Overview of major findings
The first bibliometric analysis of articles on EEG in the MCI published from 2005 to 2022 was presented, including countries, institutions, authors, journals, historical development, and research frontiers. The number of annual publications has steadily increased in recent years, surpassing 300 for the first time in 2021. The recent development track can be divided into two stages: 2005–2016 that exhibited a slow development, and 2018–2019 and 2020–2021 that exhibited a rapid development. As a result, additional in-depth studies of the EEG for cognitive impairment are likely to be published in the coming years. Of the 10 most productive institutions, 8 were in Europe and 2 were in North America. To some extent, this demonstrates the more mature application of EEG in the developed countries. The United States and Italy had the highest number of publications and citations among more than 80 countries that have published EEG-related articles for MCI. Notably, China ranked third as the only developing country among the top 10 most productive countries. This could be related to the higher prevalence of MCI in China, which has attracted researchers’ attention. According to the statistics, the overall prevalence of MCI in China was estimated to be 15.5% by 2018, involving approximately 38.77 million patients (Jia et al., 2020). When the most cited authors were taken into account, the top 10 authors made significant contributions to EEG for MCI. The three articles published by Stam CJ were all in the top five, and their high LCS confirmed his significant contributions to the field. Stam CJ concentrated on exploring functional brain networks in cognitive impairment and used several relevant approaches for EEG analysis, including graph theoretical analysis (Stam et al., 2009), small-world network model (Stam et al., 2007), and non-linear analysis (Stam, 2005). Scholars concentrated more on the brain network mechanisms underlying cognitive impairment and other neurological disorders, and the above-mentioned EEG analysis methods capable of characterizing spatially complex brain networks were further applied and developed.
Knowledge base
The more frequently a publication is cited, the more essential it is in a given field. As a result, the most frequently cited publication or publications with great impact factors can be regarded as a knowledge base in a specific field (Shen et al., 2022). Table 6 shows that the top 10 cited articles included 4 clinical trials, 4 guidelines, and 2 reviews.
First, 4 clinical trials concentrated on MCI diagnosis and the investigation of neurological mechanisms, particularly in association with AD. Compared with patients with subjective memory impairment, the likelihood of synchronization was significantly lower in the upper alpha and beta bands in AD, while these bands were not significantly different in the MCI group (Pijnenburg et al., 2004). This demonstrated that the diagnostic value of the synchronization likelihood in AD and MCI groups required further exploration. However, the conversion of MCI to AD could be predicted using a combined analysis of EEG rhythms by the low-resolution brain electromagnetic tomography and coherence analysis (Rossini et al., 2006). Similarly, it has been shown that cortical source analysis of the EEG rhythms by the low-resolution brain electromagnetic tomography has a predictive value to some extent in the early identification of MCI patients who may develop to AD (Babiloni et al., 2006c). Although these findings are valuable, a large number of high-quality studies are required to confirm them due to the insufficient sample size. EEG is effective not only in the diagnosis and prognosis of MCI, but also in investigating the mechanisms underlying cognitive impairment. For instance, a method of global field synchronization was used to evaluate neural networks in patients with MCI and AD, in order to examine the hypothesis of a “disconnection syndrome” (Koenig et al., 2005).
Second, all the 4 guidelines were developed by the task force released by the National Institute on Aging and the Alzheimer’s Association. Guidelines have been revised or updated mainly for MCI, as well as for core diagnostic criteria, pathophysiological properties, and biomarkers of AD. Reviews on EEG dynamics and cortical neural networks have been published separately, and they both provided a detailed summary of EEG studies on MCI and AD (Jeong, 2004; Babiloni et al., 2016).
Overall, according to the contents of the 10 highly cited references, the knowledge base in this area primarily concentrated on the diagnosis and identification of early MCI stages of AD. Several methods for EEG analysis are constantly explored and validated.
Research hotspots and trends
By analyzing the co-citation reference timeline view, the past research hotspots and the current research trends could be found. We will focus on the clusters that are closely related to our subjects as well as those with significant impact. Figure 6C illustrates the following findings: early research concentrated on “#2 multicentric study” and “#8 functional connectivity,” interim studies concentrated on “#3 resting-state” and “#7 EEG MRI,” whereas current studies concentrated on “#1 Lewy bodies,” “#6 bidirectional link” and “#9 deep learning.”
A multicentric study approach was used in the embryonic stage of this discipline, which not only allowed for a more comprehensive reflection of the overall status of the research question, but also demonstrated the high academic impact of researchers in the field. Functional connectivity, defined as any correlation between time series of brain activity recorded across distinct brain areas, may represent interactions between dispersed brain regions (Kramer et al., 2008). Researchers have paid further attention to alterations in neural connectivity among different brain regions rather than only on focal abnormalities in specific brain regions. Numerous studies have shown that AD patients have abnormal segregated and integrated connection patterns, which are characterized by loss of small-world network properties (Stam, 2004; Stam et al., 2007). EEG-related studies further concentrated on neural networks, which could present novel perspectives into the neural mechanism in cognitive impairment.
“Resting-state” is noteworthy in this field in the medium term. A study has shown that the cortical sources of resting-state electroencephalographic alpha rhythms deteriorated across time in subjects with amnesic MCI (Babiloni et al., 2014). Compared with healthy elderly subjects, mild AD patients had a higher power of widespread delta sources and a lower power of posterior alpha sources (Babiloni et al., 2006b). Resting-state EEG rhythms can be used as a widely available marker to track amnesic MCI patients in large clinical trials.
Subsequently, EEG analytical methods were assessed. The processing of EEG signals continues to evolve dynamically, whether in the field of MCI or in other neuroscience fields. Furthermore, multiple EEG analytical methods influenced each other and developed together. As shown in Figure 6C, the “#8 functional connectivity” cluster received attention prior to 2010, indicating that investigators concentrated on cortical connectivity during this period. At this stage, the research concentration was on methods, such as connectivity analysis and graph theoretical analysis, which mainly targeted information interactions between brain regions and brain networks.
The “#7 EEG MRI” cluster suggested that multimodal neuroimaging techniques gradually developed and formed a new trend between 2010 and 2015. Multimodal neuroimaging, the main representative of EEG-fMRI, can frequently provide more complete and complementary brain information. EEG-fMRI is a technique that proposed in Ives et al. (1993) using seminal trials on patient safety and data quality, combining the high temporal resolution of EEG and the high spatial resolution of fMRI, and it has increasingly matured in recent years (Laufs et al., 2008). An EEG-fMRI study of 14 MCI subjects and 21 cognitively healthy controls indicated that in MCI, there is a clearly aberrant coupling in several networks dominated by an anticorrelation in the posterior cingulate cortex (Michels et al., 2021).
According to “#9 deep learning” cluster, with the rapid development of artificial intelligence and computer science after 2015, EEG research is shifting to a model of multidisciplinary interdisciplinary and intelligent analysis. Machine learning has emerged as the most popular and fastest growing EEG analytical method. Machine learning can extract meaningful information from noisy, multidimensional EEG signals, paving the way for automated clinical EEG analysis (Gemein et al., 2020). Deep learning is a class of machine learning, which is an algorithm developed on artificial neural networks to perform characterization learning on data by multiple processing layers. Early diagnosis and differential diagnosis of MCI have been achieved by two different deep learning architectures, including modified convolutional and convolutional autoencoder neural networks. The outcomes demonstrate a 10% increase in accuracy rate over similar research techniques, indicating that the deep learning is a promising technique for processing EEG signals (Fouladi et al., 2022).
Mild cognitive impairment detection generally involves four main steps: data preprocessing, feature extraction, feature selection, and classification (Yin et al., 2019). In recent years, many algorithms have achieved high classification accuracy rates. For example, one study utilized the piecewise aggregate approximation technique for EEG data compression, investigated the permutation entropy and auto-regressive model features, and finally utilized the extreme learning machine technique to achieve classification accuracy of 98.78% for distinguishing MCI and healthy controls (Siuly et al., 2020). Another study utilized spectral-temporal analysis to extract features from both MCI and healthy control data, derived an appropriate feature subset using a new 3-D evaluation algorithm, and achieved an accuracy of 96.94% using a support vector machine classifier (Yin et al., 2019). Recent studies have compared different machine learning and deep learning strategies for MCI detection. For instance, a deep learning framework based on the gated recurrent unit model achieved the greatest classification accuracy of 96.91% when compared to the classical machine learning strategy of early K-nearest Neighbor et al. (Alvi et al., 2022). Additionally, many methods utilizing different feature extraction and machine learning algorithms have also reached high levels of accuracy (Buscema et al., 2015; Oltu et al., 2021). Currently, most studies focus on EEG feature extraction, while signal processing methods and more advanced machine learning algorithms remain underexplored. Thus, achieving a broader range of applications requires further optimization and exploration.
Some studies have found neurophysiological links between dementia and Lewy bodies (DLB), MCI, and AD. Cortical neuronal synchronization at delta and alpha frequencies differed between ADMCI and DLBMCI patients (Babiloni et al., 2018). However, similar abnormalities have been found in AD and DLB patients during the prodromal phase of MCI. The source connectivity may reflect a common cholinergic impairment in the prodromal state of both AD and DLB because widespread interhemispheric and intrahemispheric lagged linear connectivity solutions in alpha sources were abnormally lower in this stage (Babiloni et al., 2019).
The “#6 bidirectional link” cluster refers to the existence of a pathophysiological association between AD and epilepsy. And some studies have shown links between MCI and epilepsy as well. For example, a clinical study found that amnestic mild cognitive impairment (aMCI) patients with epilepsy developed symptoms of cognitive decline 6.8 years earlier than aMCI patients without epilepsy (Vossel et al., 2013). Neurophysiological examination of a sample of 63 aMCI patients revealed that 31.75% MCI patients had epileptiform activity (Cuesta et al., 2022). The neuroelectrophysiological link between MCI and epilepsy needs more high-quality evidence to confirm, and its common pathogenesis and clinical relevance may be a hot spot for future research.
Scholars have explored the mechanisms underlying MCI and AD itself, whereas some researchers are currently searching for the intrinsic association between various disorders (e.g., epilepsy, DLB, and Parkinson’s disease) and cognitive dysfunction. In addition, the quality of a paper cannot solely be measured by citations. Some newly published articles may have lower citation counts but this does not necessarily mean they are of poorer quality. In terms of EEG analytical methods, earlier EEG spectrum analysis can provide some descriptive information; methods such as multimodal neuroimaging or graph theoretical analysis can provide us with neural network mechanisms of disease; and the latest machine learning methods can improve diagnostic accuracy. Each of these EEG analytical methods has its own advantages. Although new methods are continuously being developed and updated, some are still limited by small sample sizes and high computational costs. Future studies should fully consider the advantages of various analytical methods and explore the most suitable analysis methods for clinical applications.