 0.9786), even after controlling for the inter-visit interval (Partial Pearson’s r_Aβ+aMCI = 0.949 and Partial Pearson’s r_CU = 0.979). In contrast, other assessments of AD based on electrophysiology present a low to moderate reliability. For example, the test-retest reliability in 1 week time period of P300 amplitude (Pearson’s r_MCI = 0.77 and Pearson’s r_CU = 0.72) and latency (Pearson’s r_MCI = 0.42 and Pearson’s r_CU = 0.57); as well as of N200 amplitude (Pearson’s r_MCI = 0.42 and Pearson’s r_CU = 0.32) and latency (Pearson’s r_MCI = 0.48 and Pearson’s r_CU = 0.49) is substantially lower than that observed for the SPR (van Deursen et al., 2009). Our results support that the SPR could be reliably used in longitudinal research, suggesting that this measurement can be potentially used to inform further evaluation strategies in clinical practice.

4.3. Cognitive function and SPR

In line with the pattern previously described in AD (Benwell et al., 2020), we found that the Aβ+aMCI group showed a significant association between lower SPR values and decreased performance on the Executive Function composite score. We also confirmed the prior finding in AD that a similar relationship is not present for the PDFr. We also tested the correlation between Executive Function and the SPR at each electrode and found that the effect was driven primarily by frontal electrodes, in line with the association between oscillatory activity in frontal areas and Executive Function (Oswald et al., 2017). However, medial posterior electrodes were also significantly correlated in line with the association between posterior alpha and global cognitive level in aging (Babiloni et al., 2006).

We found no association between SPR and the Dementia Severity composite in the Aβ+aMCI group, in contrast to the results reported in AD. These results are in line with the notion that the cognitive impairment in Aβ+aMCI is not as generalized across domains as in AD. However, we did not replicate the previously identified association between Learning and Memory and SPR in healthy individuals. This could be explained by the stronger influence of frontal shifts in oscillations on the SPR which are rather associated with Executive Function, whereas a memory impairment would be underlined by theta-gamma decoupling in the hippocampal network that sends projections to parietal/angular gyrus regions (Bréchet et al., 2021).

While learning and memory impairments have been widely associated with the dementia due to Alzheimer’s disease, the executive function capacity plays a crucial role in the functional independence of the individuals with MCI after accounting for memory and global cognition deficits (Marshall et al., 2011). Therefore, the SPR may aid long-term prognosis by reflecting pathophysiological mechanisms associated with executive function.

4.4. Study limitations

Limitations of the present study include the small sample size. In particular, for the analysis of SPR variations according to amyloid status which are only preliminary and need to be tested in future studies implementing larger sample sizes. As we excluded individuals with other neurological or psychiatric disorders, or diabetes, or with moderate or severe neurovascular disease to prevent potential covariates. Whether other disorders could mask the shift in the SPR is beyond the scope of the currently presented manuscript. Further studies may explore the ecological validity of these results by assessing whether the SPR varies across individuals diagnosed with probable Alzheimer dementia and with different co-morbid disorders. Another limitation is that, although there were robust group-level differences between the SPR of Aβ+MCI and CU individuals, there was an overlap at the individual level. For this measure to be a viable future clinical tool, future research is needed to understand the sensitivity and specificity of longitudinal EEGs for detecting individuals at risk of developing AD pathology.

We also acknowledge the limited number of shared electrodes over the medial frontal area as a limitation of the study, as it could have decreased the sensitivity of the recording to detect activity over a key region of the DMN, decreasing the influence of memory-related structures on the SPR. Another limitation of the study is that the SPR does not include gamma activity. We decided to exclude gamma activity in our analyses because the gamma band is highly susceptible to muscular-related noise in EEG recordings which we found to be more prominent in AD and aMCI populations. While these artifacts can be addressed to a certain point though more extensive processing and high-level analysis (i.e., source reconstruction techniques), these techniques are not without their limitations and run counter to the proposition that SPR could be an easy to implement measure that could be calculated relatively easily in clinical practice. In addition, the gamma band power has very low values relative to the other frequency bands and thus would not be expected to significantly change the SPR values reported in the preset study. In the present study, we used a 64-channel EEG system and a semi-automatized pipeline for EEG preprocessing for having a high-quality EEG signal. However, this challenges the translation of the results into clinical practice as clinical EEG systems have fewer channels. Thus, future studies may also the SPR by implementing clinical EEG systems and an automatized pipeline for analysis.

5. Conclusion

The SPR is an easy-to-use whole-brain measure that captures the most prominent EEG hallmarks of AD: the shift in power from higher to lower frequencies while accounting for inter-person variability. Our results show that the change in SPR described in AD occurs already during Aβ+aMCI and has a high test-retest reliability. The present study further supports the SPR correlation with Executive Function and suggests that it may be an indicator of the progression of the AD pathology. Taken together, our results suggested that the change in the SPR is present from the early stages of the disease and could aid the earlier diagnosis of AD in the population with mild cognitive impairment, providing a better opportunity for early clinical interventions and enrollment in interventional research studies. Our results encourage the validation of the SPR capacity to differentiate Aβ+aMCI from Aβ−aMCI individuals and its’ capacity to track disease progression in future longitudinal studies with larger sample sizes.

Highlights.

• The spectral power ratio is an easy-to-use whole-brain EEG measure

• The spectral power ratio has a high test-retest reliability

• The change in spectral power ratio occurs before dementia onset

• The spectral power ratio and executive function correlate from predementia stages