Evaluating Cognitive Decline Detection in Aging Populations with Single-Channel EEG Features: Insights from Studies and Meta-Analysis

doi:10.21203/rs.3.rs-5122979/v1

Download PDF

Article

Evaluating Cognitive Decline Detection in Aging Populations with Single-Channel EEG Features: Insights from Studies and Meta-Analysis

https://doi.org/10.21203/rs.3.rs-5122979/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Timely detection of cognitive decline is paramount for effective intervention, prompting researchers to leverage EEG pattern analysis, focusing particularly on cognitive load, to establish reliable markers for early detection and intervention. This comprehensive report presents findings from two studies and a meta-analysis, involving a total of 237 senior participants, aimed at investigating cognitive function in aging populations.

In the first study, 80 seniors were classified into two groups: 40 healthy individuals (MMSE > 28) and 40 at risk of cognitive impairment (MMSE 24–27). Dimensionality reduction models, such as Lasso and Elastic Net, were employed to analyze EEG features correlated with MMSE scores. These models achieved a sensitivity of 0.90 and a specificity of 0.57, indicating a robust capability for detecting cognitive decline.

The second study involved 77 seniors, divided into three groups: 30 healthy individuals (MMSE > 27), 30 at risk of MCI (MMSE 24–27), and 17 with mild dementia (MMSE < 24). Results demonstrated significant differences between MMSE groups and cognitive load levels, particularly for A0 and Gamma band.

A meta-analysis, combining data from both studies and additional data, included 237 senior participants and 112 young controls. Significant associations were identified between EEG biomarkers, such as A0 activity, and cognitive assessment scores including MMSE and MoCA, suggesting their potential as reliable indicators for timely detection of cognitive decline. EEG patterns, particularly Gamma band activity, demonstrated promising associations with cognitive load and cognitive decline, highlighting the value of EEG in understanding cognitive function.

The study highlights the feasibility of using a single-channel EEG device combined with advanced machine learning models, offering a practical and accessible method for evaluating cognitive function and identifying individuals at risk in various settings.

Biological sciences/Neuroscience/Diseases of the nervous system/Alzheimers disease

Biological sciences/Neuroscience/Diseases of the nervous system/Dementia

Biological sciences/Neuroscience/Diseases of the nervous system/Neurodegeneration

Biological sciences/Neuroscience/Cognitive ageing

Biological sciences/Neuroscience/Learning and memory

Biological sciences/Neuroscience/Neural ageing

Cognitive decline poses a significant challenge, making the implementation of timely detection methods essential [1], [2]. The advent of disease-modifying therapies such as Aducanumab [3], [4] and Lecanemab [5], which target amyloid plaques, a hallmark of Alzheimer’s Disease (AD), offers potential to alter disease progression. These FDA-approved therapies have demonstrated efficacy primarily when administered in the initial phases of AD. Identifying subtle changes in cognitive function before significant deficits occurs is paramount for maximizing the therapeutic benefits of these drugs, ultimately aiming to preserve cognitive function and improve quality of life in at-risk individuals [6]. Standard tools such as the Mini-Mental State Examination (MMSE) and the Montreal Cognitive Assessment (MoCA), are widely used in clinical settings but have notable limitations. Despite its widespread use, the MMSE is criticized for limited sensitivity and criterion validity, often leading to undetected cognitive deficits [7]. Research shows that this lack of sensitivity allows many patients with cognitive impairments to go unnoticed, undermining the need for cognitive rehabilitation [8].

Electroencephalography (EEG) provides a non-invasive window into real-time cognitive processes. It effectively identifies changes in power spectral density, along with disruptions in functional connectivity and altered coherence patterns associated with cognitive decline and AD [9], [10]. These disruptions in neural processing and connectivity underscore the complexity of neurophysiological changes linked to declining cognitive functions [11]. Research shows cognitive decline often involves reduced amplitude and synchronization of Gamma wave activity [12]. In AD patients, elevated Gamma activity was evident during performance of cognitive tasks, potentially indicating an increased resources allocation under cognitive load compared to healthy seniors [13]. Similarly, Beta power was notably higher in MCI patients than in control subjects, both at rest and during working memory tasks [14]. Furthermore, individuals with MCI and AD exhibited diminished Delta [15] and Theta [16] power during auditory and visual oddball tasks compared to healthy controls. EEG provides objective, quantifiable data to identify abnormal brain patterns, supporting timely diagnosis and intervention to enhance cognitive health in individuals at risk [17].

In recent years, machine learning (ML) and deep learning (DL) approaches have been increasingly applied to EEG data to predict early cognitive decline, yielding high accuracy results. A review analyzing 209 studies found that DL models, particularly convolutional neural networks (CNNs) and support vector machines (SVMs), can achieve accuracies exceeding 93% in distinguishing between cognitive decline stages [18]. Another review of 116 studies on the progression from MCI to AD reported that ML techniques, including SVMs, random forests, and CNNs, delivered classification accuracies up to 95% and AUC values of 0.98 for EEG-based predictions [19]. Additionally, ensemble methods and feature selection techniques like Lasso and ElasticNet were frequently used to enhance model performance [20]. While these results show potential, the effectiveness of ML and DL techniques applied to EEG data is often limited by small sample sizes, impacting generalizability and increasing the risk of overfitting. Studies with larger EEG cohorts have demonstrated the potential of these approaches to detect cognitive deficits, enhancing their applicability to broader populations and widespread use in clinical settings. For instance, a study involving frontotemporal EEG data from 120 participants showed that EEG multifractal analysis, combined with ML models, effectively detected MCI in healthy individuals, correlating well with normal MMSE scores (≥ 26) [21]. Another study analyzed resting-state prefrontal EEG biomarkers from 496 elderly individuals and used various ML methods (including WLS, Ridge, ElasticNet and Lasso) to predict cognitive impairment. These models achieved moderate AUC (0.849) and accuracy (0.754), effectively differentiating between individuals at risk of MCI and those with cognitive deficits [22]. These findings underscore the potential of advanced ML and DL methods for accurate, non-invasive early diagnosis of cognitive decline.

Our previous pilot study [23] included 50 seniors with MMSE scores ranging from 10 to 30, divided into three groups: MD (17–23), MCI-R (24–27), and healthy (28–30). EEG data was collected during an auditory cognitive assessment with varying cognitive load levels and at rest. Pre-extracted EEG features, validated in prior studies conducted on young, healthy subjects [24], [25], [26], as well as elderly populations [27], [28], showed significant correlations with MMSE scores, particularly ST4 and A0, across task difficulty levels. Furthermore, these features effectively distinguished between seniors with high vs. low MMSE scores. EEG features Theta, Delta, A0, and VC9 increased with higher cognitive load levels, indicating different activity patterns between young and senior participants in different cognitive states, particularly notable for VC9, which differentiated between all levels of cognitive load. This pilot study demonstrated that single-channel wearable EEG and ML features can effectively evaluate cognitive states and align with clinical measurements for detecting cognitive decline.

The recent FDA approval of drugs designed to slow beta amyloid buildup in AD, results in seniors increasingly seeking evaluations for eligibility for these new treatments, placing a substantial burden on clinicians. This situation intensified the demand for highly specific AI-based assessments that can accurately distinguish between healthy individuals and those who may need further evaluation. Motivated by this need, the first study presented here focuses on the high range of MMSE scores (24–30, typically considered healthy), aiming to distinguish between cognitively healthy individuals and those who may be at risk for early cognitive decline, with a cutoff score of 27. The goal was to identify subtle cognitive changes that may signal the onset of decline among elderly individuals using EEG biomarkers.

The second study aims to validate the outcomes of the pilot study by incorporating additional clinical diagnostic tools, such as MoCA. While MMSE is a reliable tool, its sensitivity can be limited by educational level variations. A review of over 50 studies indicates that MoCA exhibits greater sensitivity than MMSE in detecting subtle early-stage MCI deficits [29]. Additionally, previous work demonstrated that a single-channel EEG approach successfully extracted features comparable to MoCA scores [30]. To further enhance the clinical aspect of our assessment, we introduced functional tasks in the second study protocol. Two tasks from the Performance Assessment of Self-Care Skills (PASS), which evaluates functional status and change [31]. A study found significant associations between PASS tasks focusing on cognitive skills and performance in verbal memory and executive function, effectively differentiating MCI subjects from healthy controls [32].

Finally, we conducted a meta-analysis of data gathered in all three studies, including additional healthy controls (n = 349), to achieve a comprehensive perspective on the relationships between EEG features, cognitive assessments, and functional tasks in the elderly population.

Participants

The first study recruited 80 patients from the inpatient rehabilitation department at Dorot Geriatric Medical Center, with a mean age of 73.51 (10.45) years, evenly distributed between males and females. Both groups exhibited a diverse age range. An age difference was observed between the healthy male group and the MCI-R male group. This difference is primarily attributed to the presence of a male participant in the MCI-R group, whose age (101 years) is more than two standard deviations above the group average. When this outlier is excluded from the analysis, the age difference between the male subjects is no longer significant (t = -1.86, p = 0.07). Other than that, no significant age differences were found between the groups.

The second study included 77 patients from the same department, with a mean age of 74.17 (8.90) years, comprising 52% females and 48% males. Each group displayed a wide age range. Differences in age were observed between the MD group and the other two groups (MCI-R and Healthy), particularly among female participants. This can be explained by the established understanding that the prevalence of dementia increases with age. Previous research indicated a modest rise in MCI rates with age [33], unlike dementia, where prevalence nearly doubles every 5-year increase in age [34]. Consistent with the current study, a large-scale study of older women found that females with dementia were significantly older than cognitively healthy participants [35]. Full demographic details are provided in Tables 1 and 2.

In both studies, clinical staff identified potential participants during hospital admissions. Participants were selected based on study inclusion criteria and had MMSE scores of 24–30 (first study) or 10–30 (second study). All patients provided informed consent in line with the Declaration of Helsinki. Individuals who objected or had neurological comorbidities, scalp or skull damage, facial skin irritation, significant hearing impairments, or a history of significant drug abuse were excluded. Ethical approval for both studies was granted by the Ethics Committee (EC) of Dorot Geriatric Medical Center. The approval of the first study was granted on September 07, 2020, NIH Clinical Trials Registry number: NCT04683835. The approval of the second study was granted on March 01, 2022, NIH Clinical Trials Registry number: NCT05528445.

For the meta-analysis, we included additional 146 healthy participants (aged 18–80) who completed auditory cognitive tasks. Ethical approval was obtained from Tel Aviv University.

2.1.1 Study groups

Figure 1 illustrates the group allocation and analysis details for each part of the study.

In the first study, participants were divided into two groups based on their MMSE scores: Healthy group (MMSE scores of 28–30, n = 40); and MCI-R group (MMSE scores of 24–27, n = 40);

In the second study, participants were divided into three groups based on their MMSE scores: Healthy group (MMSE scores of 28–30, n = 30); and MCI-R group (MMSE scores of 24–27, n = 30); MD group (MMSE scores of 10–23, n = 17).

We used MMSE score cutoffs of 24 and 27 for group allocation, focusing on timely detection of cognitive decline. Previous evidence suggests that a higher cutoff score enhances diagnostic accuracy [36]. Additionally, research indicates that educated individuals scoring below 27 on the MMSE are at increased risk of developing dementia [37].

Finally, the meta-analysis included data from both studies and additional healthy participants, totaling 237 elderly individuals (allocated as in the second study) and 112 healthy young participants.

Clinical and demographic data

To enhance the validation of clinical assessments and cognitive states of participants, additional evaluations were conducted alongside the MMSE in both studies. In the first study, participants underwent Instrumental Activities of Daily Living (IADL) assessments, which measures daily living tasks across eight domains, with scores ranging from 0 (low functioning) to 23 (high functioning) [38]. The IADL is self-reported and assessed through interviews and has seldom been linked to objective measures like brain activity. However, a study using single-channel EEG effectively classified elderly subjects based on IADL scores [39].

In the second study, several clinical assessment methods were collected including the Montreal Cognitive Assessment (MoCA), the Geriatric Depression Scale (GDS) for depression diagnosis, and the Executive Clock Drawing Task (CLOX) for assessing cognitive impairment. Additionally, demographic and sleep-related data were collected in the second study.

The MoCA, scoring from 0 to 30, identifies MCI and early dementia, with a score of 26 or higher indicating normal cognitive function. Designed for the detection of MCI or early Dementia by healthcare professionals, the MoCA evaluates various cognitive domains including visuospatial abilities, memory, attention, and delayed recall [40]. The GDS, designed for elderly individuals, consists of "yes" or "no" questions about the past week's emotional experiences, scores from 0 to 15, with higher scores indicating more severe depression [41]. The CLOX task, involving drawing and replicating a clock, scores from 0 to 15, with lower scores indicating greater cognitive impairment [42].

EEG device

EEG recordings were conducted using the Neurosteer® single-channel high dynamic range EEG (hdrEEG) Recorder. A three-electrode medical-grade patch was placed on each subject’s forehead, using dry gel for optimal signal transduction. The non-invasive monopolar electrodes were positioned at the prefrontal regions, with the single-EEG-channel derived from the difference between Fp1 and Fp2 in the International 10/20 electrode system and a reference electrode in Fpz. The data were digitized continuously at a 500-Hz sampling frequency.

2.1.2 Signal processing and high-level features

In recent years, a time-frequency approach has been adopted for analyzing EEG data to characterize brain states in AD [43], [44], [45]. In line with this approach, our study employs an advanced time-frequency method to process the EEG signal, as previously described [23], [25], [28]. The EEG features are produced by a secondary layer of machine learning applied to labeled datasets previously gathered by Neurosteer, to derive several linear combinations. Specifically, the EEG features VC9 and A0 were calculated employing the linear discriminant analysis (LDA) technique [46]. LDA is designed to identify an optimal linear transformation that maximizes class separability. Previous studies employing LDA models on imaging data have demonstrated success in predicting the development of cognitive decline. Simple LDA models using MRI and PET data were shown to predict cognitive decline or stability up to four years prior to the manifestation of decline symptoms [47]. The calculation of EEG feature ST4 utilized principal component analysis (PCA) [48], a technique employed for reducing feature dimensionality before classification. Research indicates that features extracted through PCA exhibit a significant correlation with MMSE scores and effectively distinguish individuals with AD from healthy subjects [49], [50], [51]. Notably, all three EEG features were derived from datasets different from those analyzed in the current study, to avoid overfitting the data. Consequently, the weight matrices previously determined were applied to transform the data acquired in the present study.

In studies conducted on young healthy participants, VC9 feature showed increased activity with escalating levels of cognitive load manipulated by a numeric n-back task [24]. Furthermore, during an arithmetic task, VC9 activity decreased in response to external visual interruptions [26]. Additionally, in a surgery simulator task performed by medical interns, VC9 activity declined with task repetition, correlating with individual performance [25]. VC9 demonstrated greater sensitivity than Theta particularly for tasks with lower cognitive load, making it more suitable for clinical and elderly populations. Notably, in the preceding pilot study [23], higher cognitive load levels resulted in increased VC9 activity exclusively in the healthy young group compared to the healthy senior group, highlighting different activity patterns between young and senior participants across various cognitive states. In clinical settings, VC9 activity correlated with the auditory mismatch negativity (MMN) component in minimally responsive patients [52].

EEG feature A0, previously identified as a classifier for distinguishing cognitive load from rest in healthy subjects, has proven to be a robust predictor of cognitive decline in individuals with mild-to-moderate impairment [23]. Furthermore, A0 effectively differentiates between healthy controls and Parkinson’s disease (PD) patients, with higher activity observed in healthy individuals [28].

EEG feature ST4 was found to correlate with individual performance in the numeric n-back task, specifically correlating the disparity in RTs between high and low cognitive load levels to differences in ST4 activity per participant [24]. In the preceding pilot study [23], ST4 demonstrated the ability to differentiate between individuals with low MMSE scores, those with scores between 24 and 27, and those with scores above 28, as well as healthy young participants. This suggests that ST4 can detect subtle changes in cognitive states, indicating its potential as a sensitive marker of cognitive functioning.

2.1.3 Power spectrum and frequency bands

The EEG power spectrum was obtained through the fast Fourier transform (FFT) of the EEG signals within a 4-second window, using a Hamming window to minimize spectral leakage. Power spectral density was calculated from the frontal channel (Fp1-Fp2) and transformed to dB (logarithm base 10), for Delta (0.5-4 Hz), Theta (4–7 Hz), Alpha (8–15 Hz), Beta (16–31 Hz), and lower Gamma (32–45 Hz) frequency bands.

Previous research has extensively explored the impact of cognitive load on various frequency bands, particularly within the frontal lobe. Results from EEG studies reveal enhanced frontal Theta activity in high cognitive load conditions, which is increasing with the growing demands on memory retention across various cognitive tasks like the n-back [53], [54], [55]. Additionally, studies have highlighted the significance of frontal Delta power in inhibiting potential interferences that might affect performance in high-load cognitive tasks [56]. Gamma activity exhibited positive correlations with fMRI-BOLD signal in various prefrontal cortex regions, indicating modulation during cognitive processing [57]. Middle-aged adults showed heightened frontal Gamma activity than young adults during the high cognitive load level of verbal n-back task[58]. Alongside this, reduced Gamma oscillations was observed in elderly subjects (mean age 75) compared to younger subjects [59], suggesting that Gamma activity increases with age until midlife, and starts to decline in older age. Similar to Gamma, Beta EEG activity has shown positive correlations with fMRI-BOLD signal in various frontal regions and exhibited a positive load effect specifically during cognitive working memory tasks [57]. In the prefrontal cortex, heightened Beta activity aids in information erasure from working memory, cessation of long-term memory retrieval, and preserves contents during delay periods [60]. Furthermore, while behavioral performance was similar between young and healthy elderly participants in an auditory memory task study, notable differences in Beta band desynchronization during retrieval suggest age-related influences on Beta responses during working memory task [61]. Understanding how cognitive load influences frequency bands in the frontal lobe contributes valuable insights into the neural mechanisms underlying cognitive processes and can shed light on cognitive decline.

2.1.4 EEG recording and auditory battery

EEG recording followed the previously described protocols [23], [28], lasting 20–30 minutes, including a 15-minute cognitive assessment battery. This battery consisted of pre-recorded tasks: musical detection, musical n-back, and resting state tasks as outlined in prior studies [23], [28]. In the first study, each patient was re-examined under the same conditions over the next seven days, with sessions at least one day apart. In the second study, patients participated in an additional EEG session involving auditory instructions and two C-IADL sub-tasks from PASS: telephone use and medication management. Each task is rated on a 4-point scale (0–3), and patients receive three types of scores: independence, safety, and adequacy (quality) [31].

Statistical Analysis

2.1.5 Overview

The statistical analysis was conducted separately for the first and second studies, followed by a meta-analysis incorporating data from a total of 349 participants from both studies and previously collected data. In the first study, the analysis began with dimensionality reduction using Lasso, Elastic Net, Ridge, and SVM with RBF kernel models to identify key features correlated with MMSE scores. This was followed by Linear Mixed Model (LMM) analyses to assess the relationships between EEG variables, MMSE groups, and cognitive load levels.

For the first study, the LMM model included the following variables: MMSE group (numeric, between), visit (categorical, within), and cognitive load (numeric, within). Separate LMMs were then conducted for each visit, considering MMSE group and cognitive load.

In the second study, LMM analyses incorporated the MMSE group (numeric, between) and cognitive load (numeric, within) variables. Additionally, correlation models were employed to examine the associations between EEG variables and clinical test scores. Logistic regression models were applied to predict both MMSE and MoCA results based on brain activity features and collected clinical data (e.g., CLOX, GDS, and PASS scales).

The significance level for all analyses was set at p < 0.05. Post-hoc effects with Benjamini-Hochberg correction [62] were applied following significant main effects and interactions. All analyses were carried out using Python Statsmodel [63].

2.1.6 Variables

These studies included EEG variables, performance data, and clinical scales. EEG variables comprised frequency bands: Delta, Theta, Alpha, Beta and lower Gamma, as well as three EEG features: VC9, ST4, and A0 (normalized to a scale of 0-100). All EEG variables were calculated every second using a moving window of four seconds, and mean activity per condition was analyzed. Behavioral variables included mean response accuracy and mean RTs per participant. The independent variable representing cognitive load was constructed as follows: tasks performed during rest were categorized as cog_load 0; Detection task level 1 and 0-back were categorized as cog_load 1; Detection task level 2 and 1-back were categorized as cog_load 2. Finally, 2-back was categorized as cog_load3.

Demographic and clinical results

To ensure proper adjustment for age and gender, mean ages were compared within each MMSE group using the Welch Two Sample t-test, both overall and separately by gender (see Tables 1 and 2 for detailed results).

In the first study, a significant positive correlation between MMSE and IADL scores was observed (r = 0.26, p = 0.03), as expected based on previous literature [64], [65]. Significant correlations were also found between MMSE and IADL scores and A0 biomarker activity during both cognitive and resting tasks (detection task: r = -0.25, p = 0.04; n-back task: r = -0.293, p = 0.02; and resting state tasks: r = -0.39, p = 0.003), suggesting that higher A0 activity might be associated with greater cognitive decline, as indicated by lower IADL and MMSE scores.

The second study included additional demographic and clinical data (see full details in Supplementary Material C). No significant differences in education level, years of employment, average sleep hours, sleep quality, or tiredness were found between groups (MD, MCI-R, and Healthy, p > 0.05). MoCA scores showed significant differences between all groups (all ps < 0.05, see Table 2).

Table 1

Demographic information for the first study groups, including mean ages and MMSE scores for total participants, and separately for males and females. The table also includes t and p values comparing mean ages between Healthy and MCI-R groups, both overall and by gender. Additionally, t and p values comparing ages between genders are shown in the final row.
	Groups	Healthy (MMSE ≥ 28)	MCI -R (MMSE 24–27)
Total	n	40	40
	MMSE	29.03 (0.8)	25.34 (1.02)
	Age	72.23 (9.63)	75.89 (10.66)
	Age t-test	Healthy vs MCI-R t=-1.76, p = 0.08
Male	n	18	22
	MMSE	29.08 (0.82)	25.59 (1.08)
	Age	69.01 (6.4)	75.64 (11.5)
	Age t-test	Healthy vs MCI -R t=-2.15, p = 0.03
Female	n	22	18
	MMSE	29 (0.79)	25.07 (0.87)
	Age	74.54 (10.83)	76.16 (9.61)
	Age t-test	Healthy vs MCI -R t=-0.7, p = 0.48
Age males vs. females		t=-1.81, p = 0.07	t=-0.39, p = 0.69

Table 2

Demographic information for the second study groups, including mean ages, MMSE scores, and MoCA scores for total participants, as well as for males and females separately. The table provides t and p values comparing mean ages between Healthy, MCI-R, and MD groups, both overall and by gender. The final row presents t and p values comparing ages between genders.
	Groups	Healthy (MMSE ≥ 28)	MCI -R (MMSE 24–27)	MD (MMSE < 24)
Total	MMSE scores	28–30	24–27	10–23
	n	30	30	17
	MMSE	28.76 (0.71)	25.51 (1.15)	20.22 (2.21)
	Age	73.3 (6.5)	72.05 (10.37)	79.68 (7.46)
	Age t-tests	Healthy vs MCI-R t = 0.57, p = 0.57	Healthy vs MD t=-3.02, p = 0.005	MCI -R vs MD t=-2.99, p = 0.004
	MoCA scores	23.53 (2.56)	19.34 (3.58)	12.96 (4.47)
Male	n	17	14	6
	MMSE	28.76 (0.73)	25.86 (1.19)	20.04 (1.97)
	Age	74 (7.85)	71.32 (10.94)	76 (8.76)
	Age t-tests	Healthy vs MCI -R t = 0.78, p = 0.43	Healthy vs MD t=-0.45, p = 0.65	MCI -R vs MD t=-0.98, p = 0.34
Female	n	13	16	11
	MMSE	28.76 (0.69)	25.19 (1.01)	20.31 (2.34)
	Age	72.38 (4.11)	72.69 (9.82)	81.84 (5.56)
	Age t-tests	Healthy vs MCI-R t = 0.13, p = 0.89	Healthy vs MD t=-4.63, p = 0.0002	MCI -R vs MD t=-3.09, p = 0.004
Age males vs. females		t = 0.70, p = 0.48	t=-0.40, p = 0.68	t=-1.46, p = 0.18

First study results

The first study aimed to detect early cognitive decline in healthy seniors. Initially, dimensionality reduction techniques (Lasso, Elastic Net, Ridge, and SVM with RBF kernel) were used to identify features correlated with MMSE scores. Subsequently, linear mixed models (LMM) were employed to examine relationships between EEG variables, MMSE groups, and cognitive load levels.

3.1.1 Dimensionality Reduction

To identify a combination of features that would result in the highest correlation with MMSE scores, mean feature activity as well as reaction times (RTs), and accuracy were calculated for each auditory task per participant. Since the focus was on detection of timely cognitive decline in the healthy elderly population (typically associated with MMSE > 24), the aim was to differentiate between healthy individuals (MMSE > 27) and those at risk for MCI (MMSE between 24 and 27).

Multiple linear predictors and one nonlinear predictor were tested, including ridge, Lasso, and Elastic regression, linear kernel RBF, and SVM with RBF kernel. Lasso and Elastic Net yielded slightly better results than ridge regression, indicating the usefulness of both L1 and L2 penalties in feature selection. We set the number of features to analyze at 30, based on individual R² values. The data was then analyzed using cross-validated binary prediction of MMSE scores. Each cross-validation group produced an ensemble average over multiple regularization parameters to improve reliability [66]. The average R² was 0.31, corresponding to an r > 0.55. See Table 3 for results from four models.

Table 3

Performance metrics for four predictive models used to detect early cognitive decline in healthy elderly population. STDs are presented in parentheses.
Model	Sensitivity	Specificity	Precision	F1 Score	AUC
Lasso	0.90 (0.02)	0.57 (0.022)	0.67 (0.012)	0.77 (0.012)	0.73 (0.014)
Elastic	0.90 (0.015)	0.58 (0.011)	0.68 (0.007)	0.77 (0.009)	0.74 (0.01)
LinRBF	0.86 (0.013)	0.60 (0)	0.68 (0.003)	0.76 (0.007)	0.73 (0.007)
Ridge	0.73 (0.024)	0.57 (0.012)	0.63 (0.012)	0.68 (0.016)	0.65 (0.015)

These results indicate that our approach effectively predicts cognitive performance as measured by MMSE scores, achieving a good balance between sensitivity and specificity. Specifically, Lasso and ElasticNet models achieved the highest sensitivity (0.90), indicating excellent detection of true positives. Both Lasso and ElasticNet models yielded the highest F1 score (0.77), indicating a strong balance between precision and sensitivity. Elastic Net achieved the highest AUC (0.74) with Lasso closely following (0.73), demonstrating superior overall ability to distinguish between classes. Figure 2 illustrates the predictions of these two models, showing correlation of r = 0.38 and r = 0.35, with lower variability in the higher MMSE scores (27–30).

3.1.2 LMM results

For the complete LMM results of all studies, including standard deviations, p- and z-values, refer to Supplementary Materials C. This study involved two recording sessions across consecutive visits, each featuring a comparable auditory battery with tasks of varying cognitive load. The initial LMM analysis included data from both visits, with MMSE score (numeric, between-subjects), visit (categorical, within-subjects), and cognitive load (numeric, within-subjects) as variables. No significant main effects or interactions were found for any of the features analyzed. Consequently, further analyses were conducted for each visit separately.

Analysis of the first visit data revealed no main effects between the groups. However, significant interactions between group and cognitive load were found for VC9, ST4, and Theta, with the healthy group showing higher activity at higher cognitive loads: cognitive load 2 vs. rest for VC9 (p = .014), ST4 (p = .016), and Theta (p = .028); and cognitive load 1 vs. rest for VC9 (p = .016) and ST4 (p = .018). No differences in cognitive load were detected in the MCI-R group.

In the second visit, A0 showed a significant main effect of group, with higher activity in the MCI-R group compared to the healthy group (p = .033). Additionally, VC9, Theta, Delta, Alpha and Beta exhibited significant main effects of cognitive load (all ps < .001), with similar cognitive load effects observed across both groups (see Fig. 3 and Supplementary Materials C).

3.1.3 Inter-patient variability results

Refer to supplementary material B for the full details and results. Testing reliability between the two visits for all EEG features and frequency bands revealed moderate to excellent reliability (ICCs 0.5–0.8) for the n-back task, and moderate to good reliability (ICCs 0.5–0.75) for the detection task. Findings from Pearson correlations revealed significant correlations between the two visits across both detection and n-back tasks for all EEG features and frequency bands (all p values < 0.01). In summary, the low within-patient variability observed between the two visits in the first study enhances the validity of our measurement method.

Second study results

The second study included a single recording session with cognitive assessments involving musical tasks of varying cognitive loads. To gain a deeper insight into participants' clinical status, additional clinical information and measurements were collected.

3.1.4 Correlation with clinical measures

Pearson correlations were calculated between each EEG feature per cognitive load, and the MMSE score and the MoCA score (full correlation results are provided in supplementary material C). A0 and Gamma demonstrated strong correlations with MMSE scores across all tasks, and with the MoCA scores for most tasks (see Fig. 4).

3.1.5 Mixed Linear Model (LMM) results

In the LMM model with group (3 levels, categorical, between), and cognitive load (3 levels, categorical, within) variables, A0, Gamma and Beta exhibited significant differences between the groups (see Fig. 5 for individual means of A0 and Gamma per group and cognitive load). Post-hoc analyses showed that the difference between Healthy and MD groups was significant for A0 (p_adj = 0.017) and Beta and p_adj = 0.002). For Gamma, the difference between Healthy and MD groups (p_adj = 0.001), as well as MCI-R and MD groups (p_adj = 0.0431) showed significance.

Main effect for cognitive load was significant for A0, VC9, Delta, Theta, Beta and Gamma. Post-hoc analysis revealed that for most features, the differences between cognitive load levels were highly significant for the Healthy and MCI-R groups, but not significant for the MD group (see Supplementary Materials C).

3.1.6 Logistic regression model results

To incorporate the clinical data gathered in the second study, logistic regression models were created to predict the MMSE scores based on EEG features and clinical data. This approach contrasts with the linear regression models used in the first analysis, which focused on healthy participants (with MMSE > 24), aiming to identify early signs of cognitive decline. The logistic regression approach here provides a broader understanding of cognitive impairment across a wider spectrum of MMSE scores (18–30) with multiple clinical measures included.

Two linear regression analyses were conducted to identify significant predictors of MMSE score. Both regressions included potential predictors from EEG features (i.e., A0, ST4, VC9, and Delta, Beta and Gamma), demographic factors (i.e., age, gender, years of education), cognitive task performance (i.e., accuracy, response time), and clinical measures (i.e., CLOX, GDS, and PASS scales). In the first regression, we tested all predictors across the different cognitive load levels, and the second regression was repeated for each cognitive load level. All regressions were created with a backward elimination process, first inserting all variables and then sequentially removing the non-significant variables based on their p-values (> 0.05). After backward elimination, the final first model predicting MMSE score across cognitive load levels, had an R² value of 0.988, with three significant predictors: A0 (p = 0.009), Gamma (p < 0.001), and accuracy (p < 0.001). For the full results and figures, see Supplementary Materials C.

To identify the factors influencing cognitive load, a series of regression models were constructed for each cognitive load level, using the same methos as the first regression. For the highest level of cognitive load level (i.e., brain activity during 2-back), seven significant predictors were retained: A0 (p < 0.001), Gamma (p < 0.001), ST4 (p = 0.046), CLOX (p = 0.015), PASS - drugs safety (p < 0.001), years of employment (p = 0.011), and tiredness (p = 0.019). For the mid-high cognitive load level (i.e., detection level 2 and the 1-back), the significant predictors included A0 (p = 0.001), Beta (p < 0.001), CLOX (p = 0.003), age (p = 0.04), PASS - drugs safety (p < 0.001), marital status (p = 0.002), years of employment (p < 0.001), living arrangements (p = 0.011), and tiredness (p = 0.004). In the low cognitive load level (i.e., 0-back and detection level 1), the significant variables were: A0 (p < 0.001), ST4 (p = 0.026), Beta (p < 0.001), PASS - drugs quality (p = 0.024), PASS - drugs safety (p < 0.001), marital status (p = 0.002), years of employment (p = 0.009), living arrangements (p = 0.020), tiredness (p = 0.018), and accuracy (p = 0.003). Interestingly, the resting state model, the variables who were found significant were the EEG features of VC9 (p < 0.001) and Theta (p < 0.001), and CLOX (p = 0.002), PASS - drugs safety (p = 0.003), marital status (p = 0.002), years of employment (p = 0.001), living arrangements (p = 0.000), tiredness (p = 0.018), and accuracy (p = 0.003) as clinical variables.

In conclusion, while each cognitive load level displayed a distinct set of significant predictors, there were shared factors such as A0 and years of employment consistently identified across models as significant. Conversely, certain variables like Gamma played a crucial role in specific cognitive load levels but did not demonstrate universal applicability across all levels.

Meta analysis

In the final stage of our analysis, we combined the data from both studies with previously collected data [23] that included seniors with different MMSE scores and a cohort of healthy young participants. This integration enabled a thorough meta-analysis, incorporating a total of 237 elderly individuals (categorized as healthy seniors n = 121, MCI-R n = 84, and MD n = 32), along with healthy young controls (n = 112). All participants completed similar tasks with the same levels of cognitive load, allowing for analysis of differences in mean brain activity between groups and across cognitive load levels, and their interactions. The population distributions of mean A0 activity levels during both rest and cognitive resource allocation were also computed and are provided in Supplementary Material A.

Initially, our focus was directed toward the elderly population (senior participants with a valid MMSE score, n = 203). Pearson correlations were calculated for each EEG feature, MMSE score and cognitive load level. A0 exhibited a significant correlation with MMSE score (r = -0.25, p < 0.001), which remained significant across all cognitive load levels (all ps < 0.001). Similarly, Gamma band exhibited significant correlation to MMSE score (r = -0.23, p < 0.001), maintaining significance across all cognitive load level (all ps < 0.01). Refer to Fig. 6 and Supplementary Materials C for the full results.

Next, we constructed an LMM that integrated the group variable (including all senior groups and healthy young controls), with the cognitive load levels (see Fig. 7 and Supplementary Materials C for all LMM and post-hoc results). Significant main effects of group were found for A0, Delta and Gamma, indicating lower activity levels the healthier and younger the group. Subsequent post-hoc comparisons revealed significant differences for A0 between the healthy young group and all other groups (all ps = 0.001), as well as between healthy seniors and the MD group (p = 0.001), and MCI-R group (p = 0.037). For Gamma, the MD group showed significantly higher activity compared to the healthy young group (p = 0.002). Delta showed a significant difference between the MCI-R and healthy young groups (p = 0.001).

An interaction between group and cognitive load was observed for VC9, ST4, Theta, Alpha, Beta and Gamma. Simple effect comparisons indicated that differences between cognitive load levels were generally more pronounced in cognitively healthier groups, with significant differences between cognitive load levels and rest for healthy seniors and healthy young participants (all ps < 0.001). Complete results are available in Supplementary Materials C.

Timely detection of cognitive decline is crucial for effective intervention, highlighted by the recent FDA approval of two new AD drugs [3], [4]. EEG serves as a valuable tool for identifying abnormal brain activity patterns that may indicate cognitive impairment. Our previously published pilot study [23] aimed to contribute to this objective by exploring neural activity using a single-channel EEG. The current paper introduces two follow-up studies that build upon these findings, extending their scope and broadening their applicability and relevance. Furthermore, this paper presents a meta-analysis combining data from all three studies, comprising 237 seniors and 112 healthy young subjects. An auditory assessment protocol was implemented to evaluate cognitive function under varying load conditions, facilitating a comprehensive exploration of EEG pattern changes to identify reliable biomarkers for timely detection.

In the first study, 80 cognitively healthy participants (MMSE > 24) were divided into two groups based on MMSE scores, with a cutoff score of 27, aiming to detect subtle changes associated with cognitive decline in the healthy elderly population. The Lasso regression model effectively selected relevant EEG and behavioral features, achieving a sensitivity of 0.90 for identifying individuals at-risk for MCI, and a specificity of 0.57, reflecting a moderate rate of correctly identifying non-MCI individuals. The model showed a moderate positive correlation between predicted and actual MMSE scores, accounting for 31% of the variance. This approach aligns with research showing Lasso regression can predict the relationship between working memory ability and frontal brain activity through EEG signal processing [67]. Another study used Lasso regression to select functional brain indicators associated with cognitive impairment, effectively classifying participants into groups based on MoCA and MMSE scores [68]. An additional study suggested that a predictive model for MMSE scores based on Lasso regression, highlighting the effectiveness of EEG biomarkers, particularly from the prefrontal regions, in indicating early cognitive decline [22]. These findings highlight the potential of this approach, though further refinement and additional variables may improve precision.

We then evaluated significant changes in EEG biomarker activity between the groups and cognitive load levels using LMMs, analyzing the two visits separately to explore changes over time. In the first visit, significant differences in cognitive loads for features VC9, ST4, and Theta were observed only in the Healthy group, with no differences in the MCI-R group. In the second visit, the Healthy group showed more pronounced differences for VC9 and Theta, and the MCI-R group displayed significant differences which were not apparent earlier. These findings suggest a potential learning effect or adaptation over time, consistent with the multiday learning curve approach, which indicates that assessing learning over multiple days can reveal early Aβ-related memory declines before conventional AD symptoms appear [69], [70]. Furthermore, A0 activity was significantly higher in the MCI-R group compared to the Healthy group in the second visit. Further, significant differences in load conditions were observed in the Healthy group but not in the MCI-R group, suggesting cognitive load effects are more pronounced in healthy individuals and may indicate a greater risk for MCI with lower initial MMSE scores. Theta and Delta bands revealed a pronounced increase during tasks that impose cognitive load compared to rest. This is consistent with our previous findings [23], [24] as well as recent literature regarding increased frontal activity of theta [55], [71], [72] and Delta [56], [73] during performance of cognitive demanding tasks. The comparison between visits revealed high consistency of within-patient variability, with significant Pearson correlations and moderate to excellent ICCs for all EEG features and bands. In summary, the first study demonstrated good intra-group consistency and notable inter-group variability withing healthy seniors.

The second study included 77 participants, divided into three groups based on MMSE scores: Healthy (MMSE > 27), MCI-R (MMSE 24–27), and MD (MMSE < 24). They completed similar cognitive tasks with single-channel EEG recording, as well as further clinical evaluations using MoCA and PASS. Previous research has demonstrated that education level can influence individual MMSE scores [29], [74]. One limitation identified in the pilot study was the lack of information regarding the education of the senior participants. This limitation was addressed in the second study, which revealed no significant differences between groups in terms of education levels, years of employment, sleep variables, and GDS scores. Education and other demographics were not significant predictors of MMSE scores, and their inclusion improved statistical models, enhancing the results related to the novel EEG biomarkers. Logistic regression identified significant predictors of MMSE scores across different cognitive load levels, integrating EEG features, demographics, cognitive task performance, and clinical assessments. A0 and Gamma activity consistently predicted cognitive function, while factors like CLOX, PASS drug safety, and years of employment, were significant only at certain load levels. These results highlight the nuanced interplay between EEG features, clinical measures, and cognitive performance, providing a robust framework for understanding cognitive decline. Results also indicated a significant negative correlations between Gamma band activity and both MMSE and MoCA scores. Specifically, lower MMSE and MoCA scores (indicative of greater cognitive impairments) were associated with increased Gamma activity during the performance of cognitive tasks. Although previous studies showed decreased Gamma band synchronization in AD [75], [76], the increased Gamma band power observed during task performance in cognitively healthy individuals [77], [78] persists even in cognitive decline patients, possibly indicating heightened resource allocation under cognitive load [13], [79]. Studying the effects of cognitive load on brain wave patterns can provide crucial insights into the processes underlying cognitive decline, enhancing our understanding of the mechanisms involved.

The significant negative correlation between A0 activity and MMSE scores observed in the pilot study [23] was replicated in both the second study and the meta-analysis. This correlation was also extended to MoCA and IADL scores, fulfilling the primary objective of the second study. These associations to clinical measures further validate A0 as a biomarker related to cognitive state as previously described [23], [28]. A0 also demonstrated the capacity to differentiate between study groups in both studies and the meta-analysis. The MCI-R group demonstrated significantly increased A0 activity compared to the Healthy group in the first study (visit 2). This was replicated in the second study, with the edition of the MD group showing significantly higher A0 activity levels. Similarly, in the meta-analysis, differences in A0 among all groups were statistically significant (except for the comparison between the MCI-R and MD groups). These findings not only replicate but also expand upon previous results, thus achieving the primary goal of successfully identifying distinctions between cognitively healthy individuals at risk of decline (who initial scored lower with MMSE scores between 27 and 24) and healthy seniors (with MMSE over 28). These results of A0 across multiple studies and a meta-analysis, suggest its potential as a reliable biomarker for timely detection of cognitive decline.

One of the key findings from the pilot study was increased EEG activity observed with higher cognitive load levels, with more cognitively healthy group [23]. In the second study, the most significant differences in A0 across cognitive load levels were observed within the Healthy group, while the MCI-R group showed significant differences only between rest and the high cognitive load condition. No significant variations in A0 were observed in the MD group across cognitive load levels. In the meta-analysis including the healthy young participants, VC9 showed lower differences between cognitive load levels and resting state the less cognitively healthy the group was. Moreover, the MCI-R group exhibits significant differences between lower cognitive load levels (cog load1 vs. cog load2), showing lower activity for the higher load condition, a pattern not observed in the two healthy groups. This suggests that while there is an initial response to increased cognitive load in this group, activity levels plateau with further increases in load. This observation aligns with findings that seniors with cognitive decline show heightened activity during lower cognitive loads but struggle to sustain this activation as task demands increase [80], [81].

While showing promising results, further research is needed to address limitations encountered in our studies. For instance, the small sample size of the MD group (n = 32), challenges robust comparisons with the larger MCI-R (n = 84) and Healthy (n = 121) groups. Future studies should include a larger sample of MD patients to enhance statistical power and enable more comprehensive analyses and interpretations of differences between cognitive states. In future research, the significant EEG variables identified in the logistic regression model could be utilized to predict MMSE and MoCA scores of elderly participants, allowing for comparison with actual clinical assessment scores to assess their predictive power. In the second study, the PASS drugs safety score emerged as a key predictor for MMSE scores in logistic regression models across cognitive load levels. Despite its promise in assessing functional competence and distinguishing between subjects with cognitive decline and healthy controls [32], our findings did not reveal significant differences between study groups or correlations with EEG feature activity during PASS performance. Future investigations could explore alternative PASS sub-tasks, such as shopping or checkbook balancing, known for their robust discriminative capabilities [82]. Furthermore, while this paper focuses on the timely detection of cognitive decline, long-term studies could provide deeper insights into the predictive power of our biomarkers. Tracking individuals at risk for MCI over time could reveal how early biomarkers relate to the actual development of cognitive impairment, enhancing understanding of disease progression and potential early intervention.

In summary, this paper highlights the effectiveness of EEG biomarkers in detecting cognitive function among healthy elderly individuals. The integration of additional diagnostic tools and identification of key predictors further enhances our understanding of cognitive impairment. We demonstrated the capability of EEG features, particularly A0 activity, to distinguish between cognitively healthy individuals and those at risk. Collectively, our findings underscore the potential of EEG features as a non-invasive, cost-effective and reliable approach for better understanding cognitive decline and facilitating timely diagnosis to improve clinical outcomes.

Data availability

The datasets generated during and/or analyzed during the current study are not publicly available due to ethical and privacy restrictions but are available from the corresponding author on reasonable request.

Acknowledgements

The authors express heartfelt thanks extended to the study participants and the supportive staff for their contributions to this research.

Author contributions

Conception and study design L.M., N.B.M, T.Z. and N.I.; Data acquisition; T.Z., O.C., S.R, V.A and N.B.O; Supervision N.I. and A.S; Data analysis and Writing L.M., N.B.M, T.Z., O.C., and N.I; All authors read and approved the final manuscript.

Competing interests

L.M., N.B.M., and N.I. have equity interest in Neurosteer, which developed the Neurosteer EEG recorder. T.Z and O.C. are employed in Neurosteer.

Grand, J. H. G., Caspar, S. & MacDonald, S. W. S. Clinical features and multidisciplinary approaches to dementia care. J. Multidiscip Healthc. 4 10.2147/JMDH.S17773 (2011).
Zihl, J. & Reppermund, S. ‘The aging mind: A complex challenge for research and practice’, Aging Brain, vol. 3, doi: (2023). 10.1016/j.nbas.2022.100060
Budd Haeberlein, S. et al. Two Randomized Phase 3 Studies of Aducanumab in Early Alzheimer’s Disease. J. Prev. Alzheimer’s Disease. 9 (2). 10.14283/jpad.2022.30 (2022).
Salloway, S. et al. Amyloid-Related Imaging Abnormalities in 2 Phase 3 Studies Evaluating Aducanumab in Patients with Early Alzheimer Disease. JAMA Neurol. 79 (1). 10.1001/jamaneurol.2021.4161 (2022).
‘Lecanemab in Early Alzheimer’s Disease’. N. Engl. J. Med., 388, 17, doi: 10.1056/nejmc2301380. (2023).
Sperling, R. A. et al. Toward defining the preclinical stages of Alzheimer’s disease: Recommendations from the National Institute on Aging-Alzheimer’s Association workgroups on diagnostic guidelines for Alzheimer’s disease. Alzheimer’s Dement. 7 (3). 10.1016/j.jalz.2011.03.003 (2011).
Van Heugten, C. M., Walton, L. & Hentschel, U. Can we forget the Mini-Mental State Examination? A systematic review of the validity of cognitive screening instruments within one month after stroke. Clin. Rehabil. 29 (7). 10.1177/0269215514553012 (2015).
Dong, Y. et al. The Montreal Cognitive Assessment (MoCA) is superior to the Mini-Mental State Examination (MMSE) for the detection of vascular cognitive impairment after acute stroke. J. Neurol. Sci. 299 (1–2). 10.1016/j.jns.2010.08.051 (2010).
Cassani, R., Estarellas, M., San-Martin, R., Fraga, F. J. & Falk, T. H. ‘Systematic review on resting-state EEG for Alzheimer’s disease diagnosis and progression assessment’, 2018. 10.1155/2018/5174815
Dauwels, J., Vialatte, F. & Cichocki, A. Diagnosis of Alzheimer’s Disease from EEG Signals: Where Are We Standing? Curr. Alzheimer Res. 7 (6), 487–505. 10.2174/1567210204558652050 (2010).
Hamm, V., Héraud, C., Cassei, J. C., Mathis, C. & Goutagny, R. ‘Precocious alterations of brain oscillatory activity in Alzheimer’s disease: A window of opportunity for early diagnosis and treatment’, 2015. 10.3389/fncel.2015.00491
Stam, C. J. et al. Generalized synchronization of MEG recordings in Alzheimer’s disease: Evidence for involvement of the gamma band. J. Clin. Neurophysiol. 19 (6). 10.1097/00004691-200212000-00010 (2002).
Van Deursen, J. A., Vuurman, E. F. P. M., Verhey, F. R. J., Van Kranen-Mastenbroek, V. H. J. M. & Riedel, W. J. Increased EEG gamma band activity in Alzheimer’s disease and mild cognitive impairment. J. Neural Transm. 115 (9). 10.1007/s00702-008-0083-y (2008).
Jiang, Z. Study on EEG power and coherence in patients with mild cognitive impairment during working memory task. J. Zhejiang Univ. Sci. B. 6 (12). 10.1631/jzus.2005.B1213 (2005).
Başar, E., Başar-Eroǧlu, C., Güntekin, B. & Yener, G. G. ‘Brain’s alpha, beta, gamma, delta, and theta oscillations in neuropsychiatric diseases: Proposal for biomarker strategies’, in Supplements to Clinical Neurophysiology, vol. 62, doi: (2013). 10.1016/B978-0-7020-5307-8.00002-8
Güntekin, B., Saatçi, E. & Yener, G. Decrease of evoked delta, theta and alpha coherences in Alzheimer patients during a visual oddball paradigm. Brain Res. 1235 10.1016/j.brainres.2008.06.028 (2008).
Al-Qazzaz, N. K. et al. ‘Role of EEG as biomarker in the early detection and classification of dementia’, 2014. 10.1155/2014/906038
Samal, P. & Hashmi, M. F. Role of machine learning and deep learning techniques in EEG-based BCI emotion recognition system: a review. Artif. Intell. Rev. 57 (3). 10.1007/s10462-023-10690-2 (2024).
Grueso, S. & Viejo-Sobera, R. Machine learning methods for predicting progression from mild cognitive impairment to Alzheimer’s disease dementia: a systematic review. Alzheimers Res. Ther. 13 (1). 10.1186/s13195-021-00900-w (2021).
Modir, A., Shamekhi, S. & Ghaderyan, P. ‘A systematic review and methodological analysis of EEG-based biomarkers of Alzheimer’s disease’, 2023. 10.1016/j.measurement.2023.113274
Mitsukura, Y., Sumali, B., Watanabe, H., Ikaga, T. & Nishimura, T. Frontotemporal EEG as potential biomarker for early MCI: a case–control study. BMC Psychiatry. 22 (1). 10.1186/s12888-022-03932-0 (2022).
Choi, J. et al. Resting-state prefrontal EEG biomarkers in correlation with MMSE scores in elderly individuals. Sci. Rep. 9 (1), 10468. 10.1038/s41598-019-46789-2 (2019).
Molcho, L. et al. Single-Channel EEG Features Reveal an Association With Cognitive Decline in Seniors Performing. Auditory Cogn. Assessment’. 10.3389/fnagi.2022.773692 (2022).
Maimon, N. B., Molcho, L., Intrator, N. & Lamy, D. ‘Single-channel EEG features during n-back task correlate with working memory load’, arXiv preprint, no. arXiv:2008.04987, Aug. 2020, Accessed: Oct. 06, 2020. [Online]. Available: http://arxiv.org/abs/2008.04987
Maimon, N. B. et al. ‘Continuous monitoring of mental load during virtual simulator training for laparoscopic surgery reflects laparoscopic dexterity. A comparative study using a novel wireless device’. Front. Neurosci., p. 1716, (2021).
Bolton, F., Te’Eni, D., Maimon, N. B. & Toch, E. ‘Detecting interruption events using EEG’, in IEEE 3rd Global Conference on Life Sciences and Technologies (LifeTech), IEEE, Mar. 2021, pp. 33–34. doi: (2021). 10.1109/LifeTech52111.2021.9391915
Curcic, J. et al. Description of the Method for Evaluating Digital Endpoints in Alzheimer Disease Study: Protocol for an Exploratory, Cross-sectional Study. JMIR Res. Protoc. 11 (8). 10.2196/35442 (2022).
Molcho, L. et al. Evaluation of Parkinson’s disease early diagnosis using single-channel EEG features and auditory cognitive assessment. Front. Neurol. 14 10.3389/fneur.2023.1273458 (2023).
Siqueira, G. S. A. et al. Can MoCA and MMSE Be Interchangeable Cognitive Screening Tools? Syst. Review’. 10.1093/geront/gny126 (2019).
Khatun, S., Morshed, B. I. & Bidelman, G. M. ‘Single Channel EEG Based Score Generation to Monitor the Severity and Progression of Mild Cognitive Impairment’, in IEEE International Conference on Electro Information Technology, doi: (2018). 10.1109/EIT.2018.8500273
Holm, M. B. & Rogers, J. C. ‘The Performance Assessment of Self-Care Skills (PASS)’. Assessments Occup. Therapy Mental Health, (2008).
Dham, P. et al. Functional Competence and Cognition in Individuals With Amnestic Mild Cognitive Impairment. J. Am. Geriatr. Soc. 68 (8). 10.1111/jgs.16454 (2020).
Sachdev, P. S. et al. Risk profiles for mild cognitive impairment vary by age and sex: The sydney memory and ageing study. Am. J. Geriatric Psychiatry. 20 (10). 10.1097/JGP.0b013e31825461b0 (2012).
Ritchie, K. & Lovestone, S. ‘The dementias’, in Lancet, doi: (2002). 10.1016/S0140-6736(02)11667-9
Yaffe, K. et al. Mild cognitive impairment, dementia, and their subtypes in oldest old women. Arch. Neurol. 68 (5). 10.1001/archneurol.2011.82 (2011).
Crum, R. M., Anthony, J. C., Bassett, S. S. & Folstein, M. F. Population-Based Norms for the Mini-Mental State Examination by Age and Educational Level. JAMA: J. Am. Med. Association. 269 (18), 2386–2391. 10.1001/jama.1993.03500180078038 (1993).
O’Bryant, S. E. et al. Detecting dementia with the mini-mental state examination in highly educated individuals. Arch. Neurol. 65 (7), 963–967. 10.1001/archneur.65.7.963 (2008).
Graf, C. The lawton instrumental activities of daily living scale. Am. J. Nurs. 108 (4). 10.1097/01.NAJ.0000314810.46029.74 (2008).
Ou, Y. Y. et al. ‘Instrumental activities of daily living (IADL) evaluation system based on EEG signal feature analysis’, in 2013 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference, APSIPA 2013, doi: (2013). 10.1109/APSIPA.2013.6694310
Hobson, J. ‘The Montreal Cognitive Assessment (MoCA)’, doi: (2015). 10.1093/occmed/kqv078
Greenberg, S. A. ‘The geriatric depression scale (GDS) validation of a geriatric depression screening scale: A preliminary report’, Best Practices in Nursing Care to Older Adults, no. 4, (2019).
Royall, D. R., Cordes, J. A. & Polk, M. CLOX: An executive clock drawing task. J. Neurol. Neurosurg. Psychiatry. 64 (5). 10.1136/jnnp.64.5.588 (1998).
Jeong, J. ‘EEG dynamics in patients with Alzheimer’s disease’, 2004. 10.1016/j.clinph.2004.01.001
Nimmy John, T., Subha Dharmapalan, P. & Ramshekhar Menon, N. Exploration of time-frequency reassignment and homologous inter-hemispheric asymmetry analysis of MCI-AD brain activity. BMC Neurosci. 20 (1). 10.1186/s12868-019-0519-3 (2019).
Bibina, V. C., Chakraborty, U., Regeena, M. L. & Kumar, A. ‘Signal processing methods of diagnosing Alzheimer’s disease using EEG a technical review’. Int. J. Biology Biomedical Eng., 12, (2018).
Hastie, T., Buja, A. & Tibshirani, R. ‘Penalized Discriminant Analysis’, The Annals of Statistics, vol. 23, no. 1, doi: (2007). 10.1214/aos/1176324456
Rizk-Jackson, A. et al. Early Indications of Future Cognitive Decline: Stable versus Declining Controls. PLoS One. 8 (9). 10.1371/journal.pone.0074062 (2013).
Rokhlin, V., Szlam, A. & Tygert, M. A randomized algorithm for principal component analysis. SIAM J. Matrix Anal. Appl. 31 (3), 1100–1124. 10.1137/080736417 (2009).
Meghdadi, A. H. et al. February,., ‘Resting state EEG biomarkers of cognitive decline associated with Alzheimer’s disease and mild cognitive impairment’, PLoS One, vol. 16, no. 2 doi: (2021). 10.1371/journal.pone.0244180
López, M. M. et al. ‘SVM-based CAD system for early detection of the Alzheimer’s disease using kernel PCA and LDA’, Neurosci Lett, vol. 464, no. 3, doi: (2009). 10.1016/j.neulet.2009.08.061
Choi, H. & Jin, K. H. Predicting cognitive decline with deep learning of brain metabolism and amyloid imaging. Behav. Brain. Res. 344 10.1016/j.bbr.2018.02.017 (2018).
Maimon, N. B. et al. ‘EEG reactivity changes captured via mobile BCI device following tDCS intervention–a pilot-study in disorders of consciousness (DOC) patients’, in 10th International Winter Conference on Brain-Computer Interface (BCI), IEEE, Feb. pp. 1–3. (2022).
Krause, C. M. et al. ‘The effects of memory load on event-related EEG desynchronization and synchronization’, Clinical Neurophysiology, vol. 111, no. 11, doi: (2000). 10.1016/S1388-2457(00)00429-6
Raghavachari, S. et al. Gating of human theta oscillations by a working memory task. J. Neurosci. 21 (9). 10.1523/jneurosci.21-09-03175.2001 (2001).
Jensen, O. & Tesche, C. D. ‘Frontal theta activity in humans increases with memory load in a working memory task’, European Journal of Neuroscience, vol. 15, no. 8, pp. 1395–9, doi: (2002). 10.1046/j.1460-9568.2002.01975.x
Harmony, T. ‘The functional significance of delta oscillations in cognitive processing’, 2013. 10.3389/fnint.2013.00083
Michels, L. et al. Simultaneous EEG-fMRI during a working memory task: Modulations in low and high frequency bands. PLoS One. 5 (4). 10.1371/journal.pone.0010298 (2010).
Bair, M. S. et al. Age-related differences in working memory evoked gamma oscillations. Brain Res. 1576 10.1016/j.brainres.2014.05.043 (2014).
Missonnier, P. et al. Aging and working memory: Early deficits in EEG activation of posterior cortical areas. J. Neural Transm. 111 (9). 10.1007/s00702-004-0159-2 (2004).
Schmidt, R. et al. Beta oscillations in working memory, executive control of movement and thought, and sensorimotor function. J. Neurosci. 10.1523/JNEUROSCI.1163-19.2019 (2019).
Karrasch, M., Laine, M., Rapinoja, P. & Krause, C. M. Effects of normal aging on event-related desynchronization/synchronization during a memory task in humans. Neurosci. Lett. 366 (1). 10.1016/j.neulet.2004.05.010 (2004).
Benjamini, Y. & Hochberg, Y. ‘Controlling the False Discovery Rate - a Practical and Powerful Approach to Multiple Testing. (1995). Journal of the Royal Statistical Society Series B-Methodological 1995.pdf’.
Seabold, J. & Perktold, S. S., Econometric and statistical modeling with python’, in In Proceedings of the 9th Python in Science Conference, p. 61. (2010).
Cahn-Weiner, D. A., Malloy, P. F., Boyle, P. A., Marran, M. & Salloway, S. ‘Prediction of functional status from neuropsychological tests in community-dwelling elderly individuals’, Clinical Neuropsychologist, vol. 14, no. 2, doi: (2000). 10.1076/1385-4046(200005)14:2;1-Z;FT187.
Lopez, O. L. et al. Neuropsychological characteristics of mild cognitive impairment subgroups. J. Neurol. Neurosurg. Psychiatry. 77 (2). 10.1136/jnnp.2004.045567 (2006).
Naftaly, U., Intrator, N. & Horn, D. ‘Optimal ensemble averaging of neural networks’, Network: Computation in Neural Systems, vol. 8, no. 3, doi: (1997). 10.1088/0954-898x/8/3/004
Zhang, Y. et al. Prediction of working memory ability based on EEG by functional data analysis. J. Neurosci. Methods. 333 10.1016/j.jneumeth.2019.108552 (2020).
Liu, Y. et al. Classification of cognitive impairment in older adults based on brain functional state measurement data via hierarchical clustering analysis. Front. Aging Neurosci. 15 10.3389/fnagi.2023.1198481 (2023).
Jutten, R. J. et al. ‘Longitudinal multi-day learning curves (MDLCs) to capture subtle cognitive changes in preclinical Alzheimer’s disease’, Alzheimer’s & Dementia, vol. 19, no. S18, doi: (2023). 10.1002/alz.078818
Papp, K. V. et al. ‘Early Detection of Amyloid-Related Changes in Memory among Cognitively Unimpaired Older Adults with Daily Digital Testing’, Ann Neurol, vol. 95, no. 3, doi: (2024). 10.1002/ana.26833
Krause, C. M. et al. ‘The effects of memory load on event-related EEG desynchronization and synchronization’, Clinical Neurophysiology, vol. 111, no. 11, doi: (2000). 10.1016/S1388-2457(00)00429-6
Onton, J., Delorme, A. & Makeig, S. ‘Frontal midline EEG dynamics during working memory’, Neuroimage, vol. 27, no. 2, doi: (2005). 10.1016/j.neuroimage.2005.04.014
Schmiedt-Fehr, C., Dühl, S. & Basar-Eroglu, C. Age-related increases in within-person variability: Delta and theta oscillations indicate that the elderly are not always old. Neurosci. Lett. 495 (2). 10.1016/j.neulet.2011.03.062 (2011).
Ardila, A., Ostrosky-Solis, F., Rosselli, M. & Gómez, C. Age-related cognitive decline during normal aging: The complex effect of education. Arch. Clin. Neuropsychol. 15 (6). 10.1016/S0887-6177(99)00040-2 (2000).
Koenig, T. et al. Decreased EEG synchronization in Alzheimer’s disease and mild cognitive impairment. Neurobiol. Aging. 26 (2). 10.1016/j.neurobiolaging.2004.03.008 (2005).
Stam, C. J. et al. ‘Magnetoencephalographic evaluation of resting-state functional connectivity in Alzheimer’s disease’, Neuroimage, vol. 32, no. 3, doi: (2006). 10.1016/j.neuroimage.2006.05.033
Park, J. Y. et al. Gamma oscillatory activity in relation to memory ability in older adults. Int. J. Psychophysiol. 86 (1). 10.1016/j.ijpsycho.2012.08.002 (2012).
Fitzgibbon, S. P., Pope, K. J., MacKenzie, L., Clark, C. R. & Willoughby, J. O. ‘Cognitive tasks augment gamma EEG power’, Clinical Neurophysiology, vol. 115, no. 8, doi: (2004). 10.1016/j.clinph.2004.03.009
Osipova, D., Pekkonen, E. & Ahveninen, J. Enhanced magnetic auditory steady-state response in early Alzheimer’s disease. Clin. Neurophysiol. 117 (9). 10.1016/j.clinph.2006.05.034 (2006).
Schneider-Garces, N. J. et al. Span, CRUNCH, and beyond: Working memory capacity and the aging brain. J. Cogn. Neurosci. 22 (4). 10.1162/jocn.2009.21230 (2010).
Cappell, K. A., Gmeindl, L. & Reuter-Lorenz, P. A. ‘Age differences in prefontal recruitment during verbal working memory maintenance depend on memory load’, Cortex, vol. 46, no. 4, doi: (2010). 10.1016/j.cortex.2009.11.009
Rodakowski, J. et al. Can performance on daily activities discriminate between older adults with normal cognitive function and those with mild cognitive impairment? J. Am. Geriatr. Soc. 62 (7). 10.1111/jgs.12878 (2014).

Competing interest reported. L.M., N.B.M., and N.I. have equity interest in Neurosteer, which developed the Neurosteer EEG recorder. T.Z and O.C. are employed in Neurosteer.

Download PDF

Reviews received at journal
12 Nov, 2024
Reviews received at journal
11 Nov, 2024
Reviewers agreed at journal
04 Nov, 2024
Reviewers agreed at journal
02 Nov, 2024
Reviewers agreed at journal
02 Nov, 2024
Reviewers invited by journal
02 Nov, 2024
Editor assigned by journal
12 Oct, 2024
Editor invited by journal
09 Oct, 2024
Submission checks completed at journal
08 Oct, 2024
First submitted to journal
20 Sep, 2024

You are reading this latest preprint version

Evaluating Cognitive Decline Detection in Aging Populations with Single-Channel EEG Features: Insights from Studies and Meta-Analysis

Status:

Version 1

Abstract

Figures

1 Introduction

2 Methods

2.1.1 Study groups

2.1.2 Signal processing and high-level features

2.1.3 Power spectrum and frequency bands

2.1.4 EEG recording and auditory battery

2.1.5 Overview

2.1.6 Variables

3 Results

3.1.1 Dimensionality Reduction

3.1.2 LMM results

3.1.3 Inter-patient variability results

3.1.4 Correlation with clinical measures

3.1.5 Mixed Linear Model (LMM) results

3.1.6 Logistic regression model results

4 Discussion

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1