Predicting progression from amnestic mild cognitive impairment to Alzheimer's disease using longitudinal EEG data: a 12-month cohort study

Introduction: Amnestic mild cognitive impairment (aMCI), serving as a clinical precursor to Alzheimer's Disease (AD), assumes a pivotal role in the early stages of AD prevention. The longitudinal collection of data in aMCI is imperative for monitoring disease progression and guiding clinical interventions.

Methods: Utilizing a prospective cohort design, we recruited aMCI individuals and conducted a one-year follow-up study. During this period, electroencephalogram (EEG) signals were systematically collected at regular intervals, resulting in four time points for each participant. Based on the follow-up outcomes, participants were stratified into progressive mild cognitive impairment (PMCI) and stable mild cognitive impairment (SMCI) groups. We extracted spectral, nonlinear, and functional connectivity features from the EEG data at three cross-sectional time points in the initial nine months and constructed longitudinal features between these cross-sectional assessments. The longitudinal features were fed into machine learning classifiers to predict one-year follow-up outcomes.

Results: The dynamic trends of EEG features in SMCI and PMCI patients exhibited inconsistency. Utilizing the selected longitudinal features, the support vector machine (SVM) demonstrated the best prediction performance, achieving an accuracy of 94.92%, an area under the curve of 93.25%, a sensitivity of 90.20%, a specificity of 98.80%, a positive predictive value of 98.70%, and an F1-score of 93.65%.

Discussion: By capturing trend information associated with disease progression, longitudinal EEG features contributed to enhancing prediction performance in machine learning models.

1 Introduction

AD, characterized by irreversible memory impairment, aphasia, apraxia, agnosia, and changes in personality and behavioral patterns (Ossenkoppele et al., 2015), onsets insidiously with a prolonged course. Currently, more than 50 million people worldwide are afflicted with AD (Weidner and Barbarino, 2019). Lacking effective medications, this number is expected to escalate to 131 million by 2050, imposing a substantial burden on both society and families (Winblad et al., 2016). Mild Cognitive Impairment (MCI) is a transitional stage that falls between normal age-related cognitive decline and dementia (particularly AD), representing the earliest clinically detectable stage of progression toward dementia or AD (Markesbery, 2010). AMCI, identified with memory dysfunction, is a subtype of MCI with an annual progression rate to AD ranging from 10% to 15% (Cai et al., 2020) and a lifetime conversion rate ranging from 75% to 80% (Gómez-Soria et al., 2021). Therefore, the early and accurate prediction of progression in the aMCI stage becomes a crucial issue in managing the AD continuum and alleviating its burden.

The diagnosis of MCI requires a combination of comprehensive examinations, including biomarkers, neuroimaging, cognitive tests, and neuropsychological assessments (Alzheimer's disease facts figures, 2023). This process is time-consuming, labor-intensive, and cost-prohibitive. Moreover, the insidious onset can be easily misinterpreted as age-related cognitive decline, thus significantly reducing the detection rate of MCI in clinical practice. As a non-invasive examination, EEG presents the benefits of convenience, cost-effectiveness, real-time diagnosis, and wide accessibility. It has been universally applied for the diagnosis and monitoring of disease progression in MCI (Rossini et al., 2022). Compared to task-related EEG, resting-state EEG does not require examinees to perform complex instructions and actions, making it particularly suitable for elders with cognitive decline. Currently, researchers have investigated the EEG signal characteristics in MCI populations, revealing a tendency of “high to low” frequency shift, reduced complexity, and a disconnection phenomenon (Movahed and Rezaeian, 2022). However, the heterogeneity in the disease progression of MCI individuals is also manifested in EEG signals (Libon et al., 2010), presenting substantial challenges for researchers in EEG analysis (Ding et al., 2022).

Recently, machine learning has been widely utilized in the discriminative diagnosis using EEG signals in patients with MCI (Yang et al., 2019). (Lee et al. 2022) extracted spectral, nonlinear, and functional connectivity features from EEG signals in AD and MCI cases, achieving an accuracy of 86.85%. (Perez-Valero et al. 2022) applied spectral and nonlinear features from three groups—AD, MCI, and Healthy Controls (HC)—to an MLP classifier, achieving a maximum accuracy of 95%. Furthermore, researchers including (Li et al. 2022, 2021) and (Kim et al. 2022) focused their studies on the aMCI population. They employed machine learning models to classify aMCI against HC and non-aMCI groups, respectively, achieving an accuracy approaching 90%. Although researchers have attained satisfactory classification results through feature refinement and model optimization, the substantial overlap in EEG features between MCI patients and AD or HC underscores considerable room for further enhancement in classification performance (Ding et al., 2022; Ieracitano et al., 2020).

Longitudinal data, also known as panel data, involves repeated measurements for individuals. In the medical field, longitudinal studies facilitate the capture of individual trends with dynamic trajectories beyond the scope of cross-sectional studies, providing significant benefits in early diagnosis, risk prediction, treatment planning, and prognosis assessment (Cascarano et al., 2023). Though several researchers have collected EEG signals in MCI patients and conducted longitudinal studies (Miraglia et al., 2020; Vecchio et al., 2018), there is a lack of research extracting features from multiple EEG measurements and constructing longitudinal features. In the “HC-MCI-AD” disease continuum, the progression of the condition represents a dynamic evolving process (Monllor et al., 2021). Extracting longitudinal features can offer multidimensional longitudinal information for early diagnosis and prognosis improvement.

This study utilizes a prospective cohort design, recruiting aMCI participants and conducting a one-year follow-up. After the follow-up period, patients were categorized into SMCI and PMCI groups based on whether they progressed to AD, which was in alignment with definitions from previous research (Rossini et al., 2006; Vecchio et al., 2018). Each patient underwent a total of four EEG signal measurements at equidistant intervals, with spectral, nonlinear, and functional connectivity features extracted at each time point. Subsequently, the longitudinal features were constructed to capture the dynamic trends between the first three cross-sectional assessments. Lastly, extracted longitudinal features were integrated into various machine learning classifiers to predict one-year follow-up outcomes. Our study aimed to explore the advantages of longitudinal EEG features, furthermore demonstrating the necessity of repeated EEG measurements in longitudinal studies involving the AD disease continuum.

2 Methods

The study design of our prediction framework is shown in Figure 1, which consists of five main steps: EEG data acquisition, EEG preprocessing, feature extraction, prediction, and evaluation.

Figure 1

Figure 1. Study design.

2.1 Cohort

2.1.1 General study design

Between July 1, 2022, and September 22, 2024, conducting a prospective cohort design, we recruited 92 aMCI patients from the Neurology Department of the First Affiliated Hospital of Sun Yat-sen University (SYSU). Within this cohort, 65 patients underwent EEG recordings every four months during the one-year follow-up, resulting in complete sets of four EEG records. As of the follow-up endpoint on September 22, 2024, 27 aMCI patients progressed to AD (PMCI), while 38 aMCI patients did not (SMCI). A total of 65 patients (27 PMCI, 38 SMCI) were included for our analysis.

2.1.2 Definitions of SMCI and PMCI

We recruited aMCI cases and employed a one-year follow-up. Based on whether they progressed to AD after one year, they were categorized into the PMCI and SMCI groups. The diagnosis of aMCI was based on the Petersen 2011 criteria (Petersen, 2004). The inclusion criteria for this study were as follows: (1) age of 50 years and above, (2) memory complaint usually corroborated by an informant, (3) objective memory impairment for age, (4) essentially preserved general cognitive function, (5) largely intact functional activities. The exclusion criteria were: other forms of dementia or accompanying Parkinson's disease, epilepsy, psychiatric disorders, and serious organic disease. The diagnosis of AD was based on the criteria provided by the National Institute on Aging and the Alzheimer's Association (NIA-AA) in 2011 (McKhann et al., 2011). All diagnoses of aMCI and AD were established by experienced neurologists in strict accordance with the corresponding diagnostic criteria. The demographic information of the patients is shown in Table 1.

Table 1

Table 1. The demographic characteristics of participants.

2.2 EEG data acquisition

The patients had the collection of resting-state EEG signals at four-month intervals, with a margin of ±5 days. Resting-state EEG was recorded using the Nicolet EEG machine system (Natus Medical Inc., San Carlos, CA) with a sampling rate of 500 Hz. Electrodes were positioned according to the 10–20 international system, with a total of 16 channels (Fp1, Fp2, F3, F4, C3, C4, P3, P4, O1, O2, F7, F8, T3, T4, T5, and T6). All patients were right-handed, and sufficient sleep was ensured the night before the EEG collection. During recording, patients were instructed to maintain a comfortable seated posture with their eyes closed for 5 minutes. EEG technicians continuously monitored the EEG traces and promptly alerted participants if any signs of drowsiness were detected.

2.3 EEG preprocessing

EEG signals are susceptible to various artifacts, highlighting the importance of preprocessing prior to analysis. Firstly, the raw EEG data were re-referenced using an average reference, and a finite impulse response (FIR) band-pass filter was applied to filter the EEG signals within the range of 0.1–70 Hz. Also, a notch filter was used to eliminate the 50 Hz power line interference. The EEG signals were subsequently down-sampled to 250 Hz. After joint screening by two experienced EEG examiners, bad epochs were removed and bad channels were interpolated. Then, 30-second segments of continuous EEG signals with clear background rhythms and minimal interference were selected. Following, we conducted independent component analysis (ICA) to remove common artifacts such as blinks, eye movements, and cardiac interference. Examiners applied the same procedure and criteria for preprocessing each EEG signal. The above preprocessing steps were all performed using the EEGLAB toolbox (Delorme and Makeig, 2004) in MATLAB (R2023a, MathWorks).

2.4 Feature extraction

Firstly, we extracted various cross-sectional features from each EEG signal, with each feature measured three times within the initial nine months, constituting repeated measures data. Subsequently, longitudinal features were extracted from the cross-sectional features.

2.4.1 Cross-sectional feature

For each 30s EEG signal, we extracted features in three feature sets: spectral, nonlinear, and functional connectivity.

2.4.1.1 Spectral feature

Using Welch's power spectral density (PSD) estimation (Alam et al., 2020), we transformed the preprocessed EEG signals from the time domain into the frequency domain and divided them into the following five subbands: delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–45 Hz).

(1) Power spectral density ratio (PSD ratio): Considering the variation in absolute PSD values among different patients, we calculated the relative PSD values within the aforementioned subbands for each patient, resulting in the following five ratios:

$\begin{array}{ll}Ratio1=\frac{delta}{alpha}& (1)\end{array}$

$\begin{array}{ll}Ratio2=\frac{theta}{alpha}& (2)\end{array}$

$\begin{array}{ll}Ratio3=\frac{delta}{alpha+beta}& (3)\end{array}$

$\begin{array}{ll}Ratio4=\frac{theta}{alpha+beta}& (4)\end{array}$

$\begin{array}{ll}Ratio5=\frac{delta+theta}{alpha+beta+gamma}& (5)\end{array}$

(2) Power spectral density entropy (PSDE): In each subband, a sequence of PSD values can be obtained. We used the Shannon entropy method to assess the level of disorder in this sequence of PSD values (Li et al., 2023). Assuming there are PSD series with N values within the subband, the PSDE was calculated as follows:

$\begin{array}{ll}E=-{\displaystyle \sum _{i=1}^N}{p}_{i}log{p}_i& (6)\end{array}$

where E and p_i represent the PSDE of the signal and the probability of having the i sample in the signal, respectively.

2.4.1.2 Nonlinear feature

We extracted the following three nonlinear features to capture the nonlinear characteristics of the EEG signals in aMCI patients.

(1) Permutation entropy (PE): PE is an efficient quantitative complexity measure that explores the local order structure of a dynamic time series (Bandt and Pompe, 2002), particularly in EEG signals from MCI and AD patients (Seker et al., 2021; Siuly et al., 2020). Given a scalar time series, {x(i):1 ≤ i ≤ N}. Firstly, reconstruct time series:

$\begin{array}{ll}{X}_i=\left\{x(i),x(i+\tau ),\dots ,x(i+(m-1)\tau )\right\}& (7)\end{array}$

$\begin{array}{l}i=1,2,\dots ,N-(m-1)\tau \end{array}$

where τ is time delay, and m is the embedding dimension.

Then, rearrange X_i in an increasing order:

$\begin{array}{l}\{x\left(i+\left(j_1-1\right)\tau \right)\le x\left(i+\left(j_2-1\right)\tau \right)\le \dots \le \\ x\left(i+\left(j_m-1\right)\tau \right)\}& (8)\end{array}$

There are m! permutations for m dimensions. Each vector X_i can be mapped to one of the m! permutations.

Next, the probability of the j_th permutation occurring P_j can be defined as:

$\begin{array}{ll}{P}_j=\frac{n_j}{{\displaystyle {\sum }_{j=1}^{m!}}{n}_j}& (9)\end{array}$

where n_j is the number of times the j_th permutation is occurring.

The permutation entropy of the time series {x(i):1 ≤ i ≤ N} is defined by:

$\begin{array}{ll}{H}_x(m)=-{\displaystyle \sum _{j=1}^{m!}}{P}_{j}ln{P}_j& (10)\end{array}$

when the time series is random, the H_x(m) approaches its maximum value of ln(m!); when the time series is regular, the H_x(m) approaches zero.

Finally, normalizing H_x(m) by dividing ln(m!):

$\begin{array}{ll}PE=\frac{H_x(m)}{ln(m!)}& (11)\end{array}$

(2) Sample entropy (SE): The SE is proposed by Richman and Moorman (Richman and Moorman, 2000) as an improvement over approximate entropy (AE), addressing the bias present in AE. Recently, SE has been extensively utilized for feature extraction in patients with MCI and AD (Ruiz-Gómez et al., 2018; Tsai et al., 2012).

Assume we have a time-series data set of length N = {x₁, x₂, …, x_N} with a constant time interval τ. We define a template vector of length m, such that X_M(i) = {x_i, x_i+1, x_i+2, …, x_i+m−1} and the distance function ||X_m(i), x_m(j)|| (i≠j) is to be the Chebyshev distance. We define the SE to be

$\begin{array}{ll}SE=-ln\frac{A}{B}& (12)\end{array}$

where A = number of template vector pairs having ||X_m+1(i), x_m+1(j)|| < r, and B = number of template vector pairs having ||X_m(i), x_m(j)|| < r. We take the value of m to be 2 and the value of r to be 0.2 × Std.

(3) Median distance from the centroid of phase space reconstruction (M-DCPSR): Phase space reconstruction (PSR) is applied in EEG research to unveil the nonlinear dynamical properties and spatiotemporal relationships of brain electrical activity (Kaur et al., 2020; Lee et al., 2014). We innovatively propose M-DCPSR to unveil the nonlinear characteristics of EEG in the aMCI population.

Assume we have a time-series data set of length N = {x₁, x₂, …, x_N}, the calculation is as follows:

(3.1) Setting the embedding dimension of PSR as m = 3 and determine the lag of the time series (τ) using the autocorrelation function:

$\begin{array}{l}N=N-x_{mean}=\{x_1-x_{mean},x_{2-x_{mean}},\dots ,\\ x_N-x_{mean}\}& (13)\end{array}$

$\begin{array}{ll}SSd=N·N& (14)\end{array}$

$\begin{array}{ll}{R}_{xx}(\tau )=\frac{1}{SSd}{\displaystyle \sum _{i=1}^{N-\tau }}N(i)·N(i+\tau )& (15)\end{array}$

By detecting zero crossings of the autocorrelation function, the time delay τ is identified. The values of the autocorrelation function are iterated through in a loop, and when

$\begin{array}{ll}{R}_{xx}(j-1)·R_{xx}(j)\le 0& (16)\end{array}$

it indicates a zero crossing in the autocorrelation function. Subsequently, the time delay τ is determined based on the absolute values of the autocorrelation function before and after the crossing point:

$\begin{array}{ll}\{\begin{array}{ll}\tau =j,\hfill & \left|R_{xx}(j-1)\right|\le \left|R_{xx}(j)\right|\hfill \\ \tau =j-1,\hfill & \left|R_{xx}(j-1)\right|>\left|R_{xx}(j)\right|\hfill \end{array}& (17)\end{array}$

(3.2) Construct the three-dimensional coordinates of the time series in the phase space based on the τ:

$\begin{array}{ll}\begin{array}{c}\left(x_n,y_n,z_n\right)=\left\{\begin{array}{ccc}{x}_1& x_{1+\tau }& x_{1+2\tau }\\ x_2& x_{2+\tau }& x_{2+2\tau }\\ \begin{array}{c}⋮\\ x_{N-2\tau }\end{array}& \begin{array}{c}⋮\\ x_{N-\tau }\end{array}& \begin{array}{c}⋮\\ x_N\end{array}\end{array}\right\}\end{array}& (18)\end{array}$

each column of matrix represents one dimension of a three-dimensional coordinate.

(3.3) Calculate the centroid of the structure formed by all points in the phase space:

$\begin{array}{ll}\begin{array}{c}\left(\frac{{\displaystyle {\sum }_{i=1}^n}{x}_i}{n},\frac{{\displaystyle {\sum }_{i=1}^n}{y}_i}{n},\frac{{\displaystyle {\sum }_{i=1}^n}{z}_i}{n}\right)\end{array}& (19)\end{array}$

(3.4) Calculate the Euclidean distance between each point and the centroid:

$\begin{array}{ll}{d}_n=\sqrt{{\left(x_n-\frac{{\displaystyle {\sum }_{i=1}^n}{x}_i}{n}\right)}^2+{\left(y_n-\frac{{\displaystyle {\sum }_{i=1}^n}{y}_i}{n}\right)}^2+{\left(z_n-\frac{{\displaystyle {\sum }_{i=1}^n}{z}_i}{n}\right)}^2}& (20)\end{array}$

(3.5) Finally, calculate the median of these Euclidean distances, yielding the M-DCPSR for the given time series.

2.4.1.3 Functional connectivity feature

Phase lag index (PLI), which is used to measure the degree of phase synchronization between two signals, can exclude the influences of volume conduction in EEG signals. It is commonly employed as a functional connectivity feature in MCI and AD patients (Kuang et al., 2022; Nobukawa et al., 2020; Núñez et al., 2019). PLI values range from 0 to 1. A PLI of zero indicates either no coupling or coupling with a phase difference centered around 0 or π. A PLI of 1 indicates perfect phase locking at a value different from 0 or π. Due to the significant alterations observed in the alpha band among the MCI population, this study exclusively extracted the PLI within the alpha band from the patients.

Suppose the band-pass filtered signals at two electrodes are X(t) and Y(t), respectively. Through the Hilbert transform, we can obtain their analytic signals X_an(t) and Y_an(t), as

$\begin{array}{ll}{X}_{an}(t)=X(t)+i{X}_H(t)& (21)\end{array}$

$\begin{array}{ll}{Y}_{an}(t)=Y(t)+i{Y}_H(t)& (22)\end{array}$

where X_H(t) and Y_H(t) are the Hilbert transform of band-pass filtered signals X(t) and Y(t), respectively. Namely,

$\begin{array}{ll}{X}_H(t)=\frac{1}{\pi }P.V.{\displaystyle {\int }_{-\infty }^{+\infty }}\frac{X(\tau )}{t-\tau }d\tau & (23)\end{array}$

$\begin{array}{ll}{Y}_H(t)=\frac{1}{\pi }P.V.{\displaystyle {\int }_{-\infty }^{+\infty }}\frac{Y(\tau )}{t-\tau }d\tau & (24)\end{array}$

where P. V. is the Cauchy principal value.

Using analytical signals, the instantaneous amplitude A_x(t), A_y(t) and instantaneous phase ϕ_x(t), ϕ_y(t), can be calculated:

$\begin{array}{ll}{A}_x(t)=\sqrt{{X_{an}(t)}^2+{X_H(t)}^2}& (25)\end{array}$

$\begin{array}{ll}{A}_y(t)=\sqrt{{Y_{an}(t)}^2+{Y_H(t)}^2}& (26)\end{array}$

$\begin{array}{ll}{\varphi }_x(t)={tan}^{-1}\frac{X_H(t)}{X_{an}(t)}& (27)\end{array}$

$\begin{array}{ll}{\varphi }_y(t)={tan}^{-1}\frac{Y_H(t)}{Y_{an}(t)}& (28)\end{array}$

We calculate Δϕ_xy(t), which is their phase difference at time t:

$\begin{array}{ll}\Delta {\varphi }_{xy}(t)=|{\varphi }_x(t)-{\varphi }_y(t)|& (29)\end{array}$

In actual analysis, the phase difference needs to be converted to [0, 2π):

$\begin{array}{ll}\Delta {\varphi }_{rel}(t)=\Delta {\varphi }_{xy}(t)mod2\pi & (30)\end{array}$

Finally, the formula of PLI is:

$\begin{array}{ll}PLI=\left|\frac{1}{N}{\displaystyle \sum _{N=1}^N}sign(\Delta {\varphi }_{rel}(t))\right|& (31)\end{array}$

2.4.2 Longitudinal feature

We extracted features related to the centralized, dispersion, and dynamic trends of each EEG cross-sectional feature across the initial three measurements, constituting the longitudinal features.

For the t_th time point, a panel dataset X_ij(t) is formed for the j_th independent variable collected from patient i. Assuming there are N patients, each with records for P variables across T periods.

2.4.2.1 Centralized trend

The centralized feature is denoted as:

$\begin{array}{ll}MF(F_{ij})=\frac{{\displaystyle {\sum }_{i=1}^T}{X}_{ij}(t)}{T}& (32)\end{array}$

MF(F_ij) refers to the mean value of the j_th independent variable collected from patient i over the entire period T. This feature reflects the average level of the j_th independent variable collected for individual i throughout the entire period.

2.4.2.2 Dispersion trend

The dispersion features are denoted as:

$\begin{array}{ll}SDF(F_{ij})={\left[\frac{{\displaystyle {\sum }_{t=1}^T}{(X_{ij}(t)-{\overline{X}}_{ij})}^2}{T-1}\right]}^{\frac{1}{2}}& (33)\end{array}$

$\begin{array}{ll}RF\left(F_{ij}\right)=\left(F_{ij}\right)-\left(F_{ij}\right)& (34)\end{array}$

where ${\overline{X}}_{ij}$ represents the MF(F_ij), SDF(F_ij) and RF(F_ij) reflect the dispersion level of the j_th independent variable collected for individual i throughout the entire period.

2.4.2.3 Dynamic trend

The combination of numerical and graphical features reflecting the dynamic trend are denoted as:

$\begin{array}{l}AF\left(F_{ij}\right)=AREA(F_{ij}\left(1\right)-F_{ij}\left(2\right)-F_{ij}\left(3\right)-time1\\ -time3)& (35)\end{array}$

$\begin{array}{ll}Velocity(F_{ij})=\frac{X_{ij}(t+1)-X_{ij}(t)}{\delta t}& (36)\end{array}$

$\begin{array}{ll}MVF(F_{ij})=\frac{{\displaystyle {\sum }_{i=1}^T}Velocit{y}_{ij}(t)}{T}& (37)\end{array}$

$\begin{array}{ll}SDVF(F_{ij})={\left[\frac{{\displaystyle {\sum }_{t=1}^T}{(Velocit{y}_{ij}(t)-{\overline{Velocity}}_{ij})}^2}{T-1}\right]}^{\frac{1}{2}}& (38)\end{array}$

where AF(F_ij) is the area enclosed by the five points F_ij(1), F_ij(2), F_ij(3), time1, and time3 in Figure 1. This area is the sum of the three trapezoidal areas formed by the feature values at adjacent time points. The horizontal axis represents the collection time (months), and the vertical axis represents the magnitude of the original feature values. Velocity(F_ij) refers to the rate of change between two consecutive measurements of the j_th independent variable for patient i at time points t and t – 1. It corresponds to the slopes of the lines connecting F_ij(1) and F_ij(2), F_ij(2) and F_ij(3) in Figure 1. MVF(F_ij) represents the mean of these three slope values, and SDVF(F_ij) represents their standard deviation.

As shown in Table 2, for each cross-sectional feature, we extracted six longitudinal features, resulting in a total of 1968 longitudinal features. The aforementioned extracted longitudinal features were fed into the machine learning classifiers.

Table 2

Table 2. Number of extracted features.

2.4.3 Feature selection

To mitigate the risk of overfitting resulting from the high-dimensional feature space, a filter feature selection method was used. During each training phase, the F-statistic was computed for every feature using only the training data. Features were ranked by descending F-value, and the top 100 were retained for classifier development.

2.5 Statistical analysis

For comparisons of cross-sectional features between the SMCI and PMCI groups at a single time point, the Mann-Whitney U test was applied, as some features did not meet the assumptions required for parametric testing. To further examine whether the trajectories of these features over time differed between the two groups, linear mixed models (LMM) were applied.

In the LMM, group (SMCI vs. PMCI), time (time1, time2, time3, and time4), and their interaction (group × time) were included as fixed effects, while random intercepts for each subject were introduced to account for within-subject correlations due to repeated measurements. This approach enabled direct comparison of longitudinal feature trajectories between groups while addressing the non-independence of repeated measurements. All analyses were conducted using MATLAB R2023a and R 4.3.2, with a significance level of α = 0.05.

2.6 Prediction and validation

We employed eight commonly used machine learning classifiers for one-year outcomes prediction between PMCI and SMCI, including support vector machine (SVM), decision tree (DT), naive bayes (NB), linear discriminant analysis (LDA), AdaBoost (ADA), k-nearest neighbor (KNN), random forest (RF), and logistic regression (LogReg). All the parameters for machine learning models were set to the default settings in MATLAB. To evaluate model performance robustly, a repeated 10 × 5-fold cross-validation (CV) scheme was employed. Finally, we assessed the classification performance of the machine learning model using six metrics: SEN, SPE, PPV, F1-score, ACC, and AUC. The formula for the previously mentioned metrics is as follows:

$\begin{array}{ll}ACC=\frac{TP+TN}{TP+FN+TN+FP}& (39)\end{array}$

$\begin{array}{ll}SEN=\frac{TP}{TP+FN}& (40)\end{array}$

$\begin{array}{ll}SPE=\frac{TN}{FP+TN}& (41)\end{array}$

$\begin{array}{ll}PPV=\frac{TP}{TP+FP}& (42)\end{array}$

$\begin{array}{ll}F1-score=\frac{2TP}{2TP+FP+FN}& (43)\end{array}$

$\begin{array}{ll}AUC=\frac{{\displaystyle {\sum }_{in{s}_i\in positiveclass}}ran{k}_{in{s}_i}-\frac{M\times (M+1)}{2}}{M\times N}& (44)\end{array}$

where M, N are the number of positive sample and negative sample, separately. TP is the number of PMCI cases that are correctly predicted, FN is the number of PMCI cases that are incorrectly predicted as SMCI samples, FP is the number of SMCI cases that are incorrectly predicted as PMC cases, and TN is the number of SMCI samples that are correctly predicted.

3 Results

3.1 Spectral feature

As illustrated in Figure 2, at time1 and time2, there were no significant differences in PSD ratio2 between the PMCI and SMCI groups across the entire brain. However, at time3 and time4, the differences became statistically significant, with higher values observed in the PMCI group compared to the SMCI group. Similar results were obtained for PSD ratio1, ratio3, ratio4, and ratio5, as shown in Supplementary Appendix A.

Figure 2

Figure 2. The topoplot of original value and statistical differences for Power spectral density (PSD) Ratio2 in stable mild cognitive impairment (SMCI) and progressive mild cognitive impairment (PMCI) groups including 0 month (A), 4 months (B), 8 months (C), and 12 months (D). The first and second column of each sub-figure are the topoplots of mean in SMCI and PMCI groups. The third column of each sub-figure is the topoplot of statistical differences in SMCI and PMCI groups.

At time1 and time2, there were no significant differences in the PSDE for delta, theta, alpha, beta, and gamma bands between the PMCI and SMCI groups across the entire brain. However, at time3 and time4, these differences were statistically significant. Specifically, in the delta and theta bands, PMCI exhibited higher PSDE compared to SMCI across the entire brain. In the alpha and beta bands, PMCI showed lower PSDE compared to SMCI across the entire brain. For the gamma band, PMCI showed higher PSDE in the central and parietal regions compared to SMCI, while in all other channels, it was lower than SMCI. The detailed results of PSDE were presented in Supplementary Appendix A.

3.2 Nonlinear feature

As shown in Figure 3, at time1 and time2, there were no significant differences in the M-DCPSR between the PMCI and SMCI groups across the entire brain. However, at time3 and time4, these differences were statistically significant, with PMCI exhibiting lower values than SMCI. The results for SE and PE were consistent with those of M-DCPSR, as detailed in Supplementary Appendix A.

Figure 3

Figure 3. The topoplot of original value and statistical differences for Median distance from the centroid of phase space reconstruction (M-DCPSR) in stable mild cognitive impairment (SMCI) and progressive mild cognitive impairment (PMCI) groups including 0 month (A), 4 months (B), 8 months (C), and 12 months (D). The first and second column of each sub-figure are the topoplots of mean in SMCI and PMCI groups. The third column of each sub-figure is the topoplot of statistical differences in SMCI and PMCI groups.

3.3 Functional connectivity feature

As illustrated in Figure 4, at time1 and time2, there were no significant differences in the PLI between the PMCI and SMCI groups across the entire brain. However, at time3, the differences in connectivity between electrode pairs across the few brain regions were statistically significant. Furthermore, at time4, the differences in connectivity between most electrode pairs across the entire brain were statistically significant, with PMCI exhibiting lower values than SMCI.

Figure 4

Figure 4. The combination chart of original value and statistical differences for Phase lag index (PLI) in stable mild cognitive impairment (SMCI) and progressive mild cognitive impairment (PMCI) groups including 0 month (A), 4 months (B), 8 months (C), and 12 months (D). The first and second column of each sub-figure are the heatmaps of mean in SMCI and PMCI groups. The third column of each sub-figure is the topoplot of connections with statistical differences in SMCI and PMCI groups.

3.4 Composition of selected features

The selection frequency of each feature type was aggregated across all iterations of the repeated 10 × 5-fold CV. The distribution was as follows: SDF (1,197, 23.94%), SDVF (1,033, 20.66%), RF (936, 18.72%), AF (923, 18.46%), and MVF (911, 18.22%). MF features were not selected.

3.5 Prediction performance

Figure 5 and Table 3 present the prediction performance of eight classifiers using the selected 100 longitudinal features under repeated 10 × 5-fold CV. SVM achieved the best overall performance, attaining the highest mean and lowest standard deviation across ACC (94.92%), SEN (90.20%), and F1-score (93.65%). NB exhibited the highest SPE (99.05%) and PPV (98.98%), while RF achieved the highest AUC (96.91%). ADA, DT and KNN performed remarkably well, with all metrics exceeding 80%. Despite LDA and LogReg having inferior performance, their evaluation metrics generally remained around 60–70%.

Figure 5

Figure 5. The radar chart of prediction results using repeated 10 × 5-fold cross-validation.

Table 3

Table 3. The prediction results using repeated 10 × 5-fold CV (mean ± standard deviation, %).

3.6 Performance comparison

To spotlight the advantages of longitudinal features, we combined information from the initial three cross-sections to explore the necessity of repeated EEG measurements. The prediction performance for different cross-sectional combinations was presented in Supplementary Appendix B. The results demonstrated that the optimal classification performance was achieved when utilizing information from all the initial three cross-sections, followed by two cross-sections. This highlighted the enhanced model performance associated with constructing longitudinal features based on repeated EEG measurements.

4 Discussion

Conducting repeated EEG measurements during the follow-up period, this study comprehensively extracted cross-sectional features including spectral, nonlinear, and functional connectivity. By constructing longitudinal features that reflect dynamic trends and integrating machine learning methods, we established a prediction framework for both SMCI and PMCI populations, demonstrating robust predictive performance in internal validation.

The cross-sectional EEG features extracted in this study have been validated for their excellent discriminative ability across three groups: AD, MCI, and HC. We further focused our research on aMCI individuals, providing supporting evidence through follow-up studies. In this study, both groups exhibited an increasing trend in PSD ratios 1–5, as well as elevated PSDE values in the delta and theta bands. Conversely, a declining pattern was observed in PSDE values within the alpha and beta bands. These results substantiate a transition in EEG spectra from higher to lower frequencies in the MCI population, consistent with existing research conclusions (Ding et al., 2022; Kwak, 2006; Luckhaus et al., 2008; Prichep et al., 2006; Rossini et al., 2006; Yan et al., 2021). We innovatively proposed the non-linear metric M-DCPSR. Both groups exhibit varying degrees of decreasing trends in M-DCPSR, SE, and PE, confirming the reduced complexity of EEG signals in the MCI population from a follow-up study perspective (Deng et al., 2017; Mammone et al., 2018; Tsai et al., 2012; Wang et al., 2019). Additionally, both groups exhibit distinctive decreasing patterns in the PLI with the alpha band, indicating diverse progression in intra- and inter-regional disconnect phenomena, consistent with relevant research findings (Núñez et al., 2019; Ruiz-Gómez et al., 2019). The aforementioned three phenomena may be attributed to the loss of neurons, the altered anatomical structure of neuronal tracts, and the modified release of neurotransmitters (Delbeuck et al., 2003; Olsson et al., 2016; Uhlhaas and Singer, 2006).

An interesting finding of our research is that, at time1 and time2, there were no significant differences in most EEG features between the two groups. However, at time3 and time4, significant differences were observed in the majority of EEG features, indicating that the changing trends of the extracted EEG features in the two groups were inconsistent. This observation is consistent with the high-risk nature of our cohort, as all participants were recruited from the Neurology Department of the First Affiliated Hospital of Sun Yat-sen University, a regional referral center for neurological diseases in South China, where patients are predominantly referred from primary hospitals and represent a high-risk aMCI population with an increased likelihood of short-term progression. These time-dependent differences support the necessity of conducting repeated EEG measurements during follow-up and constructing longitudinal features.

(Mubeen et al. 2017) and (Ardekani et al. 2017) utilized two follow-up MRI measurements from MCI patients and constructed longitudinal features, including the rate of change. Their studies demonstrated an improvement in classification performance when employing longitudinal features compared to cross-sectional features. Our study effectively demonstrates the crucial role of longitudinal features in predicting the disease progression of patients for the AD continuum. Not only does this study extract statistical indicators such as mean, standard deviation, and range, reflecting central tendency and dispersion, but it also utilizes the concept of “combining numerical and graphical information” to extract longitudinal indicators such as the area under the feature curve and the mean and standard deviation of the velocity, further enhancing the classification performance. Feature selection revealed that metrics capturing variability (SDF and SDVF) and change magnitude (RF, AF, and MVF) predominated, whereas mean (MF) features were entirely excluded—suggesting that EEG signal instability and dynamic evolution are more informative than static averages for short-term progression prediction. Due to the gradual yet dynamic progression of MCI to AD, incorporating longitudinal features enhances the performance compared to relying solely on cross-sectional features, further emphasizing the significance of longitudinal studies (Miraglia et al., 2020; Vecchio et al., 2018). The differing trajectories of EEG features among the MCI population confirm the heterogeneity of EEG signals in this group. Our study design can be extended to other age-related diseases or neurological disorders, as well as EEG signal analysis, and various medical research domains.

Incorporating the selected longitudinal features into machine learning models, the SVM classifier demonstrated the best overall prediction performance. This aligns with its strength in modeling complex, nonlinear relationships even with limited samples (Brereton and Lloyd, 2010; Meng and Zhao, 2015), a characteristic pertinent to the dynamic nature of longitudinal EEG data. NB excelled specifically in specificity and positive predictive value, suggesting that certain EEG-derived features exhibit strong class-conditional independence beneficial for ruling out progression (Raschka, 2014). RF achieved the highest AUC, its robust performance is consistent with its known capability to handle high-dimensional data and mitigate overfitting through ensemble learning (Capitaine et al., 2021; Wang et al., 2018). In contrast, linear classifiers including LDA and LogReg showed the lowest discrimination, reflecting their inherent limitation in capturing nonlinear patterns in EEG data (Günther et al., 2012; Wagner et al., 2021; Zhang et al., 2013). The fact that all nonlinear classifiers (SVM, RF, NB, ADA, DT, and KNN) achieved metrics exceeding 80% underscores the general utility and discriminative value of the constructed longitudinal EEG features for prediction within the aMCI continuum.

From the perspective of early prediction, this study adopted a prospective cohort design, constructed EEG longitudinal features, and combined machine learning methods to establish a prediction framework for both SMCI and PMCI populations, achieving satisfactory performance. However, several limitations should be acknowledged. First, the sample size was relatively small, and all participants were recruited from a single tertiary center, which may affect the generalizability of our findings. Second, this study did not employ the Clinical Dementia Rating (CDR) scale, which is a well-established and widely used tool for staging functional impairment in the Alzheimer's disease continuum (Angioni et al., 2024). Third, the follow-up duration was one year, which is a relatively short interval within the prolonged course of AD. While a high conversion rate was observed in this enriched cohort, longer follow-up is needed to evaluate the stability of the predictive framework over time. Fourth, the performance metrics reported here are derived from internal repeated cross-validation. The absence of an independent external test set may lead to optimistic estimates of real-world generalizability (Kikuchi et al., 2022).

To address these limitations, we are expanding the cohort through a multi-center collaboration, establishing additional cohorts at Sun Yat-sen Memorial Hospital and the Third Affiliated Hospital of Sun Yat-sen University to form an independent validation dataset. We also plan to extend the follow-up period to further evaluate the long-term predictive stability of the framework. Future work will additionally focus on enhancing interpretability and the integration of multimodal biomarkers to enhance clinical applicability.

5 Conclusion

Aiming to facilitate the dynamic tracking of disease progression in aMCI patients, we introduced a prediction framework utilizing EEG longitudinal features and machine learning methods in SMCI and PMCI populations. The framework not only comprehensively extracted spectral, nonlinear, and functional connectivity features from EEG data, but also further constructed longitudinal features reflecting dynamic trends. The SVM classifier achieved an accuracy of 94.92%, with most classifiers demonstrating satisfactory classification performance, indicating the stability of this framework. Our study captured the dynamic changes in EEG features among the aMCI population, providing robust follow-up evidence for the trajectory of aMCI individuals and offering EEG biomarkers with predictive value for AD.

Data availability statement

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

Ethics statement

The studies involving humans were approved by Ethics Committee of the School of Public Health, Sun Yat-sen University (2021-No.081). The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

Author contributions

YG: Conceptualization, Writing – original draft. YFe: Writing – original draft, Investigation. CD: Writing – original draft, Visualization. SY: Validation, Writing – original draft. YFa: Resources, Writing – original draft. YZ: Resources, Writing – original draft. JY: Software, Writing – original draft. QP: Investigation, Writing – original draft. NZ: Investigation, Writing – original draft. XZ: Writing – original draft, Visualization. XL: Writing – review & editing. JZ: Writing – review & editing, Funding acquisition.

Funding

The author(s) declared that financial support was received for this work and/or its publication. This research was funded by the Basic and Applied Basic Research Foundation of Guangdong Province, China, Grant Numbers 2022A1515011237 and 2023A1515011951.

Acknowledgments

The authors would like to thank Zichao Liang, Siyang Chen, and Yuduan Han for helpful suggestions.

Conflict of interest

The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fnagi.2025.1719981/full#supplementary-material

References

Alam, R. U., Zhao, H., Goodwin, A., Kavehei, O., and McEwan, A. (2020). Differences in power spectral densities and phase quantities due to processing of EEG signals. Sensors. 20:6285. doi: 10.3390/s20216285

PubMed Abstract | Crossref Full Text | Google Scholar

Alzheimer's disease facts and figures (2023). Alzheimers Dement. 19, 1598–1695. doi: 10.1002/alz.13016

Crossref Full Text | Google Scholar

Angioni, D., Cummings, J., Lansdall, C. J., Middleton, L., Sampaio, C., Gauthier, S., et al. (2024). Clinical meaningfulness in Alzheimer's disease clinical trials. a report from the EU-US CTAD task force. J. Prev. Alzheimers Dis. 11, 1219–1227. doi: 10.14283/jpad.2024.112

PubMed Abstract | Crossref Full Text | Google Scholar

Ardekani, B. A., Bermudez, E., Mubeen, A. M., and Bachman, A. H. (2017). Prediction of incipient Alzheimer's disease dementia in patients with mild cognitive impairment. J. Alzheimers. Dis. 55, 269–281. doi: 10.3233/JAD-160594

PubMed Abstract | Crossref Full Text | Google Scholar

Bandt, C., and Pompe, B. (2002). Permutation entropy: a natural complexity measure for time series. Phys. Rev. Lett. 88:174102. doi: 10.1103/PhysRevLett.88.174102

PubMed Abstract | Crossref Full Text | Google Scholar

Brereton, R. G., and Lloyd, G. R. (2010). Support vector machines for classification and regression. Analyst 135, 230–267. doi: 10.1039/B918972F

PubMed Abstract | Crossref Full Text | Google Scholar

Cai, C., Huang, C., Yang, C., Lu, H., Hong, X., Ren, F., et al. (2020). Altered patterns of functional connectivity and causal connectivity in salience subnetwork of subjective cognitive decline and amnestic mild cognitive impairment. Front. Neurosci. 14:288. doi: 10.3389/fnins.2020.00288

PubMed Abstract | Crossref Full Text | Google Scholar

Capitaine, L., Genuer, R., and Thiébaut, R. (2021). Random forests for high-dimensional longitudinal data. Stat. Methods Med. Res. 30, 166–184. doi: 10.1177/0962280220946080

PubMed Abstract | Crossref Full Text | Google Scholar

Cascarano, A., Mur-Petit, J., Hernández-González, J., Camacho, M., de Toro Eadie, N., Gkontra, P., et al. (2023). Machine and deep learning for longitudinal biomedical data: a review of methods and applications. Artif. Intell. Rev. 56, 1711–1771. doi: 10.1007/s10462-023-10561-w

Crossref Full Text | Google Scholar

Delbeuck, X., Van der Linden, M., and Collette, F. (2003). Alzheimer's disease as a disconnection syndrome? Neuropsychol. Rev. 13, 79–92. doi: 10.1023/A:1023832305702

PubMed Abstract | Crossref Full Text | Google Scholar

Delorme, A., and Makeig, S. (2004). EEGLAB: an open source toolbox for analysis of single-trial EEG dynamics including independent component analysis. J. Neurosci. Methods 134, 9–21. doi: 10.1016/j.jneumeth.2003.10.009

PubMed Abstract | Crossref Full Text | Google Scholar

Deng, B., Cai, L., Li, S., Wang, R., Yu, H., Chen, Y., et al. (2017). Multivariate multi-scale weighted permutation entropy analysis of EEG complexity for Alzheimer's disease. Cogn. Neurodyn. 11, 217–231. doi: 10.1007/s11571-016-9418-9

PubMed Abstract | Crossref Full Text | Google Scholar

Ding, Y., Chu, Y., Liu, M., Ling, Z., Wang, S., Li, X., et al. (2022). Fully automated discrimination of Alzheimer's disease using resting-state electroencephalography signals. Quant. Imaging Med. Surg. 12, 1063–1078. doi: 10.21037/qims-21-430

PubMed Abstract | Crossref Full Text | Google Scholar

Gómez-Soria, I., Peralta-Marrupe, P., Calatayud-Sanz, E., and Latorre, E. (2021). Efficacy of cognitive intervention programs in amnesic mild cognitive impairment: a systematic review. Arch. Gerontol. Geriatr. 94:104332. doi: 10.1016/j.archger.2020.104332

PubMed Abstract | Crossref Full Text | Google Scholar

Günther, F., Pigeot, I., and Bammann, K. (2012). Artificial neural networks modeling gene-environment interaction. BMC Genet. 13:37. doi: 10.1186/1471-2156-13-37

PubMed Abstract | Crossref Full Text | Google Scholar

Ieracitano, C., Mammone, N., Hussain, A., and Morabito, F. C. (2020). A novel multi-modal machine learning based approach for automatic classification of EEG recordings in dementia. Neural Netw. 123, 176–190. doi: 10.1016/j.neunet.2019.12.006

PubMed Abstract | Crossref Full Text | Google Scholar

Kaur, S., Singh, S., Arun, P., Kaur, D., and Bajaj, M. (2020). Phase space reconstruction of EEG signals for classification of ADHD and control adults. Clin. EEG Neurosci. 51, 102–113. doi: 10.1177/1550059419876525

PubMed Abstract | Crossref Full Text | Google Scholar

Kikuchi, M., Kobayashi, K., Itoh, S., Kasuga, K., Miyashita, A., Ikeuchi, T., et al. (2022). Identification of mild cognitive impairment subtypes predicting conversion to Alzheimer's disease using multimodal data. Comput. Struct. Biotechnol. J. 20, 5296–5308. doi: 10.1016/j.csbj.2022.08.007

PubMed Abstract | Crossref Full Text | Google Scholar

Kim, J. G., Kim, H., Hwang, J., Kang, S. H., Lee, C. N., Woo, J., et al. (2022). Differentiating amnestic from non-amnestic mild cognitive impairment subtypes using graph theoretical measures of electroencephalography. Sci. Rep. 12:6219. doi: 10.1038/s41598-022-10322-9

PubMed Abstract | Crossref Full Text | Google Scholar

Kuang, Y., Wu, Z., Xia, R., Li, X., Liu, J., Dai, Y., et al. (2022). Phase lag index of resting-state EEG for identification of mild cognitive impairment patients with type 2 diabetes. Brain Sci. 12:1399. doi: 10.3390/brainsci12101399

PubMed Abstract | Crossref Full Text | Google Scholar

Kwak, Y. T. (2006). Quantitative EEG findings in different stages of Alzheimer's disease. J. Clin. Neurophysiol. 23, 456–461. doi: 10.1097/01.wnp.0000223453.47663.63

PubMed Abstract | Crossref Full Text | Google Scholar

Lee, K., Choi, K. M., Park, S., Lee, S. H., and Im, C. H. (2022). Selection of the optimal channel configuration for implementing wearable EEG devices for the diagnosis of mild cognitive impairment. Alzheimers. Res. Ther. 14:170. doi: 10.1186/s13195-022-01115-3

PubMed Abstract | Crossref Full Text | Google Scholar

Lee, S. H., Lim, J. S., Kim, J. K., Yang, J., and Lee, Y. (2014). Classification of normal and epileptic seizure EEG signals using wavelet transform, phase-space reconstruction, and Euclidean distance. Comput. Methods Programs Biomed. 116, 10–25. doi: 10.1016/j.cmpb.2014.04.012

PubMed Abstract | Crossref Full Text | Google Scholar

Li, H. B., Li, J. Y., Yang, G. Y., Liu, Y., Frost, F., and Hong, Y. L. (2023). Can one series of self-organized nanoripples guide another series of self-organized nanoripples during ion bombardment: from the perspective of power spectral density entropy? Entropy. 25:170. doi: 10.3390/e25010170

PubMed Abstract | Crossref Full Text | Google Scholar

Li, X., Liu, Y., Kang, J., Sun, Y., Xu, Y., Yuan, Y., et al. (2022). Identifying amnestic mild cognitive impairment with convolutional neural network adapted to the spectral entropy heat map of the electroencephalogram. Front. Hum. Neurosci. 16:924222. doi: 10.3389/fnhum.2022.924222

PubMed Abstract | Crossref Full Text | Google Scholar

Li, X., Yang, C., Xie, P., Han, Y., Su, R., Li, Z., et al. (2021). The diagnosis of amnestic mild cognitive impairment by combining the characteristics of brain functional network and support vector machine classifier. J. Neurosci. Methods 363:109334. doi: 10.1016/j.jneumeth.2021.109334

PubMed Abstract | Crossref Full Text | Google Scholar

Libon, D. J., Xie, S. X., Eppig, J., Wicas, G., Lamar, M., Lippa, C., et al. (2010). The heterogeneity of mild cognitive impairment: a neuropsychological analysis. J. Int. Neuropsychol. Soc. 16, 84–93. doi: 10.1017/S1355617709990993

PubMed Abstract | Crossref Full Text | Google Scholar

Luckhaus, C., Grass-Kapanke, B., Blaeser, I., Ihl, R., Supprian, T., Winterer, G., et al. (2008). Quantitative EEG in progressing vs stable mild cognitive impairment (MCI): results of a 1-year follow-up study. Int. J. Geriatr. Psychiatry 23, 1148–1155. doi: 10.1002/gps.2042

PubMed Abstract | Crossref Full Text | Google Scholar

Mammone, N., Ieracitano, C., Adeli, H., Bramanti, A., and Morabito, F. C. (2018). Permutation Jaccard distance-based hierarchical clustering to estimate EEG network density modifications in MCI subjects. IEEE Trans. Neural. Netw. Learn. Syst. 29. doi: 10.1109/TNNLS.2018.2791644

PubMed Abstract | Crossref Full Text | Google Scholar

Markesbery, W. R. (2010). Neuropathologic alterations in mild cognitive impairment: a review. J. Alzheimers. Dis. 19, 221–228. doi: 10.3233/JAD-2010-1220

PubMed Abstract | Crossref Full Text | Google Scholar

McKhann, G. M., Knopman, D. S., Chertkow, H., Hyman, B. T., Jack, C. R. Jr., Kawas, C. H., et al. (2011). 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. 7, 263–269. doi: 10.1016/j.jalz.2011.03.005

PubMed Abstract | Crossref Full Text | Google Scholar

Meng, M., and Zhao, C. (2015). Application of support vector machines to a small-sample prediction. Adv. Petrol. Explor. Dev. 10, 72–75. doi: 10.3968/7830

Crossref Full Text | Google Scholar

Miraglia, F., Vecchio, F., Marra, C., Quaranta, D., Alù, F., Peroni, B., et al. (2020). Small world index in default mode network predicts progression from mild cognitive impairment to dementia. Int. J. Neural Syst. 30:2050004. doi: 10.1142/S0129065720500045

PubMed Abstract | Crossref Full Text | Google Scholar

Monllor, P., Cervera-Ferri, A., Lloret, M. A., Esteve, D., Lopez, B., Leon, J. L., et al. (2021). Electroencephalography as a non-invasive biomarker of alzheimer's disease: a forgotten candidate to substitute CSF molecules? Int. J. Mol. Sci. 22:10889. doi: 10.3390/ijms221910889

PubMed Abstract | Crossref Full Text | Google Scholar

Movahed, R. A., and Rezaeian, M. (2022). Automatic diagnosis of mild cognitive impairment based on spectral, functional connectivity, and nonlinear EEG-based features. Comput. Math. Methods Med. 2022:2014001. doi: 10.1155/2022/2014001

PubMed Abstract | Crossref Full Text | Google Scholar

Mubeen, A. M., Asaei, A., Bachman, A. H., Sidtis, J. J., and Ardekani, B. A. (2017). A six-month longitudinal evaluation significantly improves accuracy of predicting incipient Alzheimer's disease in mild cognitive impairment. J. Neuroradiol. 44, 381–387. doi: 10.1016/j.neurad.2017.05.008

PubMed Abstract | Crossref Full Text | Google Scholar

Nobukawa, S., Yamanishi, T., Kasakawa, S., Nishimura, H., Kikuchi, M., and Takahashi, T. (2020). Classification methods based on complexity and synchronization of electroencephalography signals in Alzheimer's disease. Front. Psychiatry 11:255. doi: 10.3389/fpsyt.2020.00255

PubMed Abstract | Crossref Full Text | Google Scholar

Núñez, P., Poza, J., Gómez, C., Rodríguez-González, V., Hillebrand, A., Tola-Arribas, M. A., et al. (2019). Characterizing the fluctuations of dynamic resting-state electrophysiological functional connectivity: reduced neuronal coupling variability in mild cognitive impairment and dementia due to Alzheimer's disease. J. Neural Eng. 16:056030. doi: 10.1088/1741-2552/ab234b

PubMed Abstract | Crossref Full Text | Google Scholar

Olsson, B., Lautner, R., Andreasson, U., Öhrfelt, A., Portelius, E., Bjerke, M., et al. (2016). CSF and blood biomarkers for the diagnosis of Alzheimer's disease: a systematic review and meta-analysis. Lancet Neurol. 15, 673–684. doi: 10.1016/S1474-4422(16)00070-3

PubMed Abstract | Crossref Full Text | Google Scholar

Ossenkoppele, R., Pijnenburg, Y. A., Perry, D. C., Cohn-Sheehy, B. I., Scheltens, N. M., Vogel, J. W., et al. (2015). The behavioural/dysexecutive variant of Alzheimer's disease: clinical, neuroimaging and pathological features. Brain 138(Pt 9), 2732–2749. doi: 10.1093/brain/awv191

PubMed Abstract | Crossref Full Text | Google Scholar

Perez-Valero, E., Lopez-Gordo, M., Gutiérrez, C. M., Carrera-Muñoz, I., and Vílchez-Carrillo, R. M. (2022). A self-driven approach for multi-class discrimination in Alzheimer's disease based on wearable EEG. Comput. Methods Programs Biomed. 220:106841. doi: 10.1016/j.cmpb.2022.106841

PubMed Abstract | Crossref Full Text | Google Scholar

Petersen, R. C. (2004). Mild cognitive impairment as a diagnostic entity. J. Intern. Med. 256, 183–194. doi: 10.1111/j.1365-2796.2004.01388.x

PubMed Abstract | Crossref Full Text | Google Scholar

Prichep, L. S., John, E. R., Ferris, S. H., Rausch, L., Fang, Z., Cancro, R., et al. (2006). Prediction of longitudinal cognitive decline in normal elderly with subjective complaints using electrophysiological imaging. Neurobiol. Aging 27, 471–481. doi: 10.1016/j.neurobiolaging.2005.07.021

PubMed Abstract | Crossref Full Text | Google Scholar

Raschka, S. (2014). Naive bayes and text classification I—introduction and theory. arXiv:1410.5329. Available online at: https://ui.adsabs.harvard.edu/abs/2014arXiv1410.5329R (accessed October 01, 2014)

Google Scholar

Richman, J. S., and Moorman, J. R. (2000). Physiological time-series analysis using approximate entropy and sample entropy. Am. J. Physiol. Heart Circul. Physiol. 278, H2039–H2049. doi: 10.1152/ajpheart.2000.278.6.H2039

PubMed Abstract | Crossref Full Text | Google Scholar

Rossini, P. M., Del Percio, C., Pasqualetti, P., Cassetta, E., Binetti, G., Dal Forno, G., et al. (2006). Conversion from mild cognitive impairment to Alzheimer's disease is predicted by sources and coherence of brain electroencephalography rhythms. Neuroscience 143, 793–803. doi: 10.1016/j.neuroscience.2006.08.049

PubMed Abstract | Crossref Full Text | Google Scholar

Rossini, P. M., Miraglia, F., and Vecchio, F. (2022). Early dementia diagnosis, MCI-to-dementia risk prediction, and the role of machine learning methods for feature extraction from integrated biomarkers, in particular for EEG signal analysis. Alzheimers. Dement. 18, 2699–2706. doi: 10.1002/alz.12645

PubMed Abstract | Crossref Full Text | Google Scholar

Ruiz-Gómez, S. J., Gómez, C., Poza, J., Gutiérrez-Tobal, G. C., Tola-Arribas, M. A., Cano, M., et al. (2018). Automated multiclass classification of spontaneous EEG activity in Alzheimer's disease and mild cognitive impairment. Entropy 20:35. doi: 10.3390/e20010035

PubMed Abstract | Crossref Full Text | Google Scholar

Ruiz-Gómez, S. J., Hornero, R., Poza, J., Maturana-Candelas, A., Pinto, N., and Gómez, C. (2019). Computational modeling of the effects of EEG volume conduction on functional connectivity metrics. Application to Alzheimer's disease continuum. J. Neural Eng. 16:066019. doi: 10.1088/1741-2552/ab4024

PubMed Abstract | Crossref Full Text | Google Scholar

Seker, M., Özbek, Y., Yener, G., and Özerdem, M. S. (2021). Complexity of EEG dynamics for early diagnosis of Alzheimer's disease using permutation entropy neuromarker. Comput. Methods Programs Biomed. 206:106116. doi: 10.1016/j.cmpb.2021.106116

PubMed Abstract | Crossref Full Text | Google Scholar

Siuly, S., Alcin, O. F., Kabir, E., Sengur, A., Wang, H., Zhang, Y., et al. (2020). A new framework for automatic detection of patients with mild cognitive impairment using resting-state EEG signals. IEEE Trans. Neural Syst. Rehabil. Eng. 28, 1966–1976. doi: 10.1109/TNSRE.2020.3013429

PubMed Abstract | Crossref Full Text | Google Scholar

Tsai, P. H., Lin, C., Tsao, J., Lin, P. F., Wang, P. C., Huang, N. E., et al. (2012). Empirical mode decomposition based detrended sample entropy in electroencephalography for Alzheimer's disease. J. Neurosci. Methods 210, 230–237. doi: 10.1016/j.jneumeth.2012.07.002

PubMed Abstract | Crossref Full Text | Google Scholar

Uhlhaas, P. J., and Singer, W. (2006). Neural synchrony in brain disorders: relevance for cognitive dysfunctions and pathophysiology. Neuron 52, 155–168. doi: 10.1016/j.neuron.2006.09.020

PubMed Abstract | Crossref Full Text | Google Scholar

Vecchio, F., Miraglia, F., Iberite, F., Lacidogna, G., Guglielmi, V., Marra, C., et al. (2018). Sustainable method for Alzheimer dementia prediction in mild cognitive impairment: Electroencephalographic connectivity and graph theory combined with apolipoprotein E. Ann. Neurol. 84, 302–314. doi: 10.1002/ana.25289

PubMed Abstract | Crossref Full Text | Google Scholar

Wagner, D., Heider, D., and Hattab, G. (2021). Mushroom data creation, curation, and simulation to support classification tasks. Sci. Rep. 11:8134. doi: 10.1038/s41598-021-87602-3

PubMed Abstract | Crossref Full Text | Google Scholar

Wang, Q., Nguyen, T.-T., Huang, J. Z., and Nguyen, T. T. (2018). An efficient random forests algorithm for high dimensional data classification. Adv. Data Anal. Classif. 12, 953–972. doi: 10.1007/s11634-018-0318-1

Crossref Full Text | Google Scholar

Wang, X. W., Zhao, X. H., Li, F., Lin, Q., and Hu, Z. H. (2019). Sample entropy and surrogate data analysis for Alzheimer's disease. Math. Biosci. Eng. 16, 6892–6906. doi: 10.3934/mbe.2019345

PubMed Abstract | Crossref Full Text | Google Scholar

Weidner, W., and Barbarino, P. (2019). The state of the art of dementia research: new frontiers. Alzheimer's Dement. 15:4115. doi: 10.1016/j.jalz.2019.06.4115

Crossref Full Text | Google Scholar

Winblad, B., Amouyel, P., Andrieu, S., Ballard, C., Brayne, C., Brodaty, H., et al. (2016). Defeating Alzheimer's disease and other dementias: a priority for European science and society. Lancet Neurol. 15, 455–532. doi: 10.1016/S1474-4422(16)00062-4

PubMed Abstract | Crossref Full Text | Google Scholar

Yan, Y., Zhao, A., Ying, W., Qiu, Y., Ding, Y., Wang, Y., et al. (2021). Functional connectivity alterations based on the weighted phase lag index: an exploratory electroencephalography study on Alzheimer's disease. Curr. Alzheimer Res. 18, 513–522. doi: 10.2174/1567205018666211001110824

PubMed Abstract | Crossref Full Text | Google Scholar

Yang, S., Bornot, J. M. S., Wong-Lin, K., and Prasad, G. (2019). M/EEG-based bio-markers to predict the MCI and Alzheimer's disease: a review from the ML perspective. IEEE Trans. Biomed. Eng. 66, 2924–2935. doi: 10.1109/TBME.2019.2898871

PubMed Abstract | Crossref Full Text | Google Scholar

Zhang, R., Xu, P., Guo, L., Zhang, Y., Li, P., and Yao, D. (2013). Z-score linear discriminant analysis for EEG based brain-computer interfaces. PLoS ONE 8:e74433. doi: 10.1371/journal.pone.0074433

PubMed Abstract | Crossref Full Text | Google Scholar