The Ability of Electroencephalography Using Statistical Pattern Recognition to Predict Conversion from Subtypes of Mild Cognitive Impairment to Dementia: A 5-Year Follow-Up Study Open Access

Driven by the knowledge that pathological findings in Alzheimer’s disease (AD) probably emerge 10 to 20 years before AD progresses to dementia, there has been a shift to diagnose AD at a stage of mild cognitive impairment (MCI) or even earlier [1]. Accordingly, treatment aiming to slow down the progression of AD should be initiated during the stage of MCI, which is possible with new treatments [2, 3]. Meta-analyses show that neuropsychology, beta-amyloid (a-beta), and phosphorylated tau proteins (p-tau) in cerebrospinal fluid (CSF), as well as magnetic resonance imaging (MRI), fluorodeoxyglucose-positron emission tomography (PET), and amyloid PET, are all useful tools in diagnosing AD at a preclinical stage [4‒8]. Recent studies have also demonstrated that concentration of a-beta and a-beta 42/40 ratio in plasma correlates significantly with concentration of a-beta in CSF and amyloid PET [9‒12]. Measurements of tau proteins, especially p-tau 181, 217, and 231 in plasma, have also proven valuable in diagnosing AD [10, 13, 14]. However, the ability of the plasma biomarkers to predict progression from MCI to AD dementia (ADD) is limited [15‒17].

Most diagnostic research has focused on AD. Little is known about which biomarkers can predict conversion from MCI to dementia with Lewy bodies (DLBs), frontotemporal dementia (FTD), and vascular dementia (VaD) [16]. Thus, there is a need to explore which biomarkers can be useful to predict non-ADD. Electroencephalography (EEG) could be such a biomarker. EEG is noninvasive and inexpensive, and can be applied outside academic clinics.

Traditional qEEG utilizes fast Fourier transform (FFT) to convert the brain’s discrete time-domain signals into frequency components. Although widely used, FFT assumes signal stationarity and linearity, overlooking the brain’s dynamic, nonlinear, and nonstationary nature [18]. Alternative EEG-based methods for diagnosing ADD and tracking MCI progression have been explored in several reviews but will not be discussed here [19‒21]. One way to use qEEG as a diagnostic tool is statistical pattern recognition (SPR) method, which can be based on specific EEG features extracted from a large dataset. Previously, it has used various spectral features from conventional EEG recordings to classify the EEG of the patient into specific disease groups and thereby provide diagnostic information [22, 23]. The advantage is the automatic labeling of diagnosis, but the validity is dependent on the quality of the training dataset that should contain EEGs from a large group of patients with both MCI and dementia diagnosis as well as a control group.

Previous studies have shown that EEG can be useful in diagnosing DLB and supporting the AD diagnosis [24‒29]. Further, EEG could predict conversion from MCI to ADD and DLB, but varying results are reported [30‒37]. Spectral power is the feature that in some studies has been suggested as a potential marker of progression of MCI to AD or DLB [35, 36]. However, no larger studies investigating a combination of different EEG spectral features have been reported and there is a lack of studies examining the ability of EEG to predict conversion from different MCI subgroups to various dementia entities. In the current literature, we did not find any studies that specifically examined utility of qEEG to predict conversion from dysexecutive MCI (deMCI) to dementia.

It is known that people with amnestic MCI (aMCI) convert to ADD more often than people with non-aMCI, probably because impaired episodic memory is an early sign of AD [38, 39]. Impaired executive function or a combination of impairment in memory and executive function could also be the first signs of AD [40, 41]. Moreover, impaired executivefunction is found early in FTD and reported in an early phase of subcortical dementias, such as dementia due to small vessel disease, Parkinson’s disease, or DLB [42‒44]. In summary, as EEG is proven to be valuable in diagnosing DLB, we hypothesized that EEG could be useful in predicting conversion from deMCI to subcortical dementias, or to AD mixed with subcortical pathology. To our knowledge, conversion of deMCI to dementia with subcortical pathology has not been evaluated using quantitative EEG. Thus, the aim of the study was to test this hypothesis by using a fully automatic EEG analysis applying the SPR method.

Patients with MCI from five academic memory clinics were followed up at intervals of 1, 2, and 5 years after the baseline diagnostic assessment. The centers were located at the Department of Geriatrics at Landspitali, Reykjavik; the Department of Neurology, Rigshospitalet, Copenhagen; the Department of Neurology at Zealand University Hospital, Roskilde; the Department of Geriatrics, Haraldsplass Hospital, Bergen; and the Department of Geriatrics at Oslo University Hospital, Oslo. In all, 118 patients were recruited for participation. Five were excluded due to incorrect diagnoses, leaving 113 participants. Demographic and clinical characteristics of the 113 are shown in Table 1. During follow-up, 11 dropped out, of which 2 died and 9 withdrew. The drop-outs did not differ from the remaining 102 regarding demographic and cognitive characteristics. All drop-outs had at least one follow-up assessment after baseline, and the results from the latest assessment were used in the analyses.

The diagnostic assessment consisted of a history taken from the patients and a close family member, a battery of cognitive tests covering various cognitive domains, a physical examination, blood tests, evaluation of depression by the Mongomery Aasberg Scale (MADRS), and an MRI (n = 66) and/or CT (n = 88) of the brain [45, 46]. Scheltens’ scale was used to rate medial temporal lobe atrophy (MTA), and Fazekas scale was used to rate white matter changes on MRI [47, 48]. Further, an FDG or amyloid PET scan was conducted in 28 participants and an a-beta and p-tau in CSF were done in 41 patients. The following cognitive tests were included in the analyses at baseline and at 1, 2, and 5 years of follow-ups: the Norwegian version of the Mini Mental Status Examination (MMSE, score 0–30); the clock drawing test using Shulman’s rating (0–5); and the 10-word test from the Consortium to Establish a Registry for Alzheimer’s Disease (CERAD) battery, measuring immediate memory (0–30), delayed recall (0–10), and recognition (0–20) [49‒52]. A lower score indicates poorer performance on all tests. We further applied the Trail Making Test A and B (TMT-A and TMT-B), measuring the time to complete the tests correctly. Longer time indicates worse performance [53, 54].

The diagnoses based on the assessments and medical records were made by at least two experienced physicians in consensus and independently of the EEG results. The Winblad criteria were used to diagnose MCI, but the subclassification was slightly adjusted [55]. Patients with non-anamnestic MCI were grouped according to the most dominant symptom as confirmed by one or more cognitive tests and supported by the history from the patient and/or a proxy. As the centers applied different neuropsychological tests, we did not define specific tests to be used in the subclassification. deMCI was based on poor results on executive tests and a normal result on tests measuring episodic memory. Aphasic MCI was based on poor results on naming objects and observation, and visuospatial MCI was based on performance on visuospatial tests. Winblad’s definition of aMCI was accepted, and an unspecified MCI label was applied when it was difficult to define one single dominant symptom. ADD at follow-up was diagnosed using the DSM-5 and the McKhann criteria; VaD employed the NINDS-AIREN criteria; DLB applied the revised consensus criteria; and FTD utilized the Lund-Manchester criteria [56‒60]. Unspecified dementia was used when the patient did not fulfil any of these criteria.

EEGs were recorded using NicoletOne EEG System from Natus^® at all sites. The IS 10–20 system was used for placement of 19 electrodes, and the features extracted from the EEGs were evaluated using the average montage of the EEG signal. Two bipolar electrooculography channels and one electrocardiogram were applied to monitor artifacts depending on the center. The sample frequency for the recordings varied between sites. The EEG recording lasted for approximately 10 min and consisted of 5 min of resting with eyes closed followed by alternating eyes open and closed for at least 5 min. The patients were alerted if they became visibly drowsy.

Subsequent analysis was done in the MATLAB environment from MathWorks^® (V R2018a). As part of the preprocessing, the data were re-referenced to average montage to minimize the impact of global artifacts. Due to varying sample frequency, the recordings were downsampled to 128 Hz. For analysis, a minimum of 2.5-min recording in the resting state eyes closed was selected.

The specific section of each recoding chosen for the processing of EEGs was selected by a trained technician to secure that artifacts were minimal, and the length of the section was at least 150 s. Prior to feature extraction, the chosen segment was preprocessed by applying an 8th-order Butterworth band-pass filter with the chosen band (0.1–70 Hz) to eliminate potential low- and high-frequency disturbances for the signal. The features extracted from the EEG recording and used in the evaluation of a dementia index (DI) were retrieved according to the recommendations of the Pharmaco-EEG Society [61]. The society recommends that the signal should be segmented into two-s segments overlapping by 1 s. The signal is then analyzed segment by segment, and the feature values are estimated by evaluating the expected value over all the segments. This can be achieved by various means, for instance, by using the average value or an alternative robust measure. Using a robust measure minimizes the impact of outliers and hence reduces the influence of potential artifacts in the signal. We used the simplest robust estimate, i.e., the median of the feature values. The features used were all related to the spectral properties of the recording. Discrete FFT was applied to estimate the spectral properties of the signal [62].

The used analytic method is based on an advanced mathematical model, the SPR method, that aims to recognize and classify EEG recordings based on comparison with normative data from well-defined patients with various dementia disorders and healthy controls [22]. The analysis relies on recordings from 19 electrodes. The electrodes used are named according to the international 10–20 system [63]: Fp1, Fp2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T3, T4, T5, T6, Fz, Cz, and Pz. If the FFT components for each of the electrodes, segments, and discrete frequencies considered are denoted by σcij, where c∈{1,2,…,19} indicates the channel, i∈{1,…,N} the segment of the N segments considered, and j∈{1,…,90} the discrete frequencies $12Hz,1Hz,…,45Hz$⁠, the full spectral resolution covariance between channels c and k is then expressed by $χckij=σcij*σkij*$⁠. These covariances constitute the base features used for analysis and evaluation of the classification index values. To determine the core features relied on, principal components (PCs) were determined based on the Mentis Cura database of EEG recordings [22]. PC analysis was performed on data from subjects with a diagnosis in the database. This was done separately for each covariance. The PCs were then ranked according to their individual discriminatory properties in separating various dementia groups in the database. The discriminatory properties were determined according to the area under the curve (AUC) of the receiver operator characteristic (ROC) curve. We used the two best performing components from each of the covariances to extract the core features used for evaluation of the index.

If Pckαj denotes the two chosen PCs, α∈{1,2}, for electrode pair (c, k) at frequencies j∈{1,…,90}, the core features considered for analysis then become $Cckα=Ei{∑j=190χckijPckαj$⁠. Figure 1 illustrates typical examples of the PCs for one covariance, P111j and P112j, which describe how the spectral bands are weighted for the analysis. The PCs can be related to the classical EEG power bands, δ (1–4 Hz), θ (4–8 Hz), α (8–13), and β (13–30). Then, PC1 corresponds to the difference between the combined δ and θ power and the β power, while PC2 is a weighted measure of the total power with slightly more emphasis on α and β power. The index value for an individual recording is evaluated from these features by $I=∑ckαCckαβckα+β1AA+β2AA2+ρ$⁠, where A is the age of the subject in years. The classification coefficients β_ckα, $βiA$⁠, and ρ were determined by using a combination of genetic algorithms to optimize the number of features used. Support vector machine, an SPR technique, was applied in the Mentis Cura database, which contains EEG data from people with various dementia diagnoses and healthy controls. This was done separately for men and women, resulting in separate gender-dependent DIs. It should be noted that the DI was developed to separate people with dementia from non-demented subjects, and not to predict conversion from MCI to dementia. The results of the classification are presented as a DI score ranging from zero to 100. A higher index score indicates a higher probability of dementia.

We applied the SPSS version 29. To test for differences between groups, we used Student’s t test for continuous data and the χ² test for categorical data. We applied ROC analysis and calculated AUC, sensitivity (SS), and specificity (SP) values to test the power of EEG to separate converters from non-converters. The results on TMT-A and TMT-B correlated highly. We used TMT-B in the analysis, as this is a measurement of executive function. The three scores on the CERAD 10-word test were also highly correlated, and we used the delayed recall score, as this is a measure of episodic memory. Further, we used the average MTA score, which is atrophy on the left side plus atrophy on the right side divided by two. The drop-outs were included in the analyses, and the status (conversion or non-conversion) at time of drop-out was used. p value <0.05 was set to define statistical significance.

During follow-up, 60 (53%) converted to dementia. At 1 year, 12 had converted; between 1 and 2 years, 26 had converted; between 2 and 5 years, 22 had converted. Baseline demographic and clinical characteristics of all participants, converters and non-converters, are shown in Table 1. Concentration of a-beta, t-tau, and p-tau in CSF did not differ between converters and non-converters (data not shown).

Table 2 shows the number of the various subtypes of MCI at baseline and the diagnoses after 5 years. The majority had a diagnosis of aMCI (n = 80) at baseline, and most of the converters progressed to ADD (n = 47). Of the 17 patients with deMCI, 5 converted into ADD (mean MTA 2.3; mean Fazekas 1.9 at baseline), 2 to VaD (MTA 1.5; Fazekas 3), 1 to DLB (MTA 1.0; Fazekas 3), and 1 to unspecific dementia (MTA 1.0; Fazekas 1). Seven of the deMCI converters had hypertension, and 4 had diabetes.

The mean EEG DI among converters was 34.8 (SD 15.6), and among non-converters, it was 28.8 (SD 15.6), p = 0.03. Table 3 shows the DIs for the different subtypes of MCI. It shows clearly how the DI score of the deMCI converters differed from deMCI non-converters and that the deMCI converters had the highest DI score compared to all other MCI subgroups. Table 4 summarizes the results of the ROC analyses, comparing converters with non-converters after 2 years (38 vs. 75 patients) and 5 years (60 vs. 53 patients). Our EEG method did not separate aMCI converters from aMCI non-converters, whereas the EEG’s ability to predict conversion among deMCI patients was superior. AUC of 0.92 with SS of 89% and SP of 100% indicates that our EEG method is valid to separate converters from non-converters for this group of patients. We further found that DI separated aMCI converters from deMCI converters with AUC of 0.81 (95% CI 0.65–0.96).

In Table 5, we summarized baseline characteristics of the aMCI and deMCI patients. A distinct pattern emerged. The deMCI patients had better scores on the MMSE and the 10-word recall score and poorer scores on the TMT-B and MADRS compared to the aMCI patients. The Fazekas score differed substantially between the two groups. No differences of significance were found between the two groups regarding sex, age, and educational level (data not shown).

We found that 60 (53%) of 113 patients with MCI converted to dementia at the 5-year follow-up, and 38 converted within 2 years. The ability of EEG applying the SPR method to predict conversion to dementia was moderate to poor (AU below 0.70), when we included all MCI patients in the ROC analysis. This was unexpected, as in a previous study using the same method, we found that EEG predicted conversion from MCI to dementia with an AUC of 0.78 [23]. We have no robust explanation for this discrepancy. Further, the EEG method was not suitable to predict conversion from aMCI to dementia, which is also in contrast to the result of the previous study [23]. One explanation could be that the present study included more aMCI patients who converted to ADD compared to the sample of the previous study. This remains a hypothesis, as we did not subclassify the MCI patients in the previous study [23]. Another explanation could be that the MCI patients in the previous study were more cognitively impaired. This is supported by slightly poorer results on cognitive tests of the MCI patients who converted to dementia after 1 and 5 years in the previous study [23]. Another explanation could be that our SPR method was designed as a diagnostic tool and not trained specifically for the prognosis of MCI.

Other studies have investigated the use of EEG as a prognostic tool for patients with aMCI using different types of EEG features. They reported different results how EEG could predict conversion to ADD [30‒33, 35, 37]. Several factors may explain the differences. One possible explanation is that some previous studies may have suffered from overfitting as they were for the most part trained on small datasets, highlighting the need for validation of the findings. To sum up, our findings are in accordance with the conclusion of a systemic review suggesting that further studies are needed to understand the role of EEG as a biomarker to predict conversion from MCI to ADD [21].

The ability of our EEG method to predict conversion from the small group of deMCI to dementia was remarkably strong, as the ROC analysis showed AUC of 0.92 corresponding to an SS of 89% and SP of 100%. The EEG DI further separated converters of aMCI and deMCI patients with an AUC of 0.81. Thus, our data support the hypothesis that EEG could be useful to predict conversion from deMCI to a variety of dementia etiologies with subcortical pathology. Other studies including patients with subcortical pathology, typically DLB, have found that EEG is powerful to predict conversion from MCI to DLB, to diagnose DLB, and to separate DLB from ADD [24‒26, 29, 34, 36, 64, 65]. In addition, EEG has also been found useful, but not perfect, in diagnosing patients with major vascular cognitive impairment, which includes patients with small vessel disease [66].

By inspecting Table 2, we observe that 5 of the 9 deMCI converters received an ADD diagnosis at follow-up. Does this fit with our hypothesis? Possibly, these 5 patients had a combination of AD and a disorder causing subcortical pathology. These 5 had a mean MTA score of 2.3 and a mean Fazekas score of 1.9 and 4 had hypertension or diabetes, indicating mixed neurodegenerative and vascular pathology. A neuropathological AD study showed that in addition to the typical hippocampus pathology, 11% of AD patients had a hippocampal sparing subtype and 14% had a limbic-predominant subtype [67]. These subtypes have been described in MRI studies, and in one of them significant white matter changes were found in patients with the limbic-predominant subtype, defined as prominent atrophy in medial temporal lobes, indicating a combination of AD and small vessel disease [68‒70]. This is in accordance with the fact that hypertension and diabetes are both risk factors for ADD and small vessel disease.

How can it be explained that executive impairment in our study could be an early sign of subcortical dementia? Studies have shown that reduced executive function is associated with subcortical brain damages seen in small vessel disease, DLB, and Parkinson’s disease, because of an impaired connection between neurons in frontal lobes and subcortical areas [42‒44, 71]. In the present study, we found that deMCI patients had more subcortical white matter changes expressed by higher Fazekas scores, poorer TMT-B scores, and more depressive symptomatology (Table 5), suggesting a disruption of frontal-subcortical pathways [63‒65]. Thus, we believe that impaired executive function could be an early symptom in patients with subcortical pathology of various etiologies and explain why deMCI converters had different dementia diagnoses at follow-up. This needs to be addressed in a bigger cohort and by combining EEG with other biomarkers for improved predictive accuracy.

Taken together, we postulate that any subcortical pathology can cause executive cognitive impairment. In some patients, the pathology will not be associated with progression to dementia; in some, it will lead to progression to subcortical dementia such as VaD and DLB. If there is a combination of AD and subcortical pathology, we will have a mixed dementia, often diagnosed as ADD, because it is difficult to know how much the subcortical pathology contributes to the dementia syndrome. However, the final question arises: why and how is EEG useful in the prediction of conversion from deMCI to dementia with subcortical pathology? Probably, the EEG findings registered in DLB patients are not specific for this disease but can be found in other disorders with a disruption of the frontal-subcortical pathways [34, 65, 72‒74].

A conversion rate of 53% from MCI to dementia is high, and higher compared with rates reported in previous studies [23, 75‒79]. However, studies including memory clinic patients have reported similar conversion rates [23, 77‒79]. Memory clinic patients are highly selected, which may be the cause for the high conversion rate. Findings from the Norwegian Registry of Persons Assessed for Cognitive Symptoms (NorCog), a memory clinic register including more than 20,000 persons, can explain the selection. In this register, more than 60% were APOE e4 positive [80, 81].

This study has limitations. The sample is small, especially in the group of patients with deMCI. Thus, the results should be evaluated with caution and termed as a pilot study. Second, we did not use the traditional way of subdividing MCI. Our subclassification was based on extensive clinical experience and use of cognitive tests. We have not conducted a formal testing of its validity, but by inspecting Table 5, there is no doubt that the deMCI patients diverge largely from the aMCI patients. They performed better on the MMSE, the 10-word recall test, and poorer on TMT-B, and had higher scores on MADRS and higher Fazekas scores compared to the aMCI patients. This pattern indicates both impaired executive function and subcortical pathology. It could have been of interest to further examine the association between the DI and the performance on various cognitive tests. This is outside the scope of the study, but we suggest that it should be done in future larger studies. A third limitation is that only 58% of the patients had MRI examinations, which limited the use of MTA and Fazekas scores. However, more than 80% of the deMCI patients had MRIs. Finally, we cannot claim with certainty that the MRIs were evaluated in the same way. Although the neuroradiologists at the five academic centers were experienced in how to rate MTA and white matter changes an inter-rater reliability, checks should have been carried out. The strengths of this study are the harmonization of the clinical examinations, as well as EEG registration through many years of collaboration with the comprehensive diagnostic assessment, including a variety of diagnostic tools, and that at least two experienced clinicians settled the diagnoses in creating consensus.

First, we would like to thank Kristinn Johnsen, the former CEO of Mentis Cura who developed and described the SPR method used in the study. He further analyzed the EEG recordings without any knowledge of the clinical data. In addition, we would like to thank the Kavli Charitable Trust in Bergen, Norway, which provided funding for this study.

After being orally informed and given time to study the written information, the participants consented to participate by signing the informed consent document. The project was approved by Ethics Committees in Iceland (VSN-17/184) and Norway (REK-2016/676). In Denmark, the project was reviewed by the Ethics Committee but did not require approval (H-17023350); however, the same procedures as stated above were followed including written informed consent. Agreements were approved by regulatory authorities to transfer the Case Record Forms (CRFs), EEG, and MRI recordings between the centers.

The following authors have no conflict of interest to disclose: L.-O.W., C.S.M., P.H., T.E.G., M.L.B., B.B.A., M.N., A.R.O., and J.S. K.E were members of the journal’s Editorial Board at time of submission. D.F. was a consultant for BioArctic and has received honoraria from Esteve. He has received funding from the Swedish Research Council (Vetenskapsrådet, Grant 2022-00916), the Center for Innovative Medicine (CIMED, Grants 20200505 and FoUI-988826), the regional agreement on medical training and clinical research of Stockholm Region (ALF Medicine, Grants FoUI-962240 and FoUI-987534), the Swedish Brain Foundation (Hjärnfonden FO2023-0261, FO2022-0175, FO2021-0131), the Swedish Alzheimer Foundation (Alzheimerfonden AF-968032, AF-980580, AF-994058), the Swedish Dementia Foundation (Demensfonden), the Gamla Tjänarinnor Foundation, the Gun och Bertil Stohnes Foundation, the Åke Wiberg Foundation, StratNeuro, Funding for Research from Karolinska Institutet, Neurofonden, and the Foundation for Geriatric Diseases at Karolinska Institutet, as well as contributions from private bequests.

The study was funded by the Kavli Charitable Trust in Bergen, Norway. The funder had no role in the design, data collection, data analysis, and reporting of this study.

K.E., L.-O.W., and J.S. planned the study and wrote the research protocol; K.E. performed the analytic work and wrote the first draft of the manuscript; K.E., L.-O.W., and J.S. evaluated the analyses. L.-O.W., C.S.M., P.H., M.L.B., T.E.G., B.B.A., D.F., M.N., A.R.Ø., and J.S. gave valuable input to the draft manuscript and approved the final version, which was written by K.E.

Due to national regulations in Norway, the entire data file will not be available without special permission from the Regional Ethics Committee in South-East Norway before 2030. Further inquiries can be directed to the corresponding author.