Diagnostic Utility of Electrophysiological Markers for Early and Differential Diagnosis of Alzheimer’s, Frontotemporal, and Lewy Body Dementias: A Systematic Review

Inclusion criteria

- Participant population: AD, FTD, DLB, including preclinical and prodromal stages, presymptomatic carriers of genetic causative mutations and healthy controls. In line with the International Working Group 2 (IWG-2) criteria for AD, we only include biomarker-proven cases of AD showing in vivo evidence of amyloid and tau pathology [12], presymptomatic carriers of familiar AD (FAD) by Amyloid Precursor Protein (APP) and Presenilin (PSEN1/PSEN2) causative mutations were included. Criteria for FTD diagnosis were a diagnosis following the Rascovsky criteria for bvFTD [13], Gorno-Tempini criteria for PPA[14], carrier of a genetic mutation of FTD, and/or imaging evidence of frontotemporal atrophy. DLB was defined by the criteria of McKeith [6, 15].

- Primary research studies

- English language

- Studies using resting-state EEG and/or task-based EEG as neuroimaging modality

Exclusion criteria

- Studies that combine multiple neuroimaging modalities or EEG with other biomarkers but do not separately report the performance of EEG markers.

- Animal studies

- Population with comorbidities (Down Syndrome, Schizophrenia, …)

- Prognostic studies

- Case studies or study population <5 participants

- Sleep EEG

Lastly, risk of bias was assessed using a hybrid version of the Joanna Briggs Institute (JBI) Critical Appraisal checklist (supplementary material S2).

Resting state EEG

Diffuse slowing of the frequency spectrum towards lower frequencies is a consistent observation in both prodAD [17–20] and demAD [21, 22]. The shift in the power spectrum in demAD manifests as reduced spectral metrics in the α band, while spectral power in θ band increases [22]. Smailovic found that low α global field synchronization (GFS), a measure of global functional connectivity in the frequency domain, is linked to abnormal Aβ42 levels in demAD and prodAD, and high levels of p-tau and t-tau in prodAD [23]. Secondly, in demAD, high levels of p-tau and t-tau correlate with low global strength of scalp potential (GFP) in β and δ bands. In prodAD, high GFP in δ and θ was linked to abnormal CSF levels of Aβ42, while lower GFP α and β is linked to increased p-tau and t-tau. These results are in line with the research of Cecchetti, showing that prodAD show higher θ density than MCI without AD pathology with over 75% sensitivity and 70% specificity [21]. Similarly, Rodriguez demonstrated that frontotemporal alterations in β band discriminate between presymptomatic PSEN1 carriers and noncarriers and between FAD and HC (see figure 3)[24]. Additionally, significant progressive alterations in bispectral metrics are found along the AD continuum [25]. Specifically, with increasing severity, interactions between the δ and θ bands and other bands increase, while interactions with high frequency bands α, β1, and β2 bands diminish. Similarly, synchronization measures demonstrate reduced functional connectivity in demAD in α and β bands, while coherence metrics in θ band increase [26]. These results are similar to findings by Revilla-Vallejo, where Shannon Entropy (SE) shows higher values in the δ band and lower values in the α and β1 bands in prodAD and demAD compared to HC, suggesting less connectivity and integration in these two bands [27].

• Download figure
• Open in new tab

Figure 3.

AUROC values in AD studies. If multiple values were provided, only the highest value was reported. Studies were organized according to subpopulation. AD=Alzheimer’s Disease, prodemAD= prodromal AD, preAD= preclinical AD, Acr = Asymptomatic carriers, prodFAD= prodromal Familial AD, FAD= Familial AD, rs-EEG= resting-state EEG, tbEEG= task- based EEG, AUC= Area under the curve, Sen=sensitivity, Spe= specificity, Acc= accuracy

Another way to look at resting-state activity is through microstate analysis. Spontaneous resting state activity can be described using microstates, transient global patterns of scalp potential, reflecting coordination of neural activity among networks [28]. Microstate analysis in AD reveals alterations in the duration, occurrence and coverage of microstates varying with severity. Specifically, demAD patients showed a longer duration in microstate B [29]. In the study of Smailovic, topographic differences in microstates A and D were found between HC and patient populations SCD, prodAD and demAD [30]. Furthermore, preclinical and prodromal AD could be differentiated from the dementia stage by topographical differences in microstate A. The authors further found that topographical alterations in microstate C were linked to increased Aβ42 levels, while p-tau levels were linked to microstate B alterations.

Recently, a number of studies used EEG-based machine learning classifiers for the detection of AD. An advantage of these studies is that the outcome measures offer more insight into the potential clinical implementation, which necessitates robust results at the individual level, compared to the prior studies which reported results from group-level comparisons. Studies discriminating AD from HC show promising classification accuracies in demAD ranging from 78% to 91% and in prodAD ranging from 73-85% (see figure 3). Another measure of the model’s accuracy is the F1 score, which is calculated using precision (positive predictive value) and recall (true positive rate). Studies achieved encouraging F1 scores ranging from 0.86 to 0.88 [31, 32]. Regarding the age factor, Durongbhan [33] found encouraging results in discriminating demAD from HC below the age of 70 (sensitivity >90%, specificity >83%), in individuals over 70 years, results were a little lower (>67% sensitivity, >85% specificity). Zhao studied both linear and nonlinear synchronization and found age- and disease-related differences in network synchronization [34]. Healthy individuals below 70 years old exhibit widespread linear synchronization with dynamic variability between eyes closed (EC) and eyes open (EO) states, while in young individuals with demAD, this dynamic variability is diminished, indicating network dysfunction. Secondly, they found a high widespread nonlinear synchronization during EO with higher dynamic variability compared to demAD. In individuals above 70, both HC and demAD showed similar levels of linear and nonlinear synchronization with minimal variability between states. Combining linear and nonlinear connectivity yields highest classification accuracies below the age of 70 (80.3% EO, 74.5% EC) and above the age of 70 (EO 86.5%, EC 90.5%).

In preAD, PSD patterns can be used to model compensatory mechanisms and progression to AD [35]. More specifically, individuals with amyloid burden in absence of neurodegeneration evidenced by brain metabolism in AD core regions, show increased functional connectivity in the parieto-occipital α band. In individuals with neurodegeneration, the impact on EEG metrics varies depending the degree of amyloid accumulation. More specifically, surpassing a critical threshold of amyloid accumulation, reverses the compensatory upregulation of higher β and γ frequencies and decreased δ power seen in intermediate amyloid burden, resulting in decreased β and γ power, MSF, spectral entropy, complexity and wSMI in θ band while δ power increases. In individuals with subjective complaints (SCD) with amyloid burden, Shim observed similar patterns of increased δ power in parietal, occipital and posterior cingulate regions combined with decreased α activity in fusiform and inferior temporal areas[36]. In predicting AD pathology in individuals with SCD, the best ML model shows 88.6% accuracy (see figure 3)[37].

Task-based EEG

A first important domain affected in AD is encoding and memory. Tautvydaite [38] shows that demAD patients show neural deficits in novelty detection and encoding during both learning and delayed memory recognition of pictures. Specifically, demAD patients show a decreased P200 response to new and repeated items, reflecting attention and perceptual processing, as well as an increased P300 during delayed repetition, which might reflect difficulties differentiating new from familiar stimuli. Similarly, in a passive picture recognition task of Stothart [39], demAD patients showed a decreased neural response to familiar pictures. Early differences in visual short term memory have also been found by Pietto[40] showing a reduced N1, P2 and P3 in prodFAD. Several studies including asymptomatic carriers of an AD causative genetic mutation, explored the potential of task-based EEG for early detection, highlighting early neural functional alterations without corresponding behavioral impairments. Golob [41] observed that presymptomatic carriers of the PSEN1 or APP mutation (mean age = 33.9), show discriminable neural alterations during an auditory target detection task. These alterations included decreased slow wave amplitudes, increased P200 amplitude, and delayed N100, N200, P200 and P300 latencies. Comparison of latencies across the ERP components showed that the latencies in the asymptomatic carriers were around 10% longer than the noncarrier group. Nontarget P200 latency emerged as a potent discriminator, successfully identifying 87% of the presymptomatic carriers. Quiroz [42] observed that during a recognition memory task, presymptomatic PSEN1 carriers showed lower frontal ERP positivity alongside an increased occipital positivity, with 72.7% sensitivity and 81.8% specificity discrimination. Despite these neural differences, both groups performed equally well on the task. Control subjects during recognition memory, exhibited activation patterns reliably associated with frontally mediated processes distinguishing between studied and unstudied visual items. PSEN1 carriers on the other hand, showed increased brain activity in occipital regions associated with visual perceptual processing. Ochoa [43] demonstrated a higher connectivity during the encoding condition of the same recognition memory paradigm in PSEN1 presymptomatic carriers compared to non-carriers. Connectivity in the occipito-parietal region during the same memory encoding condition within the 500-600ms time window is able to differentiate between presymptomatic and non-carriers with 67 % sensitivity and 80 % specificity [44]. P300 latency during an auditory oddball paradigm correlates with CSF levels of p-tau181, p-tau199 and ptau231 across demAD, prodAD and HC, while N200 latency negatively correlates with Aβ42 [45]. Recent studies in preAD using a working memory paradigm point towards reduced α event-related desynchronization (ERD) and altered α spectral entropy, suggesting compensatory hyperactivity during low load and insufficient cognitive resources with increasing work load [46, 47]. In the γ band, low working load induces a higher low γ in preAD, while decreases in γ are observed during high load [48].

In the language domain, sources of ERPs during a semantic-matching task reveal distinct topographical patterns in presymptomatic carriers versus non-carriers, although their behavioral responses and N400 amplitudes remain similar [49]. Specifically, presymptomatic carriers show a notable decrease in N400 generator strength within the right inferior-temporal and medial cingulate areas, and an increase within the left hippocampus and parahippocampus compared to non-carriers. The observed shift in N400 distribution mirrored that seen in symptomatic carriers, albeit with a less pronounced reduction in generator strength.

rsEEG

Spectral analysis in FTD shows a consistent slowing of the frequency spectrum. While studies agree on a slowing in α frequencies, there is some discrepancy regarding the changes in the β band, with some indicating a decrease and others suggesting an increase [9, 50]. In the study of Herzog [51], hypoconnectivity in the δ band between frontal, temporal, parietal and posterior areas emerged as the most relevant EEG feature, showing excellent AUC values in the discrimination of bvFTD from HC (figure 4). One study found right frontotemporal hypoconnectivity in bvFTD, which correlated with deficits in a naturalistic social text task [52].

• Download figure
• Open in new tab

Figure 4.

AUROC values in FTD studies. If multiple values were provided, only the highest value was reported. Studies were organized according to subpopulation. bvFTD= behavioral variant Frontotemporal Dementia, FTD= Frontotemporal Dementia, PPA= Primary Progressive Aphasia, rsEEG=resting-state EEG, AUC= Area under the curve, Sen=sensitivity, Spe= specificity, Acc= accuracy

Microstates were investigated in two studies. One study found that microstate duration C was decreased in FTD and that the sequence of activation from C to D was reversed [53]. Another study found that microstates vary with disease progression, showing an initial increase in microstate D, and a decrease as disease duration prolongs in CHMP2B-FTD [54]. These conflicting results could be explained by the different stages and variants of FTD. More specifically, microstate C is linked to the insular-cingulate network, linked to changes in personality typical of bvFTD. In contrast, CHMP2B-FTD presents with early impairments in executive functioning, with behavioural symptoms occurring at a later stage of the disease [54]. Microstate D has previously been linked to the fronto-parietal network, the initial increase followed by a decline in microstate D activation may be interpreted as a temporary compensatory mechanism. Studies using ML classification achieve encouraging AUC values between 0.78-0.96 for bvFTD and 0.78 for PPA versus HC (see figure 4). Moral-Rubio [55] also discriminated between nfvPPA, svPPA and lPPA variants, with 58% accuracy.

Task based EEG

Only one out of 10 studies included a task-based Go/No go paradigm, showing differences in θ and δ modulation related to impaired inhibition [56].

rsEEG

The most consistent finding in DLB compared to HC is slowing of the EEG pattern. More specifically, results show a leftward shift in mean dominant frequency (DF) from α range to pre-α or high-θ (6-7.5 Hz), in anterior [57], occipital [58, 59], posterior [60], or widespread areas [61]. Similarly, increased power in low frequency bands [61, 62] and decreased power in high frequency bands [61], as well as decreased θ/α ratio [58], differences α/θ and α/δ ratio [60] are observed. Using the θ/α ratio, DLB patients can be discriminated from HC with 76.7% sensitivity, 61.7% specificity and 66.7% accuracy [63]. In DLB patients with visual hallucinations, parietal δ activity achieves 75% sensitivity, 85% specificity, 81% accuracy in discrimination from HC. Posterior α reactivity from EC to EO is reduced in DLB, suggesting altered neural vigilance in the occipital lobe, enabling discrimination of DLB with 95.2% sensitivity [64].

In terms of connectivity, DLB patients show consistent network disorganization. DLB was characterized by network randomization and reduced connectivity in the α and β band [59, 65, 66] combined with increased network complexity in the high-θ band [59]. Similarly, combining connectivity strength of the β band with graph features of α band achieved an 76% accuracy in differentiating DLB from HC in the study of Mehraram [67]. In terms of dynamic connectivity, DLB patients show reduced α band information flow between posterior to anterior areas [68]. Functional source connectivity suggests cortical disconnection in DLB as both intra- and interhemispheric lagged linear connectivity (LCC) in the α range are reduced. Interhemispheric LCC in α range achieves good discrimination [69]. In the study of Kai [62], decreases in interhemispheric coherence (Icoh) and intrahemispheric coherence (Hcoh) were besides α also pronounced for δ, β and θ frequencies. In terms of large-scale resting state networks (RSN), connectivity decreases are found in the visual and sensorimotor network [65].

Two studies on microstates found contrasting results. In the study of Schumacher [70] microstate duration of all microstates was increased with reductions in numbers of microstates per second in DLB. Lamos [71] found the opposite direction of changes in prodDLB, with increased occurrence of all microstates, and shorter mean duration and increased occurrence of microstate B, which is associated with the visual network.

Task-based EEG

Three studies included a task-based EEG paradigm. One study used an auditory oddball paradigm and found a reduced and delayed P300 amplitude, as well as a P300 latency gradient inversion in DLB [72]. A longer latency of the P300 was also found in a visual oddball task [73]. Interestingly, oddball tasks reveal differences in EEG oscillations reflecting typical DLB symptoms. More specifically, DLB patients show a decreased event-related δ synchronization combined with impaired α and β suppression during both visual and auditory oddball tasks, and decreased θ band activity in a visual task [73, 74]. From a clinical perspective, decreased event-related θ oscillations and decreased α suppression during a visual oddball task may serve as neurophysiological correlates of attention and visual dysfunction in DLB. Power in δ band was able to discriminate DLB from HC with good sensitivity and specificity [74].

AD versus DLB

Studies comparing prodAD and prodDLB consistently find a more pronounced slowing in prodDLB compared to prodAD [75–77]. To account for interindividual variability, individual α frequency peaks (IAF, maximum power peak between 6-14Hz) and transition frequency between α and θ band (TF, minimum power density between 3-8Hz) can be used. Reduced mean TF and mean IAF values are consistent observations across DLB studies [63, 69, 70]. In the study of Babiloni [78] mean TF and mean IAF were found to be lowered in both prodAD (mIAF=8.8Hz, mTF = 5.4) and prodDLB compared to HC (mIAF = 9.4Hz, mTF=96.3Hz) with more pronounced reductions in prodDLB (mIAF =7.8Hz, mTF =4.7Hz). In the discrimination of prodAD from prodDLB, sensitivity values for spectral metrics range between 41 to 78.3%, while specificity values range between 66.7 to 97% (see figure 6) [76, 78]. The leftward shift in the power spectrum in DLB was also confirmed in the study of Massa [75], where α/θ ratio was found to be decreased in prodDLB compared to prodAD, and in the study of Schumacher [76], where increases in θ/α ratio were correlated with more severe cases of prodDLB. In the study of Schumacher [76], β power showed the highest AUC (0.71) in discriminating prodAD from prodDLB, with 61% sensitivity and 81% specificity. Discrimination between DLB and AD with high β power as the most important factor, shows good sensitivity and specificity (see figure 6) [68]. Babiloni [79] found that inter- and intrahemispheric LCC values in the α band were reduced in both prodAD and prodDLB compared to HC. While intra-hemispheric LCC α2 was best to differentiate prodDLB from HC, interhemispheric LCC global α2 was best to differentiate prodAD from HC. However, discrimination between prodAD and prodDLB remained low (AUC <0.7).

AD versus FTD

While both FTD and AD show progressive connectivity alterations compared to HC, profound frequency- and location dependent differences can be found in network organization, which are linked to the core areas of the diseases. Compared to bvFTD, demAD patients show lower connectivity in the α and δ band in posterior regions and a widespread higher connectivity in the θ band [80]. bvFTD patients showed an inverse pattern, with preserved posterior alpha connectivity, but lower θ activity in posterior and frontal areas. The Minimum Spanning Tree (MST) analyses indicate that frontal networks appear to be selectively involved in bvFTD, whereas in AD, global efficiency is reduced [80]. In the study of Franctiotti [81], the typical main hub in HC is lost in FTD at dementia onset and replaced by frontal local hubs, while network organization is largely preserved in demAD. In the same study, global clustering was able to distinguish between FTD (PPA and bvFTD) and demAD with moderate sensitivity but high specificity (see figure 6). In the study of Bonanni [82], network alterations predominantly targeted the frontal region in pFTD, while in prodAD, mutual information in the left local anterior region discriminated prodFTD from prodAD with good sensitivity but low specificity (see figure 6). Using the same metric, discrimination of pFTD from HC reached 89% sensitivity and 90% specificity. For the discrimination of prodAD from HC, MI in the posterior connections achieved high sensitivity (89-100%) and specificity (85-100%). Interestingly, these effects were evident in the prodromal stages of the disease but diminished with progression, which is suggestive of hyperconnectivity as a temporary compensatory mechanism to account for the effects of neurodegeneration in the core areas of the specific disease.

• Download figure
• Open in new tab

Figure 5.

AUROC values in DLB studies. If multiple values were provided, only the highest value was reported. Studies were alphabetically organized. DLB= Dementia with Lewy Bodies, DLB VH+= DLB with visual hallucinations, DLB MMSE+= DLB with pathological Mini Mental State Examination scores, rsEEG= resting-state EEG, tbEEG= task-based EEG, AUC= Area under the curve, Sen=sensitivity, Spe= specificity, Acc= accuracy.

• Download figure
• Open in new tab

Figure 6.

AUROC values in discriminative studies. If multiple values were provided, only the highest value was reported. prodAD= prodromal Alzheimer’s Disease, prodDLB= prodromal Dementia with Lewy Bodies, prodFTD= prodromal Frontotemporal Dementia, DLB= Dementia with Lewy Bodies, demAD= dementia due to Alzheimer’s Disease, PPA= Primary Progressive Aphasia, bvFTD= behavioral variant Frontotemporal Dementia, rsEEG= resting-state EEG, tb-EEG=task-based EEG, AUC= Area under the curve, Sen=sensitivity, Spe= specificity, Acc= accuracy.