Application of machine learning in EEG-based dementia diagnosis: Classification and differential diagnosis

Abstract

Background

Electroencephalogram (EEG) is a promising, non-invasive method for identifying the presence of Alzheimer's disease (AD) by recognizing specific brain activity patterns associated with the disease. However, research on the correlation between EEG and the degree of cognitive impairment is still lacking.

Objective

In this study, we employ machine learning models to explore the potential of EEG in distinguishing different levels of cognitive impairment and various types of dementia.

Methods

A total of 431 participants, including 77 cognitively unimpaired (CU), 167 patients with mild dementia, 110 patients with moderate dementia, and 77 patients with severe dementia were enrolled. Among them, 91 patients have detailed biomarker results to support differential diagnosis, with 77 AD and 14 frontotemporal dementia. After feature extraction, the rule-based representation learning was used to train models for EEG-based classification tasks.

Results

Our model can effectively differentiate between CU and moderate-to-severe dementia (AUC 0.8475), as well as between CU and AD patients in individuals under 65 (AUC 0.8170). However, our preliminary analysis was not able to effectively distinguish between different types of dementia. It is also challenging to differentiate between CU and mild dementia groups, as well as between the moderate and severe dementia.

Conclusions

Our study suggests that EEG might be used not only in the early identification of AD, but also in the diagnosis and monitoring of the entire dementia spectrum, encompassing various stages and types of cognitive decline.

Keywords

Alzheimer's disease cognition dementia diagnosis electroencephalography machine learning

Introduction

Approximately 55 million people worldwide live with dementia.^1–3 With the gradual aging of its population, China has the largest population of people with dementia, accounting for approximately 20% of the total number worldwide.^4,5 In China, the prevalence of dementia in the elderly over 65 years old is about 5.6%.⁶ The rising prevalence of dementia seriously affects the quality of life of patients and puts a tremendous financial burden on the society, becoming an unavoidable challenge to confront.^7,8

Dementia is a syndrome characterized by the chronic and acquired loss of two or more cognitive functions, due to brain disease or injury, severely affecting the ability to perform daily activities. Neurodegenerative disorders, primarily Alzheimer's disease (AD), are leading causes of dementia. Currently, there are no available treatments that can alter the course of any neurodegenerative forms of dementia.⁹ Therefore, early diagnosis and intervention in dementia are crucial for delaying disease progression and improving the quality of life for patients.^10,11

The current diagnosis of dementia primarily relies on cognitive function assessment and the collection of medical histories indicative of decreased daily living abilities. Cognitive impairment screening questionnaires and neuropsychological tests also contribute to the evaluation and diagnosis.¹² The Mini-Mental State Examination (MMSE) is the most widely used screening tool for dementia.¹³ However, the results are susceptible to interference from subjective factors. Moreover, due to the overlap of clinical symptoms among different types of dementia, definitive diagnosis of causes of dementia remains challenging. More comprehensive examination such as supplemental neuropathological examinations, neuroimaging studies and genetic testing are required.

Taking AD as an example, the National Institute on Aging and Alzheimer's Association (NI-AAA) introduced the ATN framework in 2018 as the benchmark for AD diagnosis.¹⁴ According to the ATN framework, AD status is determined by three biomarkers (i.e., amyloid, tau, and neurodegeneration), measured from cerebrospinal fluid (CSF) and positron emission tomography (PET) imaging.¹⁴ However, to date, the definitive diagnosis of AD can only be made through postmortem pathology. In vivo fluid and neuroimaging biomarkers are limited applied due to the high cost, technical difficulty and invasiveness. Therefore, many current studies are exploring simpler testing methods and new non-fluid biomarker, to aid in the early identification of dementia and the precise determination of its severity and type.

Electroencephalography (EEG), as a low-cost, non-invasive, stable, and highly reproducible tool for measuring neural electrical activity, is attracting increasing attention nowadays.^15–17 EEG measures the electrical activity generated by the neurons, providing insights into the conduction of neural impulse, and possesses a high temporal resolution. Therefore, it is considered to have the potential to reveal the pathology of dementia from a electrophysiological perspective, and identify dementia before permanent damage occurring in the cerebral cortex.¹⁸ Numerous studies have found characteristic difference in the EEG of patients with dementia, and confirmed the potential of EEG to distinguish patients from healthy controls with varying degrees of sensitivity and specificity.^18,19 Specifically, previous researches have mainly focused on resting-state EEG (rsEEG) and AD. These studies consistently found that compared to the healthy control, rsEEG of patients with AD chiefly exhibited a decrement in alpha rhythms (10–13 hz) source activity and an increment in theta (4–8 hz) power density oscillations.^20–22 With a more comprehensive understanding of AD, we gradually realize that AD is a disconnection syndrome caused by disruptions in neuronal connections and neural signal pathways.^23,24 The focus of EEG-related research has also shifted toward cortical neural synchronization and coupling.^23–25 Spectrum-based EEG analyses transforms the temporal information of EEG into spectral information through Fourier transform and continuous wavelet transform, effectively describing neural synchronization and coupling through power density and functional connectivity. And machine learning models, due to their ability to identify complex and non-linear patterns from these features, provide excellent tools for further analyzing EEG. In fact, international expert groups have reached a consensus on the research of rsEEG and its application in AD, and have made some recommendations on the use of machine learning models in EEG research.^15,26 This supports our subsequent research.

The main objective of our study is to collect rsEEG data from patients with different levels of dementia and use machine learning models to attempt a classification task, exploring the potential of EEG-based machine learning models for assisting in diagnosis. Furthermore, we conducted further attempts by comparing the EEG data of subjects grouped by age and different types of dementia, aiming to uncover more insights. Through our research, we hope to apply EEG for the auxiliary diagnosis of dementia, such as AD, to enhance the current clinical diagnostic system.

Methods

Participants

A total of 431 participants were utilized in this study, including 77 cognitively unimpaired (CU) and 354 patients with different severities of dementia at initial study enrollment. All patients were recruited from the Peking Union Medical College Hospital (PUMCH) dementia cohort, the Dementia Clinic of the Department of Neurology of PUMCH, between October 2016 to May 2023. Inclusion criteria as follows: (1) ≥6 years of education; (2) received a detailed clinical evaluation that included medical history taking, physical and neurological examinations, a systemic of neuropsychological tests, laboratory testing (i.e., hepatic function, renal function, homocysteine, thyroid function, folic acid, vitamin B12, blood ammonia, and rapid plasma reagin test) and brain MRI (i.e., T1, T2, flair, DWI, SWI, ASL); (3) have received neuropsychological tests and electroencephalogram (EEG) within a year. And the exclusion criteria: (1) Patients with a history of major psychiatric illness (e.g., schizophrenia, bipolar disorder) or any other central nervous system disorders other than cognitive impairment. (2) Patients with significant functional disabilities or other systemic diseases (e.g., infectious, toxic, metabolic, and neoplastic diseases) that heavily impact the cognitive function and EEG results. It is worth noting that patients with severe dementia often exhibit abnormal scores on the Hospital Anxiety and Depression (HAD) scale. However, anxiety and depression can also affect cognitive assessment results, which may blur the distinction between CU and mild dementia. Therefore, in this study, HAD was used as an exclusion criterion (HAD-A ≥ 9 or HAD-D ≥ 9) only on participants with MMSE ≥ 24.^27,28

The Institutional Review Board of PUMCH approved the study. All the participants have signed the informed consent forms.

Neuropsychological examinations and grouping

Cognitive tests included the Chinese version of the MMSE²⁹ and comprehensive neuropsychological test battery (NTB).³⁰ Functional disability was assessed using the Chinese version of Activities of Daily Living Scale (ADL), which was revised and supplemented according to the scale of Lawton and Brody.³¹

Dementia was diagnosed based on clinical judgment and/or on cognitive test performance according to the DSM-5 criteria.³² In the subsequent subgroup analyses, the diagnoses of AD and FTD were based on their respective clinical diagnostic criteria.^14,33–35 The cutoff points for classifying stages of dementia based on the MMSE vary slightly among different studies.^13,36–38 In this study, the grouping of CU and different severities of dementia was based on all available information including clinical history and neuropsychological measures. Participants were categorized into different dementia stages based on a comprehensive evaluation by our specialized neurologists, which included MMSE, ADL, and NTB assessments (see Figure 1).

Figure 1.

The detailed grouping procedures. CU: MMSE ≥ 24, ADL ≤ 23, HAD-A < 9 and HAD-D < 9; mild dementia: MMSE 18–23; moderate dementia: MMSE 12–17; severe dementia: MMSE ≤ 11. For participants with MMSE ≥ 24 but ADL ≥ 24, further distinction between CU and mild dementia was made based on the NTB. And participants lacking of sufficient data to support grouping were excluded from the analysis.

EEG recording and preprocessing

EEG signal was recorded at 1000 Hz from the participants in a 40-min eye-closed resting state. Participants were required to remain awake during the entire recording. Standard 19-channels montage was utilized according to the 10–20 International System (channels: Fp1, Fp2, F3, F4, F7, F8, T3, T4, T5, T6, C3, C4, P3, P4, O1, O2, Fz, Cz, Pz).

MNE-Python³⁹ and autoreject⁴⁰ are used to build the whole preprocessing workflow and automatically remove bad epochs, using customized Python scripts (Python 3.8 on a Ubuntu 18.04 machine). Briefly, the collected EEG signals were first band-pass filtered between 0.1 Hz and 50 Hz and further filtered by a 50-Hz notch filter to remove the powerline interference. The filtered data was downsampled to 250 Hz and then re-referenced using the A1 and A2 channels. We discarded the first 20 s of every recording due to a large number of recording artifacts in this period. And after that, as recommended by a guideline published recently,²⁶ we then segmented the EEG data into a series of epochs using a 10-s time window. We removed the bad epochs via autoreject.⁴⁰ A total of 77,565 epochs were finally obtained from all participants.

CSF collection and gene detection

Some participants underwent lumbar puncture to CSF for the analysis of fluid biomarkers, while others had blood drawn for DNA sequencing. The data from these participants will be included in the subsequent analysis of different types of dementia.

The CSF samples were centrifuged at 1800×g and 4°C for 10 min and stored at the temperature of −80°C. Commercial accessible ELISA kits were used for the analysis of CSF T-tau, P-tau181, and Aβ₄₂ with INNOTEST hTAU Ag, PHOSPHO-TAU, and β-AMYLOID (1–42) (Fujirebio, Ghent, Belgium). All procedures were performed by board-certified laboratory technicians, in accordance with the manufacturer's instructions. Based on our group's previous studies and biomarker diagnostic criteria, we defined participants meeting the following criteria as definitive AD: 1. CSF T-tau/Aβ₄₂ > 0.32 or CSF p-tau/Aβ₄₂ > 0.054; 2. CSF P-tau181 levels >37 pg/mL.

Genomic DNA was extracted from fresh peripheral blood leukocytes. Whole-exome sequencing using “next-generation” sequencing (NGS) technology was performed using an Illumina HiSeq (Illumina, USA) and was verified by Sanger sequencing. Participants with previously reported pathogenic mutations, as well as those with mutations of uncertain significance but exhibiting typical clinical manifestations, were included in the analysis and grouped as AD or frontotemporal dementia (FTD).

Statistical analysis

The statistical analyses were performed by SPSS version 24.0 software (Chicago, IL, USA). Continuous variables were described as mean ± standard deviation (M ± SD) and categorical variables as numbers and percentages (n, %). The χ² test was used for categorical variables. The t-test and analysis of variance (ANOVA) was used for continuous variables. ANOVA with Bonferroni post-hoc tests was applied to compare different subgroups with the least significant different. A p-value of <0.05 was considered statistically significant.

EEG feature extraction and machine learning

We use the BrainFeatures library for feature engineering.^41,42 Specifically, we employ Continuous Wavelet Transform (CWT), Discrete Wavelet Transform (DWT), Discrete Fourier Transform (DFT) to transform the raw EEG data into frequency domain and generate features describing time and frequency. And we compute between-electrode connectivity features with the Hilbert Transform. Finally, each sample was converted into a vector consisting of 5700 feature values, which belong to the five feature domains: CWT, DWT, DFT, time, and connectivity.⁴² Detailed information of the features and feature domains can be found in the Supplemental Material.

Based on different groupings, each sample was designated as either positive or negative. And after the feature extraction, the rule-based representation learning (RRL) model was employed for classification, utilizing 5-fold cross-validation to avoid overfitting. We randomly selected 80% of the samples as a training set, and the left 20% of the samples as a testing set. For each task, five experiments were conducted using different sample partitioning methods, and the average value of each metric was calculated. The corresponding receiver operating characteristic (ROC) curves were plotted and the performance of the models was evaluated by accuracy, precision, recall, F1 score, and AUC, given as

R e c a l l (R C) = \frac{T P}{T P + F N}

P r e c i s i o n (P C) = \frac{T P}{T P + F P}

A c c u r a c y (A C C) = \frac{T P + T N}{T P + T N + F P + F N}

F 1 s c o r e = 2 * \frac{R C * P C}{R C + P C}

where TP, TN, FP, and FN represent the true positive, true negative, false positive, and false negative, respectively.

Results

Demographic information

A total of 431 participants, 197 men and 234 women, aged 67.74 ± 11.13 years, were recruited. Of these, 77 (17.87%) were CU, 167 (38.75%) had mild dementia, 110 (25.52%) had moderate dementia, and 77 (17.87%) had severe dementia. The baseline demographic and cognitive date were shown in Table 1. No significant difference in sex distribution were observed among different groups (p = 0.804), while p < 0.001 for age and p = 0.013 for education years, which could be explained by the influence of age and education level on the cognitive reserve and the disease process.

Table 1.

Demographic details and MMSE score of different groups.

	Total	CU	Mild	Moderate	Severe	χ²/F
n	431	77	167	110	77
Age (y)	67.74 ± 11.13	62.53 ± 10.47	71.16 ± 10.06	68.30 ± 11.69	64.75 ± 10.57	13.978**
Sex (male, %)	197, 45.71	36, 46.75	72, 43.11	54, 49.10	35, 45.45	0.996
Education (y)	11.91 ± 4.01	13.18 ± 3.44	11.90 ± 3.49	11.44 ± 3.28	11.35 ± 5.92	3.616*
MMSE	18.08 ± 8.15	27.62 ± 1.76	22.50 ± 2.99	13.91 ± 2.33	4.90 ± 3.02	1210.235**

p-values below 0.05 are marked by *; p-values below 0.001 are marked by **.

EEG-based classification results between Cu and various degree of dementia

In the preliminary experiments, we found that the performance of the machine learning model significantly declined in multi-class classification tasks. Thus, based on actual clinical needs, we established and trained the model to better perform binary classification tasks in distinguishing CU and varying degrees of dementia. The models’ performance and the ROC curves for five different binary classification tasks are presented in Table 2 and Figure 2.

Figure 2.

The ROC curves of the model in binary classification tasks. (a) ROC curve for the detection of CU versus dementia. (b) ROC curve for the detection of CU versus mild dementia. (c) ROC curve for the detection of CU versus moderate & severe dementia. (d) ROC curve for the detection of mild dementia versus moderate & severe dementia. (e) ROC curve for the detection of moderate dementia versus severe dementia. Since we used 5-fold cross-validation to avoid the problem of overfitting, each figure includes five light-colored lines representing the five random trials, along with one dark blue line representing the average across the five trials.

Table 2.

Performance of the model in different binary classification tasks.

	CU vs. dementia^a	CU vs. mild	CU vs. (moderate & severe)	Mild vs. (moderate & severe)	Moderate vs. severe
ACC (%)	77.32 ± 4.32	66.65 ± 3.04	78.47 ± 3.66	69.98 ± 1.97	58.03 ± 4.56
PC (%)	61.49 ± 2.52	61.17 ± 4.20	75.92 ± 4.23	70.25 ± 2.11	56.46 ± 3.79
RC (%)	57.95 ± 0.80	59.86 ± 5.21	75.41 ± 2.87	70.00 ± 2.08	56.40 ± 3.78
F1 (%)	58.71 ± 0.99	59.73 ± 5.06	75.41 ± 3.52	69.88 ± 2.01	56.40 ± 3.78
AUC (%)	71.32 ± 4.32	66.37 ± 4.64	84.75 ± 3.73	76.98 ± 2.81	59.59 ± 4.56

The “dementia” here means all the cognitively impaired groups, including the mild, moderate and severe.

Briefly, the model demonstrated moderate performance in the binary classification task of distinguishing CU from dementia (mild, moderate and severe groups), achieving an accuracy of 77.32% and a ROC-AUC of 0.713. In this context, the model had a better performance in distinguishing CU from moderate & severe dementia (ACC 78.47%, AUC 0.8475) and distinguishing mild dementia from moderate & severe dementia (ACC 69.98%, AUC 0.7698). However, the model struggled to discriminate CU from mild dementia (ACC 66.65%, AUC 0.6637) or moderate from severe dementia (ACC 58.03%, AUC 0.5959) using EEG data.

Comparison among subgroups

To further explore the differences in EEG data between different onset age groups and the role of the model in the differential diagnosis of various types of dementia, we conducted additional subgroup analysis. Based on their clinical characteristics, fluid biomarkers and genetic testing results, we selected 77 definitive AD and 14 FTD patients for the further analysis. At the same time, we used 65 years of age as the cutoff to further divide the CU and definitive AD participants into the young group and the old group, followed by a corresponding subgroup analysis.

Table 3 presents the demographic information and cognitive status of each subgroup.

Table 3.

Demographic details and MMSE score of subgroups.

	Young CU	Old CU	Young AD	Old AD	FTD	χ²/F
n	51	26	54	23	14
Age (y)	56.69 ± 6.78	74.00 ± 5.91	56.44 ± 5.58	70.26 ± 3.39	61.00 ± 8.60	56.638**
Sex (male, %)	25, 49.02	11, 42.31	26, 48.15	13, 56.52	11, 78.57	1.400
Education (y)	12.64 ± 3.21	14.23 ± 3.68	11.30 ± 3.05	12.30 ± 3.53	11.25 ± 4.34	3.770**
MMSE	27.90 ± 1.71	27.08 ± 1.74	12.22 ± 7.42	15.57 ± 6.83	17.14 ± 10.41	60.764**

p-values below 0.05 are marked by *; p-values below 0.001 are marked by **.

Different age groups

The results of the model distinguishing between CU and definitive AD of different age groups are shown in Table 4 and Figure 3.

Figure 3.
The ROC curves for the model distinguishing AD patients from CU in different age groups. (a) ROC curve for the detection of the young CU versus the early onset AD. (b) ROC curve for the detection of the old CU versus the late onset AD. (c) ROC curve for the comparison between the early onset and the late onset AD patients.

Table 4.
Comparison between different age groups.

CU (≤65 y) vs. Early onset AD CU (>65 y) vs. Late onset AD Early onset AD vs. Late onset AD

ACC (%) 74.78 ± 4.01 59.73 ± 3.28 62.36 ± 8.08

PC (%) 75.68 ± 4.29 60.82 ± 6.39 54.51 ± 10.69

RC (%) 74.66 ± 3.91 58.87 ± 4.70 54.78 ± 10.20

F1 (%) 74.52 ± 3.99 57.48 ± 3.31 54.44 ± 10.61

AUC (%) 81.70 ± 2.77 60.65 ± 4.24 57.72 ± 15.30

According to the results, we could observe that the model was nearly unable to distinguish between the normal population and the AD patients in individuals over the age of 65. However, in participants under the age of 65, the model was able to effectively discriminate the early onset AD patients from CU, achieving an ACC of 74.78% and the AUC of 0.8170.

Different type of dementia

To explore the potential application of this model in the differential diagnosis of dementia, we trained the model and attempted to make pairwise comparisons among the CU, definitive AD, and FTD groups. The performance of the model in those binary classification tasks were shown in Table 5 and Figure 4.

Figure 4.
The ROC curves for the model in differential diagnosis. (a) ROC curve for the detection of CU versus AD & FTD groups. (b) ROC curve for the detection of CU versus AD. (c) ROC curve for the detection of CU versus FTD. (d) ROC curve for the detection of AD versus FTD.

Table 5.
Comparison between different types of dementia.

CU vs. (AD & FTD) CU vs. AD CU vs. FTD AD vs. FTD

ACC (%) 66.00 ± 3.52 70.98 ± 6.94 83.83 ± 2.02 85.10 ± 2.83

PC (%) 66.28 ± 3.53 71.00 ± 6.97 55.32 ± 9.37 63.49 ± 10.20

RC (%) 66.03 ± 3.47 70.71 ± 7.22 52.58 ± 6.02 57.33 ± 9.17

F1 (%) 65.87 ± 3.49 70.61 ± 7.39 52.25 ± 7.27 57.30 ± 10.32

AUC (%) 71.99 ± 4.02 77.45 ± 7.94 64.03 ± 9.54 70.52 ± 11.99

First, when we compared AD and FTD patients as a combined group with CU (AUC 71.99), we found that the model's performance in distinguishing these groups was essentially consistent with its performance in the binary classification task of differentiating CU from all cognitively impaired groups (AUC 71.32, as shown in Table 2). The model demonstrated moderate performance in the binary classification task of distinguishing CU from AD (ACC 70.98%, AUC 0.7745). However, its ability to differentiate between AD and FTD was relatively limited (AUC 0.7052). However, in the preliminary analysis of AD versus FTD classification, the model's ability was relatively limited (AUC = 0.7052, F1 = 57.30%).

Discussion

EEG and machine learning for dementia diagnosis

An increasing number of studies have highlighted the potential application of EEG in the diagnosis of dementia, but most of them have primarily focused on AD.^18,19,43 In the present study, based on the results of cognitive assessments and neuropsychological tests, we considered dementia patients with varying degrees of cognitive impairment as a heterogeneous group, aimed to explore the potential of using EEG and machine learning methods to distinguish dementia patients from CU individuals. Based on our results, we found that the model effectively distinguished between CU and the cognitively impaired group. Moreover, when the scope of participants was further narrowed to individuals under 65 years old or those with moderate to severe dementia (MMSE < 18), independently of the age, the models’ performance improved significantly. When we aimed to apply the model for differential diagnosis, we found that although it effectively distinguished between CU and the patients (AD and FTD groups), it struggled to differentiate between AD and FTD.

Previous studies have found that, compared to cognitively normal individuals, AD patients exhibit a widespread reduction in alpha power and an increase in theta power in their EEG patterns.^15,20,22 Functional connectivity and neural synchrony may also serve as potential biomarkers for identifying distinct EEG patterns associated with AD.^21,25,26 Our model has the ability to distinguish the heterogeneous group of dementia patients from normal individuals, rather than being limited to AD alone. This may suggest that changes in EEG patterns are not only present in cognitive decline related to AD, but that there could be a shared pathophysiological basis underlying dementia caused by various neurodegeneration.^44,45

EEG features across different dementia severity levels

In our study, the comparison of EEG characteristics across different dementia severity levels contributes to enriching the existing literature and enhancing our understanding. At present, research on EEG in the progression of dementia primarily focuses on distinguishing between MCI and dementia patients, as well as predicting the possibility of progression from MCI to dementia. Fewer studies have focused on whether there are differences in EEG patterns between patients with varying degrees of dementia. In a cross-sectional retrospective study conducted in Taiwan, researchers classified AD patients into different severity levels based on Clinical Dementia Rating (CDR) and MMSE scores.²² However, although the study suggested that the amplitude modulation of low-frequency oscillations in EEG becomes more pronounced as AD progresses, it did not include a comparison of EEG patterns between AD patients at different severity levels, but focused only on comparing AD patients with MCI and normal controls. In another study conducted in Korea, although the researchers aimed to investigate the relationship between EEG patterns and cognitive decline using MMSE scores, they only divided the participants into three groups with cutoffs at 28 and 24.⁴⁶ This study lacked more detailed cognitive groupings and comparisons. However, the aforementioned studies also suggest the potential for using EEG to differentiate between varying degrees of dementia and to monitor disease progression.^22,46 In this study, the model we developed was able to effectively differentiate between mild and moderate-to-severe dementia based on EEG patterns (ACC 69.98%, AUC 0.7698). Furthermore, the model achieved its highest performance in the binary classification task of distinguishing CU from moderate-to-severe dementia (ACC 78.47%, AUC 0.8475). This further indicates that there are indeed significant differences in EEG patterns among patients with varying degrees of dementia. And EEG biomarkers may be able to explain cognitive decline and could potentially be applied in monitoring disease progression. Unfortunately, our model did not perform well in distinguishing between CU and mild dementia patients, nor in differentiating between moderate and severe dementia patients. At present, its clinical applicability remains limited.

EEG between early-onset and late-onset AD

One of the key innovations in our study is the comparison between early-onset and late-onset AD, given the impact of age on cognitive reserve and the clinical differences between these groups. Our results showed that, compared to the age-matched CU group, the model performed better at identifying the EEG patterns of early-onset AD than late-onset AD. We hypothesize that the early-onset AD is associated with more distinct EEG patterns and more severe brain function alterations. And this premise is supported by previous studies suggesting that the early-onset AD patients exhibit more pronounced slowing of spontaneous oscillatory activity.^47,48 On the other hand, healthy aging can also lead to EEG changes similar to those seen in AD.^49–51 Thus, age-related EEG alterations may reduce the differences between CU and AD patients’ EEG patterns, which could explain why the model performs better at distinguishing CU from AD patients in groups aged 65 years or younger.

Challenges in differential diagnosis

We also explored a highly discussed question: whether EEG-related biomarkers can differentiate AD from other dementia subtypes. The preliminary results indicate that it is still difficult to distinguish between AD and FTD groups based on EEG patterns. Considering the significant disparity in the number of participants between the two groups (77 versus 14), the F1-score of 57.30% provides a more reliable reflection of the model's actual performance, indicating that it is not well-suited for this classification task. In addition to the difference in sample size, the potential shared pathophysiological mechanisms by which different types of dementia affect brain function and EEG patterns may also be one of the reasons why differential diagnosis is challenging. However, an increasing number of recent studies suggest that different types of dementia exhibit distinct EEG biomarker alteration patterns, and these specific changes can be identified through EEG measurements.^45,52,53 However, these studies inevitably face a common issue: the sample sizes for dementia types other than AD are often too small, making the overall results prone to bias due to sample size disparities. And to date, no rsEEG biomarker has been identified that can reliably differentiate individual patients. At present, solely relying on EEG for the differential diagnosis of dementia is not feasible. Future research needs to collect sufficient data from patients with other types of dementia to address this limitation.

Clinical implications of EEG

In summary, we believe that EEG remains a valuable clinical tool for detecting AD and other types of dementia. Firstly, EEG can non-invasively reflect brain function, providing a direct measure of synaptic integrity and neural signal transmission. This fills a gap in the current ATN research framework for AD, which lacks an explicit representation of electrophysiological changes. Secondly, due to the high temporal resolution of EEG, we believe that compared to MRI, which primarily provides high spatial resolution and structural insights, EEG has the potential to detect dementia-related changes at an earlier stage. Moreover, as discussed above, an increasing number of studies suggest that EEG has the potential to differentiate between different types of dementia. Although our study did not confirm this capability, it remains a potential advantage of EEG over MRI. Thirdly, considering its cost-effectiveness, EEG has the potential to be a scalable tool for large-scale dementia screening. Especially in resource-limited and technologically underdeveloped regions, EEG is more accessible compared to MRI. With the advent of disease-modifying therapies for AD, EEG can serve as a non-invasive, real-time monitoring tool to assess treatment efficacy and disease progression.^54–56 Initially, EEG could only be interpreted visually, which inevitably introduced errors and subjective bias. The development of machine learning models has provided tools for precise EEG data analysis. We believe that in the future, machine learning models will play a crucial role in EEG-related research.

Study limitations and future directions

There are several limitations in our study. First, the internationally recognized method for grading dementia severity primarily relies on the CDR scale. Due to the lack of corresponding assessment results, our study primarily used MMSE and NTB as the basis for grading the severity of dementia.^36,57 Secondly, in our study, the sample size for dementia types other than AD was limited, preventing us from drawing reliable conclusions on whether EEG can be used for the differential diagnosis of dementia. Future research should focus on accumulating more samples for other dementia subtypes. Thirdly, in our study, the EEG preprocessing pipeline lacks Independent Component Analysis (ICA), which is considered the gold standard for removing physiological artifacts. In this study, after acquisition, the EEG recordings were first visually inspected, and segments with obvious artifacts were removed to ensure that the retained data were generally clean and the subsequent results reliable. Considering that ICA might inadvertently remove neural signals, we deliberately chose not to apply it during the experimental design. Instead, we employed autoreject and bandpass filtering for noise reduction. However, this approach inevitably results in the partial retention of Electrocardiogram and Electrooculogram and affected the quality of our data. Fourth, the RRL model utilizes a 5700-dimensional feature vector and such an uncurated high-dimensional feature space may capture noise correlations and reduce clinical translatability. Moreover, the RRL model learns rule-based patterns from the feature vectors and assigns importance to the rules rather than to individual features, making it difficult to identify and interpret the most relevant features in relation to AD-specific pathology. In future studies, it may be beneficial to integrate feature selection methods such as LASSO or SHAP to first identify the important features prior to feeding them into the RRL model. This approach may enhance the electrophysiological understanding of AD-related pathological changes.

EEG requires further research and supporting evidence before it can be effectively integrated into the existing diagnostic framework for dementia. We believe that more precise cognitive assessments and classification systems, a larger and more diverse dataset, and optimized EEG data processing methods will further enhance data quality and improve the performance of our model. Furthermore, as a high-temporal-resolution tool capable of capturing dynamic changes in brain function and neural activity, EEG holds great potential for longitudinal studies. We believe that future research should focus on longitudinal data comparisons.

Conclusion

In this study, we found that using EEG and RRL models, we could effectively distinguish dementia patients from normal individuals. Our model achieved high performance in differentiating between CU and moderate-to-severe dementia, as well as between CU and AD patients in individuals under 65. However, our preliminary attempts were still insufficient to effectively distinguish between different types of dementia.

Supplemental Material

sj-docx-1-alz-10.1177_13872877251360331 - Supplemental material for Application of machine learning in EEG-based dementia diagnosis: Classification and differential diagnosis

Supplemental material, sj-docx-1-alz-10.1177_13872877251360331 for Application of machine learning in EEG-based dementia diagnosis: Classification and differential diagnosis by Yixuan Huang, ZhenYu Li, Fangzhou Liu, Bo Li, Chenhui Mao, Liling Dong, Shanshan Chu, Wei Jin, Jianyong Wang and Jing Gao in Journal of Alzheimer's Disease

	CU (≤65 y) vs. Early onset AD	CU (>65 y) vs. Late onset AD	Early onset AD vs. Late onset AD
ACC (%)	74.78 ± 4.01	59.73 ± 3.28	62.36 ± 8.08
PC (%)	75.68 ± 4.29	60.82 ± 6.39	54.51 ± 10.69
RC (%)	74.66 ± 3.91	58.87 ± 4.70	54.78 ± 10.20
F1 (%)	74.52 ± 3.99	57.48 ± 3.31	54.44 ± 10.61
AUC (%)	81.70 ± 2.77	60.65 ± 4.24	57.72 ± 15.30

	CU vs. (AD & FTD)	CU vs. AD	CU vs. FTD	AD vs. FTD
ACC (%)	66.00 ± 3.52	70.98 ± 6.94	83.83 ± 2.02	85.10 ± 2.83
PC (%)	66.28 ± 3.53	71.00 ± 6.97	55.32 ± 9.37	63.49 ± 10.20
RC (%)	66.03 ± 3.47	70.71 ± 7.22	52.58 ± 6.02	57.33 ± 9.17
F1 (%)	65.87 ± 3.49	70.61 ± 7.39	52.25 ± 7.27	57.30 ± 10.32
AUC (%)	71.99 ± 4.02	77.45 ± 7.94	64.03 ± 9.54	70.52 ± 11.99

Footnotes

Acknowledgements

We would like to thank the EEG department at Peking Union Medical College Hospital for collecting all EEG data from the subjects. We would also like to express our gratitude to the School of Basic Medicine, Peking Union Medical College, for testing the fluid biomarkers of the participants.

ORCID iDs

Yixuan Huang

ZhenYu Li

Fangzhou Liu

Bo Li

Chenhui Mao

Liling Dong

Jianyong Wang

Jing Gao

Ethical considerations

The Institutional Review Board of PUMCH approved the study.

Consent to participate

All the participants have signed the informed consent forms.

Author contribution(s)

Yixuan Huang: Data curation; Investigation; Visualization; Writing – original draft.

ZhenYu Li: Formal analysis; Methodology; Software.

Fangzhou Liu: Formal analysis; Investigation; Software.

Bo Li: Data curation.

Chenhui Mao: Formal analysis; Funding acquisition; Project administration.

Liling Dong: Formal analysis; Project administration; Supervision.

Shanshan Chu: Data curation.

Wei Jin: Data curation.

Jianyong Wang: Conceptualization; Methodology; Software.

Jing Gao: Conceptualization; Funding acquisition; Project administration; Supervision; Writing – review & editing.

Funding

The authors disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was financially supported by the National Key Research and Development Program of China (nos. 2020YFA0804500 and 2020YFA0804501), the CAMS Innovation fund for medical sciences (CIFMS) (nos. 2021-I2M-1-020 and 2020-I2M-C&T-B-010), the National Natural Science Foundation of China (nos. 81550021 and 30470618), and the Science Innovation 2030-Brain Science and Brain-Inspired Intelligence Technology Major Project (no. 2021ZD0201106).

Declaration of conflicting interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Data availability statement

The data supporting the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.

Supplemental material

Supplemental material for this article is available online.

References

Fang

Xie

Schenkel

, et al. A research agenda for ageing in China in the 21st century (2nd edition): focusing on basic and translational research, long-term care, policy and social networks. Ageing Res Rev 2020; 64: 101174.

Pérez Palmer

Trejo Ortega

Joshi

. Cognitive impairment in older adults: epidemiology, diagnosis, and treatment. Psychiatr Clin North Am 2022; 45: 639–661.

Prince

Wimo

Guerchet

, et al. World Alzheimer report 2015. The global impact of dementia: An analysis of prevalence, incidence, cost and trends. London: Alzheimer's Disease International, 2015.

Jia

Chu

, et al. Prevalence, risk factors, and management of dementia and mild cognitive impairment in adults aged 60 years or older in China: a cross-sectional study. Lancet Public Health 2020; 5: e661–e671.

Zhang

Eggink

Zhang

, et al. Needs and views on healthy lifestyles for the prevention of dementia and the potential role for mobile health (mHealth) interventions in China: a qualitative study. BMJ Open 2022; 12: e061111.

Huang

Wang

, et al. Prevalence of mental disorders in China: a cross-sectional epidemiological study. Lancet Psychiatry 2019; 6: 211–224.

GBD 2016 Neurology Collaborators . Global, regional, and national burden of neurological disorders, 1990–2016: a systematic analysis for the Global Burden of Disease Study 2016. Lancet Neurol 2019; 18: 459–480.

Liu

Jiang

Peng

, et al. Association between physical activity and mild cognitive impairment in community-dwelling older adults: depression as a mediator. Front Aging Neurosci 2022; 14: 964886.

Tisher

Salardini

. A comprehensive update on treatment of dementia. Semin Neurol 2019; 39: 167–178.

10.

Weller

Budson

. Current understanding of Alzheimer's disease diagnosis and treatment. F1000Res 2018; 7: F1000 Faculty Rev-1161.

11.

Sanford

. Mild cognitive impairment. Clin Geriatr Med 2017; 33: 325–337.

12.

Arvanitakis

Shah

Bennett

. Diagnosis and management of dementia: review. JAMA 2019; 322: 1589–1599.

13.

Folstein

McHugh

. “Mini-mental state”. A practical method for grading the cognitive state of patients for the clinician. J Psychiatr Res 1975; 12: 189–198.

14.

Jack

Jr Bennett

Blennow

, et al. NIA-AA research framework: toward a biological definition of Alzheimer's disease. Alzheimers Dement 2018; 14: 535–562.

15.

Babiloni

Arakaki

Azami

, et al. Measures of resting state EEG rhythms for clinical trials in Alzheimer's disease: recommendations of an expert panel. Alzheimers Dement 2021; 17: 1528–1553.

16.

Meghdadi

Stevanović Karić

McConnell

, et al. Resting state EEG biomarkers of cognitive decline associated with Alzheimer's disease and mild cognitive impairment. PLoS One 2021; 16: e0244180.

17.

Rossini

Di Iorio

Vecchio

, et al. Early diagnosis of Alzheimer's disease: the role of biomarkers including advanced EEG signal analysis. Report from the IFCN-sponsored panel of experts. Clin Neurophysiol 2020; 131: 1287–1310.

18.

Paraskevaidi

Martin-Hirsch

Martin

. Progress and challenges in the diagnosis of dementia: a critical review. ACS Chem Neurosci 2018; 9: 446–461.

19.

Mitsukura

Sumali

Watanabe

, et al. Frontotemporal EEG as potential biomarker for early MCI: a case-control study. BMC Psychiatry 2022; 22: 289.

20.

Musaeus

Engedal

Høgh

, et al. EEG Theta power is an early marker of cognitive decline in dementia due to Alzheimer's disease. J Alzheimers Dis 2018; 64: 1359–1371.

21.

Smailovic

Koenig

Kåreholt

, et al. Quantitative EEG power and synchronization correlate with Alzheimer's disease CSF biomarkers. Neurobiol Aging 2018; 63: 88–95.

22.

Chu

Lei

, et al. A holo-spectral EEG analysis provides an early detection of cognitive decline and predicts the progression to Alzheimer's disease. Front Aging Neurosci 2023; 15: 1195424.

23.

Babiloni

Blinowska

Bonanni

, et al. What electrophysiology tells us about Alzheimer's disease: a window into the synchronization and connectivity of brain neurons. Neurobiol Aging 2020; 85: 58–73.

24.

Teipel

Grothe

Zhou

, et al. Measuring cortical connectivity in Alzheimer's disease as a brain neural network pathology: toward clinical applications. J Int Neuropsychol Soc 2016; 22: 138–163.

25.

Babiloni

Lizio

Marzano

, et al. Brain neural synchronization and functional coupling in Alzheimer's disease as revealed by resting state EEG rhythms. Int J Psychophysiol 2016; 103: 88–102.

26.

Babiloni

Barry

Başar

, et al. International federation of clinical neurophysiology (IFCN) – EEG research workgroup: recommendations on frequency and topographic analysis of resting state EEG rhythms. Part 1: applications in clinical research studies. Clin Neurophysiol 2020; 131: 285–307.

27.

Zigmond

Snaith

. The hospital anxiety and depression scale. Acta Psychiatr Scand 1983; 67: 361–370.

28.

Bjelland

Dahl

Haug

, et al. The validity of the hospital anxiety and depression scale. An updated literature review. J Psychosom Res 2002; 52: 69–77.

29.

Zhang

Hong

. The mini-mental state examination in the Chinese residents population aged 55 years and over in the urban and rural areas of Beijing. Chinese J Neurol 1999; 32: 149–153.

30.

Harrison

Minassian

Jenkins

, et al. A neuropsychological test battery for use in Alzheimer disease clinical trials. Arch Neurol 2007; 64: 1323–1329.

31.

Lawton

Brody

. Assessment of older people: self-maintaining and instrumental activities of daily living. Gerontologist 1969; 9: 179–186.

32.

American Psychiatric Association. Diagnostic and statistical manual of mental disorders. 5th ed. Arlington, VA: American Psychiatric Publishing, 2013.

33.

McKhann

Knopman

Chertkow

, et al. The diagnosis of dementia due to Alzheimer's disease: recommendations from the national institute on aging-Alzheimer's association workgroups on diagnostic guidelines for Alzheimer's disease. Alzheimers Dement 2011; 7: 263–269.

34.

Rascovsky

Hodges

Knopman

, et al. Sensitivity of revised diagnostic criteria for the behavioural variant of frontotemporal dementia. Brain 2011; 134: 2456–2477.

35.

Gorno-Tempini

Hillis

Weintraub

, et al. Classification of primary progressive aphasia and its variants. Neurology 2011; 76: 1006–1014.

36.

Chua

Choo

RWM

NHL

, et al. Mapping modified mini-mental state examination (MMSE) scores to dementia stages in a multi-ethnic Asian population. Int Psychogeriatr 2019; 31: 147–151.

37.

Juva

Sulkava

Erkinjuntti

, et al. Staging the severity of dementia: comparison of clinical (CDR, DSM-III-R), functional (ADL, IADL) and cognitive (MMSE) scales. Acta Neurol Scand 1994; 90: 293–298.

38.

Lim

Chin

Lam

, et al. Clinical dementia rating: experience of a multi-racial Asian population. Alzheimer Dis Assoc Disord 2005; 19: 135–142.

39.

Gramfort

Luessi

Larson

, et al. MEG And EEG data analysis with MNE-python. Front Neurosci 2013; 7: 267.

40.

Jas

Engemann

Bekhti

, et al. Autoreject: automated artifact rejection for MEG and EEG data. Neuroimage 2017; 159: 417–429.

41.

TNTLFreiburg. brainfeatures, https://github.com/TNTLFreiburg/brainfeatures (2019, accessed 16 February 2025).

42.

Gemein

LAW

Schirrmeister

Chrabąszcz

, et al. Machine-learning-based diagnostics of EEG pathology. Neuroimage 2020; 220: 117021.

43.

Ieracitano

Mammone

Hussain

, et al. A novel multi-modal machine learning based approach for automatic classification of EEG recordings in dementia. Neural Netw 2020; 123: 176–190.

44.

Giustiniani

Danesin

Bozzetto

, et al. Functional changes in brain oscillations in dementia: a review. Rev Neurosci 2023; 34: 25–47.

45.

Jiao

Zhou

, et al. Neural biomarker diagnosis and prediction to mild cognitive impairment and Alzheimer's disease using EEG technology. Alzheimers Res Ther 2023; 15: 32.

46.

Choi

You

, et al. Resting-state prefrontal EEG biomarkers in correlation with MMSE scores in elderly individuals. Sci Rep 2019; 9: 10468.

47.

de Waal

Stam

Blankenstein

, et al. EEG Abnormalities in early and late onset Alzheimer's disease: understanding heterogeneity. J Neurol Neurosurg Psychiatry 2011; 82: 67–71.

48.

de Waal

Stam

de Haan

, et al. Young Alzheimer patients show distinct regional changes of oscillatory brain dynamics. Neurobiol Aging 2012; 33: 1008.e1025–1031.

49.

Javaid

Kumarnsit

Chatpun

. Age-related alterations in EEG network connectivity in healthy aging. Brain Sci 2022; 12: 218.

50.

Perinelli

Assecondi

Tagliabue

, et al. Power shift and connectivity changes in healthy aging during resting-state EEG. Neuroimage 2022; 256: 119247.

51.

Scally

Burke

Bunce

, et al. Resting-state EEG power and connectivity are associated with alpha peak frequency slowing in healthy aging. Neurobiol Aging 2018; 71: 149–155.

52.

Nardone

Sebastianelli

Versace

, et al. Usefulness of EEG techniques in distinguishing frontotemporal dementia from Alzheimer's disease and other dementias. Dis Markers 2018; 2018: 6581490.

53.

Engedal

Snaedal

Hoegh

, et al. Quantitative EEG applying the statistical recognition pattern method: a useful tool in dementia diagnostic workup. Dement Geriatr Cogn Disord 2015; 40: 1–12.

54.

Cecchetti

Agosta

Canu

, et al. Analysis of individual alpha frequency in a large cohort from a tertiary memory center. Eur J Neurol 2024; 31: e16424.

55.

Briels

Stam

Scheltens

, et al. In pursuit of a sensitive EEG functional connectivity outcome measure for clinical trials in Alzheimer's disease. Clin Neurophysiol 2020; 131: 88–95.

56.

Jovicich

Babiloni

Ferrari

, et al. Two-year longitudinal monitoring of amnestic mild cognitive impairment patients with prodromal Alzheimer's disease using topographical biomarkers derived from functional magnetic resonance imaging and electroencephalographic activity. J Alzheimers Dis 2019; 69: 15–35.

57.

Perneczky

Wagenpfeil

Komossa

, et al. Mapping scores onto stages: mini-mental state examination and clinical dementia rating. Am J Geriatr Psychiatry 2006; 14: 139–144.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.00 MB

0.03 MB