Systematic Review of EEG Coherence in Alzheimer’s Disease

Eligibility criteria

Studies selected for the review included original, full-text articles published in English, investigating MSCOH in patients with AD. The following eligibility criteria were chosen:

The study must have collected data on quantitative EEG data in human participants (not being a review, conference letter, computer simulation, animal study, etc.).

Search

The following bibliographic databases were searched to identify eligible studies: PubMed (NLM interface), using predefined Mesh-terms; Scopus; and Embase. In the MeSH database, no mesh-term for EEG-coherence exists. The following search strings were used: “Alzheimer Disease” [Mesh]) AND “Electroencephalography” [Mesh] AND “coherence”, “Cognitive Dysfunction” [Mesh]) AND “Electroencephalography” [Mesh] and “coherence”. Using Embase the following search strings were used: “Alzheimer OR Alzheimer’s disease) AND (electroencephalography OR EEG) AND coherence”. Using Scopus the following search strings were used: “eeg AND coherence AND Alzheimer”, “eeg AND coherence AND alzheimer’s”, “eeg AND coherence AND alzheimer’s AND disease”, “electroencephalography AND coherence AND alzheimer’s disease”.

Study selection

We performed deduplication due to multiple bibliographic databases being used. Titles and the abstracts were then screened based on the inclusion criteria listed above was performed by two researchers (MFF and CSM) independently. Discrepancies were resolved in discussion with another author (ICZ). Studies that were eligible for inclusion based on title and abstract were read at full text length. Again, if disagreements or doubts were encountered, the third author (ICZ) was consulted for inclusion of studies.

Data extraction

Data from included articles were systematically extracted. A summary of retrieved data about study aim(s), demographics, and details on EEG recording and preprocessing are shown in Table 1.

Due to the heterogeneity of the included studies, we chose not to perform a meta-analysis.

Demographics

OPEN IN VIEWER

Fig. 1

Literature search and study selection

24 of the included studies examined AD participants diagnosed according to the 1984 NINCDS-ADRDA criteria [12]. Two studies used the 2011 NIA-AA criteria [13] and in 1 study AD participants had been diagnosed according to the 2018 NIA-AA criteria [14] requiring evidence of amyloid- and tau-pathology. Two studies solely used the DSMIII-R criteria [15], two studies used the DSM-IV criteria [16], of which one also stated use of ICD-10 [17]. Details on diagnostic classification criteria were not available for 3 studies [18–20] and only for a subgroup of participants in another [21], relying partly on the NINCDS-ADRDA criteria. The number of examined participants in the studies varied greatly from the minimum number of participants eligible for study inclusion (n = 10) to almost 400 [22].

A total of 19 studies reported a mean age of AD participants in the range of 70–80 years. In addition, two studies had subgroup of AD participants with a mean age in the 70–80 [21, 23]. In 10 studies, the mean age of AD participants was in the range of 60–70 years while four studies reported the mean age of AD participants to be below 60 years [24–27]. No study reported a mean participant age above 80. Five studies reported significant differences in mean age of HC compared to AD [1, 28–30] and 12 studies explicitly showed no significant difference in age [4, 31–38]. A visual summary on the relationship between age and MMSE-score of included studies can be seen in Fig. 2.

Cognitive scores

OPEN IN VIEWER

Fig. 2

Relationship between age, MMSE score, and sample size of included studies. Relationship between age of participants and mean MMSE-score of included studies. Size of the dots reflects study population size. Blue: Alzheimer’s disease dementia, light brown: mild cognitive impairment, green: healthy controls.

The lowest mean MMSE score reported for AD participants was 8.9 [34] and the highest was 27.6 [30]. Four studies did not report any details on MMSE scores of participants [18, 40] and for some studies the details on MMSE scores of participants were sparse [25, 41–43].

MSCOH findings

A summary of MSCOH findings across all included studies is shown in Fig. 3. Most studies classified MSCOH according to the conventional alpha (8–13 Hz), beta (13–25 Hz), delta (1–2, 1–3 or 1–4 Hz), and theta bands (3–7, 4–8 or 4–9 Hz) with minor variations among the studies on the definition of the different frequency bands (see Fig. 3 for details). Some studies subclassified the alpha and beta bands [1, 41–47]. One study calculated MSCOH for each frequency individually [48].

OPEN IN VIEWER

Fig. 3

Magnitude-Squared Coherence (MSCOH) differences between AD and HC divided into frequency bands.

34 of 35 studies investigated MSCOH in the alpha band. The most consistent finding was significantly reduced coherence in the alpha band reported in 24 studies. In addition, four studies [30, 43] reported a non-significant reduction of alpha MSCOH in AD compared to HC. Four studies did not find any significant difference in alpha band MSOCH among groups [29, 46], two studies reported a significant difference in the alpha band between AD and HC but did not state the direction [18, 19], one study did not investigate MSCOH in the alpha band [40].

31 of 35 studies investigated MSCOH in the beta band. 18 studies reported significantly reduced beta (or similarly classified) band MSCOH in AD compared to HC [1, 49]. Four studies reported non-significant trends of lower beta MSCOH in AD participants [20, 50], 6 studies found no difference between groups [4, 48], 1 study found a trend of increased beta MSCOH in AD [38]. Two studies reported a significant difference in the alpha band between AD and HC but did not state the direction [18, 19]. Four studies did not investigate MSCOH in the beta band [29, 47].

31 of 35 studies investigated MSCOH in the theta band. 9 studies reported significantly lower theta MSCOH in AD compared to HC [20, 49] and five studies reported trends in the same direction [24, 50]. Two studies reported the opposite with theta MSCOH being increased in AD compared to HC [1, 30]. One study found a significant difference between groups but did not state the direction [19].

29 of 35 studies investigated MSCOH in the delta band. Five studies reported significantly lower MSCOH in AD [21, 49] and five studies reported trends of lower delta MSCOH in AD [20, 46]. One study reported significantly increased MSCOH in the delta band [50], two other studies reported trends in the same direction [34, 39], two studies reported significant differences between the groups without stating the direction [18, 19]. 14 studies did not report any significant or trend in findings for delta MSCOH.

MSCOH in the gamma band was investigated in five studies [20, 46] of which one study reported a significant decrease in AD compared to HC [40] and one study reported a non-significant decrease [20]. In addition, some studies reported on ‘gamma coherence’ which according to Fig. 3 would be classified as beta-coherence [21], others reported MSCOH findings for a broad common beta-band including frequencies above 30 Hz [28,31, 33].

13 studies [1, 47] out of 35 explicitly reported use of correction for multiple comparison when appropriate, others [35, 47] fixed the level of statistical significance at p < 0.01 or lower (see Supplementary Table 1 for details).

Diagnostic utility

9 studies used MSCOH to discriminate between patients with AD, MCI, and HC. Details on classification using solely MSCOH or measures derived from MSCOH were available for 5 studies [1, 49]. One study reported a high overall accuracy >90% [1], one study had an overall accuracy of 84% and 93.54% for AD versus HC [19]. Four studies showed accuracies below 80% [37, 49] with one of these studies only investigated the ability of MSCOH to separate AD from multi infarct dementia [37] with a classification accuracy of 76%. All the mentioned studies used different types of classifiers including multiple discriminant analysis for both the raw [19] and Z-scores from coherence [49] or simply a cut-off [37, 38]. None of the included studies have performed validation of their respective classifiers in a separate cohort.

Preprocessing findings

Most studies used between 16 and 21 electrodes with the highest number of electrodes being 22 in a subsample of one study [21] and the lowest being 8 electrodes [48]. In 2 studies the number of electrodes was not reported [32, 35]. The choice of reference was mentioned in 25 of 35 studies. Sample rate varied between 102.4 and 1024 Hz with most studies having collected data at a rate ranging from 127 Hz to 200 Hz. Down sampling was used in 3 studies [1, 30].

Recording duration ranged from 20 s [21] to 20 min with 8 studies not reporting recording length. In addition to resting state, 5 studies added photic stimulation [24–27, 48], 1 study subjected participants to observation of a chaos pendulum [38], 1 study added auditory stimulation [48], and 2 studies added cognitive tasks during the experiment [40, 42].

Epoch length was reported in 34 studies, ranging from 500 ms to 24.56 s. 28 studies reported the number of epochs used for MSCOH calculations ranging from 2 [38] to more than 100 [1, 30].

No included study reported use of spatial filters. The applied method for artifact removal was described in 25 of 35 studies (see Supplementary Table 1). All these studies relied on visual inspection (VI) of (raw) EEG data. 4 of the 25 studies used Independent Component Analysis in addition to visual inspection [1, 46] (Supplementary Table 1) and one study used Principal Component Analysis [21].

Alpha coherence differences

The literature on MSCOH in patients with AD is heterogeneous with respect to severity of the disease, demographics, and methods for preprocessing of the EEGs. Despite of this, there was a pattern of decreased alpha coherence in participants with AD in 24 out of 34 studies investigating this frequency band [1, 47–50]. This finding is reported consistently across differences such as sample size [22, 50] (ranging from 10 to >200 participants in each group), disease severity (MMSE 8.9–27.6 in the AD group) [30, 34], and age of participants [24, 33]. The consistent findings of decreased alpha MSCOH in AD strongly suggest that alpha MSOCH may serve as a diagnostic marker of AD. Indeed, a recent study show that MSCOH can separate patients with AD from HC with accuracies >90% [1]. However, alpha coherence does not seem to be specific for patients with AD as it has been shown to decrease in other neurodegenerative disorders like Lewy body dementia [51]. Therefore, future research into MSCOH as a diagnostic marker should take account of the topography of alpha MSCOH changes between different neurodegenerative disease.

Neurobiological correlates of coherence

How alpha MSCOH correlates with the underlying pathophysiology as AD develops is unknown but one proposed hypothesis is cholinergic dysfunction. In support of this hypothesis, reduced alpha coherence or lower beta coherence has been reported in HC experimentally treated with scopolamine, a drug known to decrease brain cholinergic activity [8, 52]. However, studies investigating neurophysiological changes after treatment with choline esterase inhibitors are scarce [53–55]. Evidence from clinical populations indicate a high variability of serum concentration of donepezil [55, 56], the most widely prescribed choline esterase inhibitor. The relationship between serum concentration of choline esterase inhibitors and MSCOH has not yet been explored in a longitudinal clinical setting. Reduced alpha coherence is correlated with both cognitive dysfunction assessed by neuropsychological examination [3, 50] and disease severity [5]. Cortical atrophy could be associated with decreased alpha MSCOH, which has been shown to be present early in the disease [57] which emphasizes the relevance of MSCOH as a marker of disease. It is unlikely that the decrease in MSCOH is entirely due to atrophy since cholinesterase inhibitors lead to cognitive improvement in patients, particularly in those with the APOE4 genetic risk factor who also show a more pronounced decrease in alpha coherence [58–61]. Therefore, cholinergic integrity may be partially linked to alpha coherence, but more studies are needed.

Low frequency coherence

Results were more mixed or contradictory regarding 1–8 Hz coherence. Thirteen studies reported reduced low-frequency MSCOH whereas four studies reported the opposite. There is no obvious explanation for this discrepancy, but we hypothesize that it may be due to either different stages of the disease or difference in the preprocessing steps. In support of low-frequency being due to different stages of the disease, one study has suggested that low-frequency MSCOH may be a marker of progression in patients with mild cognitive impairment [30] but only a very small sample size was investigated.

High frequency coherence

Most included studies reporting on coherence in the beta-band (or similarly classified) found a similar decrease of coherence in AD. Though compared to findings on alpha coherence more studies reported no or insignificant decreases. Of the 5 studies examining coherence in the gamma-band (or similar), 2 studies reported differences among the groups. The validity of such reporting must however be interpreted with caution due to a risk of apparent gamma-activity being caused by muscular artifacts [62, 63].

Diagnostic utility

With respect to diagnostic utility, 5 studies [1, 49] used MSCOH for disease classification. Here, two studies showed accuracies above 90% for discrimination between AD and HC [1, 19], albeit one of these achieved a lower overall classification accuracy of 84% when participants with MCI were considered [19]. Three studies showed accuracies in the range of 75–78% [37, 49]. The study with high overall and AD/HC specific accuracy [1] used more advanced techniques and more importantly showed that classification accuracy was improved by including other frequency bands besides alpha coherence. In contrast, the studies with lower accuracies relied on fewer coherence values for classification [37, 38] or values dependent on significant findings from the statistical analysis [49]. Therefore, methods that include more MSCOH values in different frequency bands may improve the diagnostic accuracy. In addition, a study found that reducing the number of included 1-s epochs strongly affected the results, which suggests that the signal-to-noise ratio is strongly affected by the length of the recordings. No studies have so far performed separate validation of the classifier.

Thus, we emphasize that the evidence in support of clinical utility of MSCOH is presently limited and cross-validation inadequately utilized, which warrants caution with respect to generalizability of the results. Future validation studies are needed and should recruit a large representable sample of patients and control participants referred to a memory clinic (i.e., subjective cognitive decline) and classifications metrics require rigorous cross-validation. In this review, the inclusion criteria prior to literature search studies only examining AD at the stage of MCI where excluded. Therefore, we did not review MSCOH as a classifier at such an early stage of the disease, which arguably would be the most valuable with respect to establishing a timely diagnosis by cost-efficient means.

Diagnostic criteria and non-EEG modalities

Variability in the diagnostic criteria is unlikely to have affected the results significantly, as 24 out of 35 studies used the 1984 McKhann criteria and further 2 the 2011 NIA-AA criteria. Only 1 study applied the 2018 NIA-AA criteria which imply demonstration of amyloid or tau pathology via biomarkers, and 5 (4 not specified and 1 study only for subgroup) studies did not specify the applied criteria for the whole study group. In general, a variety of methods like CSF sampling, PET scans and MRI are validated as relevant diagnostic markers for neurodegeneration and amyloid pathology in AD [64, 65]. All these methods are costly and/or invasive, and some only available at academic centers. Compared to considerably more expensive diagnostic methods (i.e., PET scans) EEG derived MSCOH may serve as a non-invasive cost-efficient method relevant for both academic and especially non-academic centers.

Preprocessing differences and consequences

There are several steps in both the recording and subsequent preprocessing that can lead to bias in the results. Most importantly, all the included studies have used visual inspection for removing artifacts, which is likely to introduce bias due to inter-rater variability. Therefore, future studies should move towards more automatic artifact rejection as more automatic artifact removal toolboxes become available [66, 67]. So far, one study has compared the different types of automatic pre-processing toolboxes [68], but more studies are needed. In addition, the number of included EEG data has shown to strongly affect the results and the studies included between 8 seconds [43] and >100 s of data [1]. Future studies may consider analyses to show the differences in coherence depending on the amount of included data. Lastly, the number of electrodes were for all but one of the included studies 21 or less. Future studies need to investigate the required duration of the recording and the number of electrodes needed to use EEG as a diagnostic marker.

Volume conduction

Volume conduction [69–71] is a phenomenon causing adjacent recording electrodes to be influenced by activity from the same source creating spurious correlations [72]. Volume conduction effects on coherence estimates are greater with shorter inter-electrodes distances, but the effect can be attenuated by a number of preprocessing techniques such as applying a surface Laplacian spatial filtering [73], which generally also benefits from a higher spatial resolution afforded by using more recording electrodes [74]. Furthermore, several EEG measures have been shown to be better to mitigate the effects of volume conduction [75] including (weighted) phase lag index and the imaginary part of coherence [71]. In general, MSCOH is strongly affected by volume conduction but is still widely applied in studies investigating EEG in patients with AD. Even though volume conduction might have biased MSCOH results reported in the literature, the magnitude of this problem is unlikely to invalidate the overall direction of the reported coherence differences between AD and HD as these align with cognitive, behavioral, and other measures to suggest that coherence differences are caused by the pathological changes in AD. However, increasing use of analytical strategies mentioned earlier to mitigate volume conduction effects might further enhance the diagnostic utility of coherence analyses.

Limitations

Due to the heterogeneity of included studies regarding both demographic factors, preprocessing and MSCOH data reporting we did not attempt to conduct a meta-analysis on MSCOH findings across frequency bands. As noted, many included studies examined small sample sizes (10–20 participants in each group), which suffer from low statistical power. Many studies also perform multiple statistical tests between different EEG electrodes for multiple frequency bands whilst not employing controls for multiple comparisons.

Conclusion

Based on a systematic review of the literature, MSCOH is reduced in the alpha band in patients with AD compared with HC. This finding is consistent regardless of demographics, EEG recording protocols, and preprocessing. A possible explanation for the reduction of alpha coherence is cholinergic dysfunction due to disturbance of the cholinergic brain pathways and/or possible related to atrophy. Although less consistent, a similar trend in reduction of coherence in the beta-band was reported. Low-frequency coherence findings were less uniform, but more studies are needed to understand changes in the early stages of the disease. Two studies found classification accuracies above 90%; however, validation of MSCOH for diagnostic utility in patients with AD is needed.