Alzheimer’s disease (AD) is a neurodegenerative disorder that accounts for nearly 70% of the more than 46 million dementia cases estimated worldwide. Although there is no cure for AD, early diagnosis and an accurate characterization of the disease progression can improve the quality of life of AD patients and their caregivers. Currently, AD diagnosis is carried out using standardized mental status examinations, which are commonly assisted by expensive neuroimaging scans and invasive laboratory tests, thus rendering the diagnosis time consuming and costly. Notwithstanding, over the last decade, electroencephalography (EEG) has emerged as a noninvasive alternative technique for the study of AD, competing with more expensive neuroimaging tools, such as MRI and PET. This paper reports on the results of a systematic review on the utilization of resting-state EEG signals for AD diagnosis and progression assessment. Recent journal articles obtained from four major bibliographic databases were analyzed. A total of 112 journal articles published from January 2010 to February 2018 were meticulously reviewed, and relevant aspects of these papers were compared across articles to provide a general overview of the research on this noninvasive AD diagnosis technique. Finally, recommendations for future studies with resting-state EEG were presented to improve and facilitate the knowledge transfer among research groups.
The term dementia is used to characterize several neurodegenerative disorders caused by damage and death of neurons, provoking a disturbance of cognitive and behavioral functions. Among the different forms of dementia, Alzheimer’s disease (AD) is the most common, accounting for nearly 70% of the dementia cases worldwide. It mostly affects people over 65 years of age and the rate of incidence grows exponentially with age [
In 2015, 46 million people were diagnosed with dementia worldwide and this number is projected to grow to 66 million by 2030, and to 115 million by 2050 [
According to symptomatology, AD has been divided in three stages: preclinical, mild cognitive impairment, and dementia due to AD [
The pathophysiological process of AD is thought to start up to 20 years before clinical symptoms can be detectable [
Accurate diagnosis is a true challenge, as AD pathophysiological processes may start up to 20 years before clinical symptoms can be detectable [
Today, definite AD diagnosis is only possible postmortem when analysis reveals the structural brain damage characteristic of the disease. Typically, accuracies up to 90% have been reported with current diagnosis methods, such as neurological tests and medical records. The current clinical diagnostic criteria for AD were developed by the National Institute on Aging and the Alzheimer’s Association (NIA-AA) [
Relying on neurological tests and the evaluation of medical records require experienced clinicians and lengthy sessions, rendering AD diagnosis irreproducible and time consuming. In response to these drawbacks, in the last few years, there has been an increase in the use, research, and development of biomarkers [
EEG is a technique that consists of recording the changes in time of the electrical activity in the cerebral cortex, produced by postsynaptic potentials from thousands of neurons with similar spatial orientation. These electric potentials are measured by electrodes placed on the scalp. The spatial resolution of EEG is related to the number of electrodes used and their placement, or layout, on the scalp. The most utilized layout is the international 10-20 system, commonly consisting of 21 electrodes; higher density variants of the 10-20 system such as 10-10 and 10-5 systems are utilized as well, usually with 64 and 128 electrodes, respectively, [
Since EEG signals reflect functional changes in the cerebral cortex, EEG-based biomarkers can be used to assess neuronal degeneration caused by AD progression (biomarkers in category N according to [
Over the last decades, many studies have investigated the effects of AD and its progression on EEG signals. Studies have made use of EEG signals under diverse recording conditions, which can be divided into two major groups:
In the literature on AD, recent reviews have been published covering the use of event-related EEG for AD diagnosis [
Regarding resting-state analysis for AD diagnosis, some recent reviews have also been written. However, none of these have treated exclusively the specific topic of EEG-based AD diagnosis. For instance, some reviews do not study EEG as the main technique for diagnosis [
This systematic review will focus on recent studies on resting-state EEG for AD diagnosis, describing and comparing the crucial stages in EEG-based AD diagnosis, such as EEG signal acquisition, preprocessing, artifact handling, and feature extraction and classification. Moreover, pointing out common practices, differences and consensus in the utilization of resting-state EEG reported limitations and recommendations for several experimental stages ranging from population characteristics to results reporting for future studies. We hope that this review will boost the research on this topic, leading to more reliable EEG resting-state AD diagnosis techniques. The remainder of this article is organized as follows. Section
A survey on English peer-reviewed journal articles published between January 2010 and February 2018 was performed for this review. Four major bibliographic databases were queried, namely, PubMed, Web of Science, IEEE Xplore, and Scopus, using the following search terms:
Electroencephalogr Alzheimer Diagnos
These search terms were combined in the following rule: (1 OR 2) AND 3 AND 4. Resulting journal articles were selected or rejected based on the criteria presented in Table
Eligibility criteria.
Studies using EEG to assess AD progression |
Studies using EEG to AD diagnosis |
Studies using EEG to perform differential diagnosis between AD and other dementias |
Studies on AD-related epilepsy |
Studies without resting-state EEG recordings |
Studies focused on dementias other than AD |
Studies focused on the effects of AD treatment drugs |
Studies on animals (nonhuman studies) |
Studies not treating MCI as a prodromal stage for AD |
Review articles |
Along with the mentioned search terms, studies that used other modalities besides EEG were further analyzed and divided in two types: (i) those that used features from other modalities combined with EEG features for AD diagnosis and (ii) those that used other modalities to verify and compare the results obtained with only EEG. The former papers were excluded, as we want to focus only on EEG-based diagnosis of AD; the latter were included.
Eligibility assessment was performed by at least two independent researchers by reading the article title and, when the article title did not provide enough information to be selected or rejected, the abstract was also read. In the cases where the assessors independently disagreed on the inclusion or exclusion of a paper, the final decision was made after a discussion between the two. Lastly, some articles were rejected after careful reading of the papers when it became clear that it did not meet the inclusion criteria. In order to keep track of the relevant information while reading the articles, a data extraction sheet was developed. For each article selected, 21 data items were extracted and grouped into five categories: study rationale, study population, experiment setup, EEG processing, and reported outcomes.
The category
Extracted items from each article.
Category | Data item | Description |
---|---|---|
Study rationale | Study goal | Application or aim of the article |
Other dementias | Differential diagnosis of different types of dementias with respect to AD | |
Study population | Sample size | Size of the population in the study |
Group matching | Groups matched (or not) by sample size, age, gender, and education level | |
Following of MCI participants | Follow-up of MCI participants, when required | |
Experiment setup | Other modalities | Other modalities utilized beside EEG |
Number of electrodes and layout | Electrode number and positioning system | |
External channels | Report the acquisition (or not) of EOG and ECG signals | |
Resting-state recording state | EEG recorded only in resting state or with task performing too | |
Experiment duration | Session duration of each experiment | |
EEG processing | Preprocessing | Survey on preprocessing techniques |
EEG bandwidth | Bandpass filtering of EEG signal and type of filters used | |
Artifact handling | Artifact rejection and/or correction methods | |
Effective sampling frequency | Sampling frequency of EEG data for feature extraction | |
EEG epoching | Epoching process, length, and quantity of epochs | |
Effective EEG signal duration | Length of EEG signal used for feature extraction | |
Source localization | Survey on source localization methods when required | |
EEG feature types | Survey on the types of EEG features used | |
Reported outcomes | Discriminative studies | Methods for discriminative task and reported results |
Assessment studies | Methods for assessment task and reported results | |
Reported limitations | Limitations reported in the study |
This review was written following the PRISMA statement scheme for reporting systematic reviews [
A total of 921 journal articles were found in the database queries, with 714 unique records remaining after duplicates were removed. Through title and abstract screening, 289 and 158 articles were rejected respectively, as they did not meet the inclusion criteria. A total of 267 articles met all the previously established inclusion criteria. After full-text examination, only 112 articles were included in the systematic review. Figure
Diagram showing the selection process of articles from PubMed, IEEE Xplore, Web of Science, and Scopus.
Distribution of selected articles according to world regions.
Number of reviewed articles by publication year.
According to the reported aim of the articles, two major goals were identified: (1) discriminative (or diagnosis), i.e., explore the difference in EEG-based features among populations, MCI, mild AD, severe AD, other types of dementias, and healthy normal elderly controls (Nold) and (2) progression assessment, i.e., find correlates between EEG-based features and clinical markers related to the MCI-to-AD conversion and AD severity progression. The majority (72) fell exclusively in the diagnosis category, whereas 18 articles were included in the progression assessment category and 22 studies were double aimed. The articles belonging to each study goal and populations investigated are presented in Table
Study goal description.
Study type | Study goal | Articles |
---|---|---|
Diagnosis (72) | AD vs Nold (48) | [ |
MCI vs AD (2) | [ | |
MCI vs Nold (4) | [ | |
AD vs Nold vs others (6) | [ | |
AD vs MCI vs Nold (12) | [ | |
Progression assessment (18) | AD (3) | [ |
AD vs Nold (1) | [ | |
MCI vs AD (3) | [ | |
MCI (11) | [ | |
Diagnosis and progression assessment (22) | AD vs Nold (11) | [ |
AD vs MCI vs Nold (4) | [ | |
AD vs Nold vs others (2) | [ | |
AD vs others (1) | [ | |
MCI vs AD (3) | [ |
As mentioned previously, dementia is a term that involves different disorders and diseases, one of them is AD, which accounts for the great majority of dementia cases. Having similar symptoms, around 10% of dementia cases are difficult to diagnose with reasonable confidence and it is not uncommon in clinical practice to mix dementia diagnoses [
Combination of AD diagnosis with other dementias.
Type of dementia | Articles |
---|---|
VaD | [ |
FtD/FTLD | [ |
DLB | [ |
PDD | [ |
PDD/DLB | [ |
PDD/DLB/FtD | [ |
PDD/DLB/FtD/VaD | [ |
VaD: vascular dementia; FtD: frontotemporal dementia; FTLD: frontotemporal lobar degeneration; DLB: Lewy body dementia; PDD: Parkison’s disease dementia.
The number of participants reported in each paper varied greatly, ranging from 12 to 654 subjects as shown in Figure
Number of subject histogram.
Datasets used repeatedly in the selected studies.
Datasets used in more than one study | Articles |
---|---|
22 subjects (11 AD, 11 Nold) | 4 [ |
24 subjects (10 mild AD, 14 Nold) | 2 [ |
27 subjects (20 probable AD, 7 Nold) | 3 [ |
28 subjects (14 probable AD, 14 Nold) | 3 [ |
34 subjects (22 probable AD, 12 Nold) | 2 [ |
34 subjects (17 AD, 17 Nold) | 3 [ |
48 subjects (17 early AD, 16 MCI, 15 Nold) | 4 [ |
62 subjects (3 databases: (a) 17 mAD, 24 Nold; (b) 5 mAD and 5 Nold; (c) 8 mAD and 3 Nold) | 3 [ |
74 subjects (74 MCI) | 9 [ |
79 subjects (79 probable AD) | 4 [ |
220 subjects (120 AD, 100 Nold) | 2 [ |
In most studies, the number of participants per group is well balanced between healthy controls, AD patients and, in some cases, MCI patients. Notwithstanding, 15 studies only included one group and this was the case of most studies in the progression category (Table
Group matching according to the number of subjects, age, gender, and education.
Group matching | Articles |
---|---|
One group only (15) | [ |
Number, age, gender, education (8) | [ |
Number, age, gender (10) | [ |
Number, age, education (7) | [ |
Age, gender, education (4) | [ |
Number, age (18) | [ |
Number, gender (2) | [ |
Number, education (3) | [ |
Age, gender (3) | [ |
Age, education (3) | [ |
Number (8) | [ |
Age (12) | [ |
Not paired or no information (19) | [ |
Forty studies included MCI participants, as detailed in Table
When EEG recordings were utilized along with other techniques such as MRI, PET, and CSF analyses, only studies that reported only-EEG-based diagnosis or assessment were considered. A total of 92 articles used exclusively EEG for their studies. Nevertheless, as other biomarkers have been explored and validated by the clinical community (e.g., CSF and MRIs), comparisons between these modalities and EEGs are very useful. Table
Combination of EEG with other modalities.
Modality | Biomarkers | Articles |
---|---|---|
MRI (11) | Cortical thickness, hippocampal atrophy, and other cortical density alterations | [ |
MRI and SPECT (5) | Regional blood perfusion and other cortical density alterations | [ |
SPECT (1) | Anomalous activities of cerebral neurons in NAT (neuronal activity topography) | [ |
MRI and genetic (1) | Comparison of Genetic (ApoE) and neuroimaging alterations | [ |
Genetic data (1) | ApoE genotype; PSEN1 E280A mutation | [ |
PET (1) | Disease processes revealed by cortical hypometabolism | [ |
MRI: magnetic resonance imaging; SPECT: single-photon emission computed tomography; ApoE: apolipoprotein E; PET: positron emission tomography.
The reported number of electrodes used for EEG signal acquisition in the reviewed studies varies greatly, from as low as one to as high as 256 electrodes (Table
Number of electrodes used by each selected study.
Electrode |
Articles |
---|---|
1–16 (14) | [ |
17–32 (89) | [ |
33–64 (2) | [ |
65–128 (5) | [ |
129–256 (2) | [ |
During EEG recordings, it is common practice to acquire simultaneously electrooculogram (EOG) and electrocardiogram (ECG) signals to monitor eye movement and heart activity, respectively. EOG and ECG are helpful as reference for cleaning the EEG signals as ocular and heart activity artifacts will be easier to detect and clean. Forty-one studies mention the registration of EOG signals in their studies [
Resting-state EEG can be recorded under two different conditions: sleeping and resting awake (either open or closed eyes). From the reviewed articles, the most common recording condition was resting awake eyes closed (EC), reported in 109 studies. None of the reviewed articles acquired EEG during sleep, or solely during resting awake eyes open (EO). Taking into account that the vast majority of participants are elderly and around half of them suffer from AD or MCI, resting-awake conditions are the most comfortable recording condition for participants, as they are not required to perform any mental task, which could be confusing or frustrating for these individuals [
Recording conditions.
Condition | Articles |
---|---|
Resting-awake EC (85) | [ |
Resting-awake EC + EO (13) | [ |
Resting-awake EC + EO + sensory stimulus (3) | [ |
Resting-awake EC + EO + cognitive tasks (8) | [ |
Resting awake, eye condition not reported (3) | [ |
The total duration of the EEG recording session was reported in 82 articles. This is important as long sessions can have detrimental effects for wet electrodes and cause alterations in participant mood and compliance [
Signal duration.
Description | Articles |
---|---|
[ | |
5–9 min (39) | [ |
10–20 min (17) | [ |
[ | |
Not informed (30) | [ |
In a broad sense, EEG signal preprocessing stands for the manipulations performed on the raw acquired data in order to prepare it for feature extraction in the next processing phases [
Filters.
Filter/preprocessing | Articles |
---|---|
Notch filter for power grid interference (35) | [ |
Resampling (12) | [ |
Rereference to common average (28) | [ |
Interpolation of bad channels (3) | [ |
The most common approach is the use of digital bandpass filters to enhance EEG-related spectral components. As each study had specific interest in different spectral components, diverse bandwidths have been reported. Lower bound of the EEG bandwidth is usually in the range of 0.1 to 4 Hz; however, the upper bound varies in a wider range, from 20 to 200 Hz. The most common upper limit was 70 Hz (31 studies) and the most used lower limit was 0.5 Hz (36 studies). Tables
Different upper limit bandwidths used by the selected EEG studies.
Upper limit (Hz) | Articles |
---|---|
≤25 (4) | [ |
26–50 (57) | [ |
51–75 (36) | [ |
≥76 (8) | [ |
Not reported (7) | [ |
Different lower limit bandwidths used by the selected EEG studies.
Lower limit (Hz) | Articles |
---|---|
≤0.5 (32) | [ |
0.5— |
[ |
≥1 (26) | [ |
Not reported (18) | [ |
Moreover, filtering can be performed with finite impulse response (FIR) or infinite impulse response (IIR) filters. The types of filters used in the various studies are presented in Table
Filter type.
Filter | Articles |
---|---|
FIR (26) | [ |
HOLS (1) | [ |
IIR (19) | [ |
Not reported (68) | [ |
EEG signals are inherently noisy and susceptible to blink, eye movements, heartbeats, cranial muscle, and power line artifacts. As mentioned previously, the process of cleaning EEG data from artifacts is pivotal in the EEG signal preprocessing pipeline. Analyzing clean EEG signals is an important prerequisite to avoid errors in the feature extraction step and to prevent misclassification of mental activity [
Artifact removal techniques.
Category | Method | Articles |
---|---|---|
Manual (65) | Epoch selection | [ |
Semiautomated (8) | ICA | [ |
ICA (IWASOBI) | [ | |
ICA (JADE) | [ | |
ICA in a sample and then ICA templates used to automatic removal | [ | |
ICA and wavelet denoising | [ | |
Automated (19) | FASTER | [ |
Notch filter on blink frequency | [ | |
LR to EMG electrodes | [ | |
wICA | [ | |
BSS-SOBI-CCA and wICA | [ | |
No filtering or no description (20) | — | [ |
BSS-SOBI-CCA: blind source separation based on second-order blind identification and canonical correlation analysis; ICA: independent component analysis; wICA: wavelet ICA; LR: linear regression; EMG: electromyographic.
While EEG devices can digitize data at high sampling frequencies (in the order of kHz), EEG signals are often downsampled as the processing of signals with excessive temporal resolution results in extra (and perhaps not as useful) computation load. As such, Table
Sample frequency.
Frequency (Hz) | Articles |
---|---|
125 or 128 (22) | [ |
200 or 256 (60) | [ |
500 or 512 (12) | [ |
1000 or 1024 (11) | [ |
Not informed (7) | [ |
The EEG signal is not stationary; however, it presents quasi-stationarity behavior for epochs (segments) ranging approximately from 1 to 60 s [
Epoch duration.
Duration (s) | Articles |
---|---|
0.3–1 (8) | [ |
1.1-2 (27) | [ |
2.1–5 (22) | [ |
5.1–10 (21) | [ |
10.1–20 (8) | [ |
21–70 (7) | [ |
Not informed (19) | [ |
The EEG epoch length is quite consistent across studies, with 56 studies using 5-second epochs or less. On the other hand, the number of epochs used in EEG analysis varies greatly from study to study. The utilization of overlapping epochs to extract EEG features and their averaging in the feature domain has been shown to improve the features of SNR and, consequently, increasing the classification performance [
Number of epochs.
Number of epochs | Articles |
---|---|
1–3 (12) | [ |
4–10 (14) | [ |
11–50 (20) | [ |
51–150 (20) | [ |
151–500 (7) | [ |
Not informed (39) | [ |
Study used several EEG epoching approaches (Section
Effective EEG duration.
EEG duration (s) | Articles |
---|---|
8–30 (12) | [ |
31–70 (20) | [ |
71–150 (9) | [ |
151–300 (20) | [ |
301–600 (9) | [ |
601–1500 (3) | [ |
Not informed (39) | [ |
EEG source localization methods estimate the location and distribution of active (electric) current sources within the brain based on the potential recorded through scalp electrodes. Going from activity recorded with electrodes to the current sources is an ill-posed inverse problem, since the number of unknown parameters is greater than the number of known parameters. In the last decades, this has been proven useful as a noninvasive neuroimaging technique, with high temporal and low spatial resolution that allows the characterization of “inside-the-brain” activity. A review on EEG source localization can be found in [
As already reported in Section
Slowing features.
Category | Description | Articles |
---|---|---|
Current source density | Source localization solutions | [ |
Spectral | Barlow’s metrics | [ |
Individual alpha peak (IAP) | [ | |
Individual alpha3 alpha2 | [ | |
Individual beta peak | [ | |
PSD (absolute and relative band power) | [ | |
PSD (band power ratios) | [ | |
PSD (central frequency) | [ | |
PSD (frequency peak in bands) | [ | |
PSD (mean frequency in bads) | [ | |
PSD (median frequency in bands) | [ | |
PSD (modelling parameters) | [ | |
Wackermann’s metrics | [ | |
Spectrotemporal | Wavelet (continuous) parameters | [ |
Wavelet (continuous) sparsification | [ | |
Wavelet (discrete) parameters | [ | |
Wavelet maximum frequency | [ |
Complexity features.
Category | Description | Articles |
---|---|---|
Entropy | Auto mutual information | [ |
Epoch-based entropy | [ | |
Fuzzy entropy | [ | |
Multiscale entropy | [ | |
Multivariate multiscale entropy | [ | |
Quadratic sample entropy | [ | |
Sample entropy | [ | |
Shannon entropy | [ | |
Spectral entropy | [ | |
Tsallis entropy | [ | |
Wavelet entropy | [ | |
Other | Bispectrum analysis | [ |
Central tendency measure | [ | |
Correlation dimension | [ | |
Distance-based LempelZiv complexity (dLZC) | [ | |
Hjorth activity, mobility, and complexity | [ | |
Lempel-Ziv complexity | [ | |
Visibility graphs | [ | |
Wavelet compression coefficients | [ |
Synchronization features.
Group | Description | Articles |
---|---|---|
Directed model based | Direct transfer function | [ |
Direct directed transfer function | [ | |
Full frequency transfer function | [ | |
Granger causality | [ | |
Kullback–Leibler divergence | [ | |
Lateral asymmetry index (LAI) | [ | |
Phase slope index (PSI) | [ | |
Sugihara causality | [ | |
Directed model free | Relative wavelet entropy | [ |
Peak interregional transfer entropy delays (PITED) | [ | |
Nondirected model based | Coherence | [ |
Coherence (wavelet) | [ | |
Correlation | [ | |
Correlation (amplitude envelopes) | [ | |
Detrended cross-correlation analysis (DCCA) | [ | |
Global field synchronization (GFS) | [ | |
Global phase synchronization | [ | |
Global synchronization index | [ | |
Lagged linear connectivity (LLC) | [ | |
Multivariate phase synchronization (MPS) | [ | |
Omega complexity | [ | |
Phase lag index (PLI) | [ | |
Phase synchrony | [ | |
S-estimator | [ | |
Stochastic event synchrony | [ | |
Nondirected model free | Coherence entropy coefficient | [ |
Correlation entropy coefficient | [ | |
Mutual information | [ | |
Permutation disalignment index | [ | |
Synchronization likelihood | [ | |
Wavelet entropy coefficient | [ | |
Others | Canonical correlation | [ |
Global field power (GFP) | [ | |
Graph theory metrics | [ | |
Static canonical correlation | [ |
Neromodulatory features.
Description | Articles |
---|---|
Amplitude envelope, spectral analysis | [ |
Amplitude envelope, statistics | [ |
Nonbiological features.
Description | Articles |
---|---|
ANN extracting spatial content from EEG | [ |
Back-predictive model | [ |
Linear predictive model | [ |
Paraconsistent artificial neural network (PANN) using morphological analysis of EEG | [ |
Symmetric predictive model | [ |
As reported in the study goal (Section
The reported findings for discriminative studies fall into three categories: (1) studies reporting statistical significance of used features, (2) studies reporting classification performance among populations, and (3) studies reporting both statistical significance and classification performance. Table
Classification, statistical analysis, or both.
Description | Articles |
---|---|
Statistical (35) | [ |
Classification (36) | [ |
Both (23) | [ |
Papers where statistical significance was reported used a variety of parametric and nonparametric methods for statistical analysis. Table
Statistical analysis strategy in the selected studies.
Description | Articles |
---|---|
ANOVA | [ |
Anterior hub ratio | [ |
chi squared | [ |
Correlation | [ |
Correlation |
[ |
Correlation |
[ |
Cost function | [ |
Graph analysis | [ |
Kruskal-Wallis | [ |
LDA | [ |
Lilliefors test | [ |
Log- |
[ |
Mahalanobis D2 | [ |
MANCOVA | [ |
Mann–Whitney | [ |
MANOVA | [ |
Mean and standard deviation | [ |
PCA | [ |
Quadratic univariate regressions | [ |
SNK | [ |
LDA: linear discriminant analysis; MANCOVA: multivariate analysis of covariance; SNK: Student–Newman–Keuls.
In studies where classification performance was reported, three important aspects were taken into account: feature selection, cross-validation, and classification algorithm.
Feature selection.
Feature selection methods | Articles |
---|---|
AUC maximization | [ |
BFE | [ |
Consistency-based filter (CBF), correlation-based feature selection (CFS), filtered subset evaluator (FSE), Chi squared (CS), gain ratio (GR), relief- |
[ |
Correlation-based pursuit | [ |
FCBF | [ |
Fit-curve model | [ |
Genetic | [ |
Logistic regression | [ |
Manual | [ |
OFR | [ |
[ | |
PCA | [ |
Ranking by Fisher ratio score | [ |
Reverse sequential feature selection | [ |
SVD | [ |
SVM classifier (best performers) | [ |
BFE: best feature extraction; FCBF: fast correlation-based filter; OFR: orthogonal forward regression; SVD: singular value decomposition.
Cross validation methods.
Description | Articles |
---|---|
5-fold CV | [ |
10-fold CV | [ |
100-fold CV | [ |
500-fold CV | [ |
Dataset split in train and test set splits | [ |
LOSO | [ |
Leave one epoch out | [ |
CV: cross-validation; LOSO: leave one subject out.
Classifying Strategy.
Classifier | Articles |
---|---|
ANN | [ |
ANOVA | [ |
Autoregressive models | [ |
Back predictive model | [ |
Decision tree | [ |
k-nearest neighbor | [ |
LDA | [ |
LR | [ |
LRA | [ |
Nave Bayes | [ |
PANN | [ |
Parzen classifier | [ |
PCA | [ |
PDM-based model | [ |
PNN | [ |
QDA | [ |
ROC | [ |
SMO | [ |
SVM | [ |
Takagi-Sugeno neurofuzzy inference system | [ |
ANN: artificial neural network; LDA: linear discriminant analysis; LR: logistic regression; LRA: logistic regression analyses; PANN: paraconsistent artificial neural network; PDM: principal dynamic mode; PNN: probabilistic neural network; QDA: quadratic discriminant analysis; SMO: sequential minimal optimization.
A total of 39 studies aimed to find correlates between EEG-based features and AD progression. Table
AD progression assessment.
Description | Articles |
---|---|
ANOVA | [ |
ANCOVA | [ |
ANOVA 2 way | [ |
Chi squared | [ |
Correlation | [ |
Correlation (Pearson) | [ |
Correlation partial | [ |
Correlation (Spearman) | [ |
Genetic search multiple markers | [ |
K-means | [ |
LDA | [ |
Linear regression | [ |
Mahalanobis D2 | [ |
Mann–Whitney | [ |
Quadratic ordinary least squares regression models | [ |
R2 | [ |
Scheffes test | [ |
[ | |
[ | |
Wilcoxon rank-sum test | [ |
ANCOVA: analysis of covariance.
By compiling the different limitations reported in all the reviewed articles, it is possible to have an idea of the issues that need to be addressed in the following years to advance EEG-based research on AD. Firstly, the most reported limitations are related to the population participating in the studies, specifically, the small size of the dataset and cohorts (Section
Regarding the EEG processing, emphasis is often put on the manual selection of clean EEG epochs, which introduces human biases and cannot be reproduced. A limitation reported in a very recent study [
Lastly, limitations related to the reported outcomes include the uncertainty of AD diagnosis using MMSE and other neuropsychological tests. Several studies measure the classification accuracy between AD or healthy controls using the results from these tests. However, neuropsychological tests do not provide 100% sure diagnosis; they do not work well in all dementia stages, and as they have lower sensitivity, it is difficult to detect early stages of AD [
Reported limitations.
Category | Description | Articles |
---|---|---|
Population | Small number of subjects in the study | [ |
Merged databases are different due to local implementations | [ | |
Lack of different stages in AD cohort | [ | |
AD cohort includes participants taking antidementia drugs | [ | |
Lack of population matching, age, gender, and/or education | [ | |
Possible preclinical AD in N cohort | [ | |
Prodromal AD was applied in aMCI with A |
[ | |
EEG experiment setup | No severe AD as hard to perform EEG recordings | [ |
Presence of dominant alpha activity during EC condition | [ | |
Differences in datasets due manual artifact handling | [ | |
Low number of electrodes for source localization methods | [ | |
Low number of electrodes for connectivity analysis | [ | |
Low number electrodes for advanced AAR methods | [ | |
Reported results | Lack of research for other dementia types | [ |
Lack of longitudinal approach for N, MCI, AD populations | [ |
After the discussion in previous subsections, various aspects worth to be addressed in future resting-state EEG-based studies are presented in Table
Recommendations.
Recommendations for future EEG-based AD studies |
Provide detailed population characteristics |
Describe how the AD diagnosis was performed |
Mention whether the MCI participants were followed-up |
Detail EEG experiment in duration and phases |
Use standard EEG layouts |
Mention not only the quantity of channels but their location |
Define EEG processing in more detail |
Use standard features such as PSD features as baseline |
Describe artifact handling strategies |
Throughout this review, we found that several studies do not present a detailed characterization of the cohorts participating in the study. Variables such as age, gender, and education level have been demonstrated to be confounding factors in AD [
Regarding the study setup, a frequent issue that arises from this systematic review is the huge amount of different experimental setups that have been reported across the reviewed articles. While all the databases utilized by the studies included in this review used the same resting-awake eyes-closed protocol, the recording duration was extremely variable (Section
One direct consequence of this experimental variability is that most of the reviewed studies performed their analysis just on one dataset. While testing the efficiency of the developed methods on other datasets is highly advisable to verify if the results are realistic and can be generalized, this variability makes that practically impossible. As such, it is recommended that a standardization effort on EEG data collection and experimental protocol be put in place to facilitate cross-site, cross-country, and cross-database validation.
For EEG processing, in turn, the most used artifact-handling approach was the meticulous visual inspection by expert clinicians, which is inherently irreproducible and prone to errors. Consequently, even when EEG data is collected in the same conditions, the manual rejection of artifacts hinders the comparisons among different approaches of the same experimental setup. Studies could make use of AAR methods to report their results with manual selected EEG signals and contrast them to the ones obtained with automatically cleaned signals, as was done in [
Moreover, when recording EEG signals, less than half of the studies use EOG and very few use ECG electrodes. Registering eye and heart movements can help with artifact removal and thus should become standard during data recording. In addition, a clear description of the EEG signal epoching process should be provided and aspects like epoch length, epoch overlap, and number of epochs used need to be mentioned in the article. When source localization is performed, higher-density montages are desirable (≥25 electrodes) [
Lastly, not every reviewed paper mentioned the limitations found during the study, this could be enlightening for the design of future studies. A solution for some of the reported limitations (Section
In this systematic review, a total of 112 journal articles published between January 2010 and February 2018 on the utilization of EEG for AD diagnosis and progression assessment were surveyed. In these papers, the most often reported goal was to discriminate between healthy controls and AD participants (59 articles). From these articles, crucial aspects were grouped under five main categories: study rationale, study population, experiment setup, EEG processing, and reported outcomes. Such aspects were reviewed, compared, and discussed, with the final goal of providing an overview of the state of the art on resting EEG for AD diagnosis and assessment.
In this review, limitations reported in the reviewed articles were also collected and discussed, with the aim of having an idea of the issues that need more attention in order to advance the use of EEG in AD research. Among these reported limitations, the limited number of datasets available to researchers appeared to be the most common one. Ultimately, it is hoped that this review will boost the research of EEG as a noninvasive, less-expensive, and potentially portable technique for AD study, assessment, and diagnosis, particularly for low- and middle-income countries which lack access to costly neuroimaging equipment.
The authors declare that there is no conflict of interest regarding the publication of this paper.
Raymundo Cassani, Mar Estarellas, and Rodrigo San-Martin contributed equally to the paper, listed in alphabetical order by last name.
Professor Francisco J. Fraga was partially supported by the Fundação de Amparo à Pesquisa do Estado de São Paulo (FAPESP), grant # 2017/15243-7.