A Cross-Device Ensemble Learning Model for Continuous Cognition Assessment Based on Resting-State EEG

doi:10.21203/rs.3.rs-3119061/v1

Download PDF

Article

A Cross-Device Ensemble Learning Model for Continuous Cognition Assessment Based on Resting-State EEG

https://doi.org/10.21203/rs.3.rs-3119061/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 02 Jan, 2024

Read the published version in npj Aging →

Version 1

posted

You are reading this latest preprint version

One critical manifestation of neurological deterioration is the sign of cognitive decline. Causes of cognitive decline include but are not limited to: aging, cerebrovascular disease, Alzheimer's disease, and trauma. Currently, the primary tool used to examine cognitive decline is scale. However, scale examination has drawbacks such as its clinician subjectivity and inconsistent results. This study attempted to use resting-state EEG to construct a cognitive assessment model that is capable of providing a more scientific and robust evaluation on cognition levels. In this study, 75 healthy subjects, 99 patients with Mild Cognitive Impairment (MCI), and 78 patients with dementia were involved. Their resting-state EEG signlas were collected twice, and the recording devices varied. By matching these EEG and traditional scale results, the proposed cognition assessment model was trained based on Adaptive Boosting (AdaBoost) and Support Vector Machines (SVM) methods, mapping subjects' cognitive levels to a 0-100 test score with a mean error of 4.82 (< 5%). This study is the first to establish a continuous evaluation model of cognitive decline on a large sample dataset. Its cross-device usability also suggests universality and robustness of this EEG model, offering a more reliable and affordable way to assess cognitive decline for clinical diagnosis and treatment as well. Furthermore, the interpretability of features involved may further contribute to the early diagnosis and superior treatment evaluation of Alzheimer's disease.

Health sciences/Health care/Diagnosis

Biological sciences/Neuroscience/Cognitive ageing

Cognitive decline is an outward manifestation of neurological decline caused by factors other than natural aging (Petersen et al., 1999). The causes of cognitive decline include but are not limited to, many neurological disorders such as Alzheimer's disease, epilepsy, cerebrovascular disease, tumors, other injuries or illnesses such as trauma, diabetes, HIV, thyroid abnormalities, etc. (Petersen, 2004). According to WHO estimates, people suffering from cognitive decline will reach 152 million worldwide by 2050.

The main symptoms of cognitive decline are 1. memory loss. 2. repetitive questioning. 3. abstract logic impairment. 4. language impairment. 5. loss of hobbies and interests. 6. emotional deficits and moodiness. 7. reduced judgment. 8. spatial orientation impairment. 9. misplaced objects. 10. comprehension impairment (Charles & Carstensen, 2010; Gauthier et al., 2006; Rao et al., 1991; West, 1996). Cognitive decline places a huge economic burden on society due to its overall impairment of a person's ability to function autonomously. Early detection of cognitive decline has become the key to its treatment as many studies have shown that early intervention is relatively effective.

The current diagnosis of cognitive decline relies heavily on scales (Crum et al., 1993; Thapa et al., 2020). Alzheimer's disease, a primary neurodegenerative disease, is one of the leading causes of cognitive decline and even dementia (Albert et al., 2011). A more accurate and scientific diagnosis requires PET imaging and tests for abnormal protein deposits in the cerebrospinal fluid, which can be costly. Therefore, in the face of many people with cognitive decline, there is an urgent need for a convenient and stable assessment method for early screening and developmental assessment of cognitive decline.

EEG has many applications in cognitive decline research as a high temporal resolution neural signal that can be easily acquired (Brown et al., 2012; Jeong, 2004; Palop & Mucke, 2009). Various temporal, spatial, and frequency-based EEG computational features are currently available to evaluate cognitive decline at different levels (D'Rozario et al., 2020; Jafari et al., 2020). Although there is a solid research base in the direction of cognitive segmentation, there are currently no continuous macroscopic models for evaluating cognitive decline based on extensive data sets. Considering that different features contain different neural activity information, the present study will combine three dimensions of temporal, spatial, and frequency EEG computational features using the AdaBoost method. The study uses scale results to constrain the evaluation model so that the model can give continuous scores.

The study analyzed the cognitive scale test results for the three subject groups (Table 1 and Fig. 2) and discovered that the two-scale tests indicated a decreasing trend in scores with decreasing cognitive capabilities. The mean MoCAB scores for the HC, MCI, and Dementia groups were 24.39, 18.01, and 10.68, respectively, while the MMSE scores were 27.47, 25.18, and 17.86, respectively. It can be seen that as the level of cognition decreases, the range of distribution of the scale test results increases. This suggests that even among those with a clinical diagnosis of dementia, a few subjects have high-scale test scores, which may be associated with high levels of schooling, difficult work, etc. As a result, depending exclusively on scale results carries some risk.

Regarding the frequency domain dimension, cognitive decline implies an overall slowing of activity. A decrease in high-frequency activity and an increase in low-frequency activity are typical of the frequency domain (Fig. 3). The change from the HC to MCI stage is relatively moderate. In contrast, the change from MCI to dementia is more pronounced. In terms of the spatial distribution of power, the temporoparietal lobe of HC has relatively strong activity in the resting state, while MCI and dementia have weaker activity. Overall activity in HC is highly homogeneous, while frontal lobe activity is significantly reduced in dementia.

Observed in the spatial dimension, the decrease in overall information transfer brought about by the cognitive decline implies a silent and inactive network (Fig. 4). The results point to a higher outflow of information from the central regions of the network (C1, C2, C3, C4, CP1, CP2) in the resting state, which may imply central control and moderation. This phenomenon diminishes with decreasing cognition, and the average information transfer across the network decreases. In addition, it was observed that cognitive decline did not imply a monotonic network degeneration and that enhanced information between specific regions was observed from the HC to the MCI stage, e.g., a slight increase in outflow information in C5, C6, FP1, FP2 compared to the HC network (4.93% increase on average)

Observed from the perspective of spatial networks, cognitive decline reduces network connectivity. This section calculates the sum of incoming and outgoing information between any two nodes as the strength of the connection between these two nodes. The research filters the network's major connections according to a criterion of greater than 80%. It presents them normalized (Fig. 5). The results indicate that the number of major connections in the network decreases with cognitive decline and follows the principle of retaining strong connections and reducing weak ones.

In the time dimension, cognitive impairment appears to have destabilized network activity (Fig. 6). This implies shorter state durations and more significant gaps between states. In the resting state, the average brain is in a dynamic and stable process that involves constant switching between many modes. The findings, on the other hand, suggest that dementia patients are missing states 12 and 15 (Fig. S1), which could be related to structural damage produced by brain atrophy and neuronal death. From a translational standpoint, cognitive deterioration decreased state duration, implying more translations and systemic dissociation. When the average transformation distance results are combined, it is known that brains with cognitive impairment appear to be more prone to state leaps rather than continuous and smooth evolution. The temporal representation results hint at cognitive deterioration, rendering the conscious experience discrete, which may also be associated with impaired central control and a lack of attention.

In this work, many features were extracted from the EEG of subjects with different cognitive levels in three dimensions: time domain, space domain, and frequency domain. Features are composed of eight weak classifiers according to the feature types, and the classification effect of each feature individually is shown in Table 1. The results point to a relatively good classification performance for the micro-state phase pipe features and a classification accuracy of only 0.64. This indicates that the amount of information carried by a single-dimensional feature is insufficient for the triple classification task of HC, MCI, and dementia.

Table 1

Classified expressions of different characteristics
Weak classifier	Single classifier Accuracy (baseline)	Final weights	The final number of features
Fractal characteristics	0.44 (0.33)	0.02624	34
Complexity measures	0.56 (0.33)	0.14783	26
Microstate feature	0.67 (0.33)	0.31355	125
Spectral characteristics	0.51 (0.33)	0.11679	113
Bispectral features	0.52 (0.33)	0.10355	82
Network Properties	0.64 (0.33)	0.22268	95
Phase synchronization	0.47 (0.33)	0.07846	37

The strong classifier consisting of 7 weak classifiers after feature filtering and a combination of 7 weak classifiers could achieve a better triple classification, and a classification accuracy of 0.9333 (28 out of 30) was obtained on the test set. As can be seen from the confusion matrix (Fig. 7), all patients with dementia were correctly predicted. One HC was predicted as MCI, and one MCI was predicted as dementia. While studies usually consider HC, MCI, and dementia as a continuum of development, this study also looked further at the cognitive test scores (mean MoCAB and MMSE) of the mis-scored sample. The cognitive test score for the misclassified HC sample was 23, while the MCI misclassified sample had a cognitive test score of 18. Considering that the full score on the cognitive test is 30, the lowest score for the HC sample was 22, while the lowest score for the MCI sample was 15, both misclassified samples were also at the lower end of their respective categories.

A continuous cognitive evaluation model can be obtained by restricting and weighing the outcomes of the generated trichotomous model based on the results of the cognitive scale tests (MoCAB and MMSE). After mapping the true and predicted scores to a score between 1 and 100 using the procedure section of the algorithm, the mean error in the final prediction of cognitive level for the 30 individuals was 4.63 (Fig. 8). Examining the two misclassified samples again, the true cognitive score for the sample classified as MCI was 67.97 and the predicted value was 62.16, while the true value for the sample misclassified as dementia was 33.15 and the anticipated value was 27.33. The results demonstrate that after the continuous evaluation, the cost of misclassification was lowered, and the prediction of the neighboring category shows that there may be a cognitive effect—the forecasts for the neighboring categories point to a potential pattern of perceived deterioration in this sample. Overall, the findings indicate that the approach accurately predicts cognitive function from resting EEG. The mean prediction error was 4.82 on a scale of 1-100 (100 being the healthiest and one being seriously demented), resulting in a less than 5% error rate.

This study proposes a method that can predict the cognitive level of subjects based on resting-state EEG and validates the feasibility of the method on a large sample data set, obtaining relatively good predictions. The evaluation model has easy input data collection advantages, low algorithm complexity, and high model interpretability. The method may contribute to the early screening of cognitive decline, the assessment of the effectiveness of therapeutic interventions for neurodegenerative diseases associated with cognitive decline, mainly Alzheimer's disease, the development of drugs for related diseases, and the evaluation of drug efficacy.

One question worth discussing is why many single types of features do not distinguish the three groups well simultaneously, while once they are combined in a specific form, satisfactory classification accuracy can be obtained. Schapire proves that if the target is weakly learnable, it must be possible to make it strongly learnable in some form based on the significant number theorem (Schapire, 1990). In practice, to achieve this goal, the errors of each weak classifier need to be independent of each other. That is, different weak classifiers can provide different information to achieve the error reduction of strong classifiers. The core purpose of this study's sample weight adjustment strategy using the Adaboost method is to ensure an even distribution of correctly classified samples, ultimately making the number of correctly classified samples more than the number of incorrectly classified samples. The results (Fig. 8) of the weak classifier for the predicted samples point out that the starting classification distribution of the samples in this problem is relatively uniform, making it easier to improve the accuracy during further iterations of the weights. In addition, the study also correlated the classification result labels of each weak classifier (Fig. 9). The results showed that each weak classifier did not perform the same, and there was no weak class of classifiers with similar information contributions that could be replaced or removed.

Specifically, in this study, the reason for the success of integrated learning is that the different features pinpoint different levels of variation and increase the amount of valid information at each iteration. The results show that the microstate sequence-related features perform relatively well in the differentiation of the three populations, which may be due to the high temporal resolution of the EEG that allows for the mining of fine-grained systemic transformations that are highly correlated with disease and abnormal activity (Khanna et al., 2015). Spatial network features performed best in distinguishing MCI from dementia but were average for distinguishing HC from MCI. This may be because the structural changes accompanying early neurological decline are not very significant, while the structural differences become increasingly significant as the system collapses (Elmore, 2007).

Another advantage of this work is that the data used contains three different acquisition devices (NeuroScan, Nuracle, BP). The combination and processing based on the feature level are not device dependent and do not require device consistency (Table 2 and Fig. 11), so a multi-device acquisition dataset has no impact on the model building of this work. This study also researches the distribution of the true and predicted values of the prediction sample by an amplifier to ensure the approach is pervasive and cross-device capable. The results showed that the distribution tests between the true and predicted values of the test samples from the three amplifiers did not show any difference (p > 0.1). Furthermore, the absolute prediction error for all three amplifiers is close to the absolute prediction error for the full sample (Total error: 4.63; NeuroScan error: 4.13; Neuracle error: 4.33; BP error: 5.23). The advantage of multiple devices, however, is the increased diversity of data, which also contributes to the generalizability and credibility of the model and establishes a sound basis for future algorithm applications (Cheng et al., 2012).

Table 2

Predictive distribution test for different amplifiers
Amplifiers	Mean absolute error	True Value Distribution Inspection	Predicted Value Distribution Inspection
NeuroScan	4.13	NeuroScan & Nuracle p = 0.32	NeuroScan & Nuracle p = 0.37
Nuracle	4.33	NeuroScan & Nuracle p = 0.40	NeuroScan & Nuracle p = 0.31
BP	5.23	NeuroScan & Nuracle p = 0.21	NeuroScan & Nuracle p = 0.20

Considering the strong correlation between cognitive decline and aging, the contribution of age in the model obtained by fitting based on cognitive decline is also worth discussing; after all, the study does not expect to obtain only one age predictor. For this reason, this study used the same Adaboost method to fit the age and predict the test sample's age. After obtaining the predicted age corresponding to the actual clinical labels of the subjects, it was found that the results of the model fitting seemed to have some age correlation. However, it can be seen from the figure that the results of the model fit do not all depend on age, as there are normal subjects with a high predicted age and demented subjects with low predicted age (Fig. 12).

The method has strong interpretability, which is crucial for clinical purposes. The results of the weak classifier show that the time domain features are relatively well classified, followed by the space domain features and again by the frequency domain features. This fully reflects the high temporal resolution of the EEG signal, which also indicates that the three categories of HC, MCI, and dementia, that is, the process of cognitive decline, differ more significantly in the transformation of neural activity patterns in the temporal dimension. This part of the results is also reflected in the paper. In particular, early cognitive decline, which is poorly identified subjectively, can be better detected by this model.

This study also describes the pattern of cognitive decline in three dimensions, spatiotemporal and frequency, to enhance the interpretability of the evaluation system. Combining the results of this study with previous work, the study concluded that cognitive decline follows four rules: strong and constant strong non-monotonic changes, decreased efficiency, and reduced stability.

The strong and constant strong means the entire neural system does not change uniformly, with strong connections and strong active areas surviving the degradation process better. These strong active areas or strong connections may characterize the person's essential activity capacity. For example, preserving parietal connections (Sun et al., 2021) may be related to central control and basic perception.

Another critical point is that cognitive decline is not a monotonic change, and there could be repair and compensation of the neural system, especially in the early stage(Cabeza et al., 2002). As can be seen from the results, cognitive decline occurs at an accelerated rate, with MCI to dementia changes far exceeding HC to MCI. There is even a tiny amount of additional connectivity and enhanced information from HC to MCI. This may also account for the mediocre performance of linear models for prediction in some studies (Wolz et al., 2011).

Decreased efficiency is also a primary phenomenon. Cognitive decline characterizes the degradation of the nervous system, with reduced activity and information implying reduced processing of external input (Petersen & Posner, 2012). This correlates with the slowness of response, attention deficit, and repetition of questions and answers exhibited by patients with cognitive decline.

Finally, cognitive decline will reduce stability. The normal brain is a stable dynamic equilibrium system, and cognitive decline disrupts this equilibrium (Buckner et al., 2009). Combined with the findings, the jumps in state and the diminished persistence indicate a system disruption. Some patients with cognitive decline exhibit emotional abnormalities, and communication difficulties may be associated with this.

In summary, the study in this paper also provides an idea of constructing a generic EEG activity space with large samples and labels, such that the mapping can characterize various neural activity changes. With the help of some techniques to increase the signal-to-noise ratio of EEG (Alyasseri et al., 2022; Sun et al., 2022), such a framework could extract information from resting-state scalp EEG more effectively. The cognitive decline mapping proposed in this study is only one part of a larger neural activity space. The construction of such a mapping system will be helpful for neurological function assessment, disease development evaluation, clinical target research, and drug effect testing.

Participants

The participants were patients with complaints of cognitive decline from the Memory Specialist Clinic of the Department of Neurology, Tongji Hospital, Tongji University. The Ethics Committee of Tongji Hospital, Shanghai, China, approved the study. All subjects were informed and signed an informed consent form. Patients were required to meet the following criteria: 1) no brain tumor, epilepsy, neurosyphilis, no new cerebral infarction at the time of consultation, and no other central nervous system diseases (infections, clear history of demyelinating diseases, etc.). 2) no major medical or physical diseases such as hepatic encephalopathy or myocardial infarction. 3) no previous history of severe mental disorders, psychoactive substances, or drug abuse; 4) able to cooperate with the examination, complete the full neuropsychological assessment and sign an informed consent form. 5) no contraindications to cranial MRI or electroencephalography. A total of 252 participants were enrolled in this study, including 75 healthy controls,99 MCI, and 78 patients with dementia (Table 3). In addition, EEG was acquired twice per participant, so the data involved in the model building was 150 healthy subjects, 198 MCI, and 156 dementia patients.

Table 3

Subject information
	HC	MCI	dementia
Number	75	99	78
Females / Males	38 / 37	43 / 46	40 / 38
Age (years)	67.9 ± 13.5	70.6 ± 14.2	72.7 ± 13.8
p-value of t-test for age distribution	0.014 (HC and dementia)	0.095 (HC and MCI)	0.021 (MCI and dementia)
MoCAB	24.39 ± 5.39	24.39 ± 5.39	24.39 ± 5.39
MMSE	27.47 ± 3.47	27.47 ± 3.47	27.47 ± 3.47

EEG data acquisition and pre-processing

EEG data acquisition is performed using a 64-lead EEG recording system from Brain Products (www.brainproducts.com), NeuroScan (www.compumedicsneuroscan.com), and Neuracle (www.neuracle.tech), with electrode positions using International 10–20 EEG Society standards. The equipment sampling rate is 1000Hz. The electrode contact resistance was less than 5 KΩ. Resting EEG was recorded with eyes open for at least 5 min before the task EEG was performed on all subjects. EEG signal pre-processing processing included pre-positioning electrodes, removal from 50 Hz industrial frequency interference, 0.5-100Hz band-pass filtering, rejection of useless channels, threshold method to reject bad segments, ICA to remove EMG artifact components and whole brain average re-referencing. EEG signals were processed in this research by MATLAB 2018b platform (ww2.mathworks.cn).

Cognitive scale tests

Each participant was tested on the Montreal Cognitive Assessment Basic (MoCAB) (Nasreddine et al., 2005) and the Mini-Mental State Examination (MMSE) (Tombaugh & McIntyre, 1992) scales to determine their cognitive level. MoCAB is a quick cognitive impairment assessment tool, and the MMSE is a technique for detecting cognitive and intellectual deterioration in older persons. The scales test the capability of visuospatial, executive, naming, memory, attention, language, abstraction, delayed recall, orientation through charts, answers for questions and choices. Both scales have been translated into several languages and are widely used around the world.

Time domain feature extraction (fractal features)

The time domain feature extraction mainly considers the time-varying characteristics of the EEG signal. In this part, except for the microstate calculation, 50 trials 2s segments are segmented for superimposed averaging. Fractal features characterize the waveform of the EEG, and in this work, the fractal features of nine EEGs were calculated. The specific calculations are shown in Table 5, assuming an EEG signal of X.

Table 4

Fractal features
Feature	Description	Calculation
Peak value (pk)	Difference between the maximum and minimum values	max(X)-min(X)
Rectification average value (av)	Average activity intensity	$\stackrel{-}{\left\|X\right\|}$
Effective value (rm)	Description of power	$\sqrt{\stackrel{-}{{X}^{2}}}$
kurtosis	The extent that the signal deviates from a normal distribution	$\frac{\frac{1}{n}{\sum }_{i=1}^{n}{({X}_{i}-\stackrel{-}{X})}^{4}}{{\left(\frac{1}{n}{\sum }_{i=1}^{n}{({X}_{i}-\stackrel{-}{X})}^{2}\right)}^{2}}$
skewness	Signal symmetry	${\left(\frac{\frac{1}{n}{\sum }_{i=1}^{n}{({X}_{i}-\stackrel{-}{X})}^{3}}{\frac{1}{n}{\sum }_{i=1}^{n}{({X}_{i}-\stackrel{-}{X})}^{2}}\right)}^{\frac{3}{2}}$
margin	Maximum deviation	$\frac{pk}{{\stackrel{-}{\sqrt{\left\|X\right\|}}}^{2}}$
form factor	Shock characteristics of the signal	$\frac{rm}{av}$
Impulse Factor	Shock characteristics of the signal	$\frac{pk}{av}$
crest factor	extremes of peaks in the signal	$\frac{pk}{rm}$

Time domain feature extraction (signal complexity)

Complexity is a statistical measure of the amount of information carried by a signal, and Lempel-Ziv signal complexity (Lempel and Ziv, 1976) is used in this study. It evaluates the probability of emerging signal patterns and the feature of neural activity transitions. Assume that the EEG signal X = [x₁, x₂, …, x_n], x_n denotes the amplitude, while n denotes the time.

$${S}_{n}=\left\{\begin{array}{c}0 {x}_{n}\le {T}_{1}\\ \genfrac{}{}{0pt}{}{1 {{T}_{1}<x}_{n}\le {T}_{2}}{⋮}\\ l {{ T}_{l}<x}_{n}\end{array}\right.; \text{C}=\frac{c{log}_{l}n}{n}$$

In the formula, the original signal (formula1) is segmented into a new sequence S according to l segments. T_l denotes the threshold value obtained by dividing the difference between the maximum and minimum values of the EEG signal according to l-segment averaging. And then, the normalized complexity of the S sequence is calculated (Lempel & Ziv, 1976), in which c is the complexity of sequence S.

Time domain feature extraction (microstate feature)

In this study, the microstate analysis method is used to extract the transformation information of the EEG. In this paper, the segments are segmented according to 100ms, and the standard deviation of the signal within each 100ms is calculated as a single microstate. Each segment of EEG data is divided into a total of 1500 microstates. Each 100ms (0.1s) characterizes a microstate, and the research selects 150 s of high-quality segments from the recorded 5min data and divides them into a sequence of 1500 microstates. Finally, 756000 microstates from all 504 segments of EEG are clustered to obtain a category label for each microstate (divided into 20 classes). The sequence features of each individual microstate sequence were extracted to obtain the state transition information of each data, and the specific calculation features and methods are shown in Table 5. In the table ${S}_{i}$ represents the microstate I, and${ S}_{in}$ is one element of microstate vectors ${S}_{i}={[S}_{i1,}{S}_{i2,}\dots ,{S}_{in}]$.

Table 5

Microstate sequence features
Feature	Description	Calculation
occurrences	Number of occurrences of each state	$\sum {S}_{i}$
Status transformation	Number of interconversions between any two types of states	${\sum }_{i=1}^{20}{\sum }_{j=1}^{20}{S}_{i}{\to S}_{j}$
Conversion distance	Distance between two mutually transitional states	$\sqrt{{\sum }_{1}^{n}{({S}_{in}-{S}_{jn})}^{2}}$
Distribution entropy	Distribution properties of sequences	hctsa toolbox(Fulcher & Jones, 2017)
Approximate entropy	Degree of self-similarity of sequences	hctsa toolbox(Fulcher & Jones, 2017)
auto-mutual information change	Changes in the auto-mutual information with the addition of noise.	hctsa toolbox(Fulcher & Jones, 2017)
generalized linear self-correlation	The generalized linear self-correlation function of a time series.	hctsa toolbox(Fulcher & Jones, 2017)
AR model	Fits an AR model	hctsa toolbox(Fulcher & Jones, 2017)
GARCH model	Fits an GARCH model	hctsa toolbox(Fulcher & Jones, 2017)
KPSS stationarity	The KPSS stationarity test	hctsa toolbox(Fulcher & Jones, 2017)
PP stationarity	Phillips-Peron unit root test	hctsa toolbox(Fulcher & Jones, 2017)
Lyapunov index	Dispersion rate of chaotic systems	hctsa toolbox(Fulcher & Jones, 2017)
Poincare section	Parametric description of non-linear systems	hctsa toolbox(Fulcher & Jones, 2017)
binary statistics	Statistics on a binary symbolization of the time series.	hctsa toolbox(Fulcher & Jones, 2017)

Frequency domain feature extraction

The frequency domain is a classical observational dimension of the signal view. The frequency domain information could generally reflect how much the signal contains different rate activity components. In the signal decomposition process, 50 trials 2s segments of each EEG signal were selected for superimposed averaging, resulting in a frequency domain resolution of 0.5 Hz. At the frequency domain level, the five features shown in Table 6 were calculated in this study. The following standard frequency bands are used in this work: delta band (0.5 Hz − 3 Hz), theta band (4 Hz − 7 Hz), alpha band (8 Hz − 12 Hz), beta band (13 Hz − 20 Hz), gamma band (21 Hz − 40 Hz).

Table 6

Frequency features
Feature	Description	Calculation
Fast Fourier Transform	Frequency domain power distribution of signals	$fft\left(X\right)$
Center of gravity frequency	The point at which the sum of the power equals half the total power	0.5*$\sum fft\left(X\right)$
RMS frequency	root mean square of the spectrum	$\sqrt{\stackrel{-}{{X*fft\left(X\right)}^{2}}}$
variance frequency	frequency of variance	$std\left(fft\left(X\right)\right)$
Bispectral frequencies	Signal symmetry	Method in paper(Glass et al., 1997)

Spatial domain feature

The spatial features calculate the connections and interactions between different brain regions, reflecting the activity patterns of the nervous system. In this part of the study, the EEG was averaged by superimposing 20 trials 2s segments, and then the information transfer between individual leads was calculated based on the directed transfer function (DTF) to form a directed graph. Subsequent calculations were based on the BCT toolbox (Rubinov & Sporns, 2010) to extract spatial features in the EEG, with the feature parameters shown in Table 7. In phase lag index computation, N denotes the time point, ${\varnothing }_{rel}$ denotes the phase difference between the two channel signals at time tn, and sign is a sign function

Table 7

Spatial domain features
Feature	Description	Calculation
directed transfer function	Directed information flow of different signal	eConnectome toolbox (He et al., 2011)
Phase lag index	The degree of phase synchronization of the two signals	$\left\|\frac{1}{N}{\sum }_{n=1}^{N}sign(\varDelta {\varnothing }_{rel}\left({t}_{n}\right))\right\|$
Node degree	Number of neighbors connected to the node	BCT toolbox (Rubinov & Sporns, 2010)
Node strength	Weighted node degree considering connection strength	BCT toolbox (Rubinov & Sporns, 2010)
Clustering coefficient	Network local information processing efficiency	BCT toolbox (Rubinov & Sporns, 2010)
Average path	The average value of the shortest path to pass information between any two nodes	BCT toolbox (Rubinov & Sporns, 2010)
Small World Properties	Distributed information processing capability within the network	BCT toolbox (Rubinov & Sporns, 2010)
Rich Club Properties	Network core critical node interaction capability	BCT toolbox (Rubinov & Sporns, 2010)

Feature selection and model training

In this study, the EEG features of HC, MCI, and dementia were extracted from three spatial and temporal frequency perspectives, which describe the differential information contained in the EEG from different perspectives. Many features are extracted from the research's time, space, and frequency domains, and each feature may contain multiple parameters. To avoid excessive spatiotemporal complexity, this study uses the ReliefF (Robnik-Sikonja & Kononenko, 2003) feature selection method, which is a method of assigning different weights to features by considering the distance between like and unlike, and the core idea of ReliefF is that if a feature makes the closest distance of a sample point to like less than the closest distance of unlike, then the feature is beneficial for classification and the weight of the feature is increased. This study uses the ReliefF method. This study uses the ReliefF method to rank the features in descending order of weight and selects the top N features that maximize the correct classification rate as the final training result of the model.

Different information has different abilities to differentiate between different groups of people at the feature level, so a specific approach is needed to combine and train the valid information to build a cognitive-level assessment model. In terms of feature combination, SVM was first used to construct models for Fractal characteristics, Complexity measures, Microstate features, Spectral characteristics, Bispectral features, Network Properties, and Phase synchronization, and eight weak classifiers was constructed. This study combined the eight weak classifiers into one strong classifier for HC, MCI, and dementia classification using the AdaBoost idea. The SVM method was also used to predict the three categories' EEG and cognitive scale scores, and the results were combined using Adaboost.

AdaBoost procedure (Classification and regression):

Step 1: Each sample is given an initial weight w = 1/N(A total of N samples), and all weak classifiers are used for classification.

Step 2: Select the classifier with the smallest error rate (the sum of the weights of all misclassified sample points is the error rate e) so that its weight is 0.5*log((2-e)/e).

step3: Update the sample weights, the correct sample weight is w/2*(1-e), the wrong sample weight is w/2*e.

Step 4: Repeat Step 2 until all classifiers are assigned weights or several iterations are reached.

According to previous studies, healthy individuals, mild cognitive impairment, and dementia are a continuum of development and are accompanied by a gradual decline in cognitive ability (Langa & Levine, 2014; Morris et al., 2001). This means that these three groups represent the full range of the cognitive level space, and if the interval 0-100 is used as the cognitive level score, the healthy group would have the highest score, the mild cognitive impairment group would have the medium score, and the dementia group would have the lowest score. Based on this criterion, the study divided the 100 equivalents into three ranges. That is, 66.6–100 for the normal group, 33.3–66.6 for the mild cognitive impairment group, and 0-33.3 for the dementia group.

In this study end of the AdaBoost algorithm, all classifiers were given weights, and the final cognitive assessment scores were obtained by summing the classification and cognitive regression results. The final cognitive scores are shown in Formula 2, with the classification results being a default score of 66.6 for HC, 33.3 for MCI, and 0 for dementia.

cognitive score = classification results + 0.5 * (MoCAB + MMSE) / 30*33.3 (2)

The formula implies that the scale scores obtained by EEG fitting are first mapped to 0–33 and then mapped to 0-100 according to subjects' clinical labels (HC, MCI, dementia). The final cognitive scores are shown in Eq. 1, with a default score of 66.6 for HC, 33.3 for MCI, and 0 for dementia. In the formula, 30 represents a score of 30 on the cognitive scale.0.5 represents the mean value of the MoCAB and MMSE scores. 504 EEG cases were included in the study, of which 474 cases were used as training samples and 30 cases were set aside as test samples. The following principles were followed when reserving the test sample: 1. random sampling. 2. 10 cases from HC, 10 from MCI, and 10 from dementia. 3. ensure no sample was taken from the same subject twice. Otherwise, re-sample.

Author contributions

Sun Jingnan involved in experimental design Data collection Data analysis Code design Article writing.Sun Yike and Shen Anruo participated in part of the data processing and article writing. Li Yunxia provided data of human patients. Gao Xiaorong provided guidance on working methods and revised the article. Lu Bai conceived the original idea and supervised the project.

Acknowledgments

We thank the patients with cognitive impairment involved in this work for their cooperation in data collection, and Dr. Li Yunxia for her support. Thanks also to Chen Xiaogang for editing the article and improving the quality of the study.

Competing interests

All authors declare no financial or non-financial competing interests.

Data availability

The datasets used or analyzed during the current study are available from the corresponding author upon reasonable request.

Code availability

The underlying code for this study is not publicly available for proprietary reasons.

Albert, M. S., DeKosky, S. T., Dickson, D., Dubois, B., Feldman, H. H., Fox, N. C., Gamst, A., Holtzman, D. M., Jagust, W. J., Petersen, R. C., Snyder, P. J., Carrillo, M. C., Thies, B., & Phelps, C. H. (2011). The diagnosis of mild cognitive impairment due to Alzheimer's disease: Recommendations from the National Institute on Aging-Alzheimer's Association workgroups on diagnostic guidelines for Alzheimer's disease. Alzheimers & Dementia, 7(3), 270-279. https://doi.org/10.1016/j.jalz.2011.03.008
Alyasseri, Z. A. A., Khader, A. T., Al-Betar, M. A., Yang, X. S., Mohammed, M. A., Abdulkareem, K. H., Kadry, S., & Razzak, I. (2022). Multi-objective flower pollination algorithm: a new technique for EEG signal denoising. Neural Computing & Applications. https://doi.org/10.1007/s00521-021-06757-2
Brown, R. E., Basheer, R., McKenna, J. T., Strecker, R. E., & McCarley, R. W. (2012). CONTROL OF SLEEP AND WAKEFULNESS. Physiological Reviews, 92(3), 1087-1187. https://doi.org/10.1152/physrev.00032.2011
Buckner, R. L., Sepulcre, J., Talukdar, T., Krienen, F. M., Liu, H. S., Hedden, T., Andrews-Hanna, J. R., Sperling, R. A., & Johnson, K. A. (2009). Cortical Hubs Revealed by Intrinsic Functional Connectivity: Mapping, Assessment of Stability, and Relation to Alzheimer's Disease. Journal of Neuroscience, 29(6), 1860-1873. https://doi.org/10.1523/jneurosci.5062-08.2009
Cabeza, R., Anderson, N. D., Locantore, J. K., & McIntosh, A. R. (2002). Aging gracefully: Compensatory brain activity in high-performing older adults. Neuroimage, 17(3), 1394-1402. https://doi.org/10.1006/nimg.2002.1280
Charles, S. T., & Carstensen, L. L. (2010). Social and Emotional Aging. Annual Review of Psychology, 61, 383-409. https://doi.org/10.1146/annurev.psych.093008.100448
Cheng, H. Y., Weng, C. C., & Chen, Y. Y. (2012). Vehicle Detection in Aerial Surveillance Using Dynamic Bayesian Networks. Ieee Transactions on Image Processing, 21(4), 2152-2159. https://doi.org/10.1109/tip.2011.2172798
Crum, R. M., Anthony, J. C., Bassett, S. S., & Folstein, M. F. (1993). POPULATION-BASED NORMS FOR THE MINI-MENTAL-STATE-EXAMINATION BY AGE AND EDUCATIONAL-LEVEL. Jama-Journal of the American Medical Association, 269(18), 2386-2391. https://doi.org/10.1001/jama.269.18.2386
D'Rozario, A. L., Chapman, J. L., Phillips, C. L., Palmer, J. R., Hoyos, C. M., Mowszowski, L., Duffy, S. L., Marshall, N. S., Benca, R., Mander, B., Grunstein, R. R., & Naismith, S. L. (2020). Objective measurement of sleep in mild cognitive impairment: A systematic review and meta-analysis. Sleep Medicine Reviews, 52, Article 101308. https://doi.org/10.1016/j.smrv.2020.101308
Elmore, S. (2007). Apoptosis: A review of programmed cell death. Toxicologic Pathology, 35(4), 495-516. https://doi.org/10.1080/01926230701320337
Fulcher, B., & Jones, N. S. (2017). hctsa: A Computational Framework for Automated Time-Series Phenotyping Using Massive Feature Extraction. Cell Systems, 5(5), 527-+. https://doi.org/10.1016/j.cels.2017.10.001
Gauthier, S., Reisberg, B., Zaudig, M., Petersen, R. C., Ritchie, K., Broich, K., Belleville, S., Brodaty, H., Bennett, D., Chertkow, H., Cummings, J. L., de Leon, M., Feldman, H., Ganguli, M., Hampel, H., Scheltens, P., Tierney, M. C., Whitehouse, P., Winblad, B., & Int Psychogeriatric Assoc Expert, C. (2006). Mild cognitive impairment. Lancet, 367(9518), 1262-1270. https://doi.org/10.1016/s0140-6736(06)68542-5
Glass, P. S., Bloom, M., Kearse, L., Rosow, C., Sebel, P., & Manberg, P. (1997). Bispectral analysis measures sedation and memory effects of propofol, midazolam, isoflurane, and alfentanil in healthy volunteers. Anesthesiology, 86(4), 836-847. https://doi.org/10.1097/00000542-199704000-00014
He, B., Dai, Y. K., Astolfi, L., Babiloni, F., Yuan, H., & Yang, L. (2011). eConnectome: A MATLAB toolbox for mapping and imaging of brain functional connectivity. Journal of Neuroscience Methods, 195(2), 261-269. https://doi.org/10.1016/j.jneumeth.2010.11.015
Jafari, Z., Kolb, B. E., & Mohajerani, M. H. (2020). Neural oscillations and brain stimulation in Alzheimer's disease. Progress in Neurobiology, 194, Article 101878. https://doi.org/10.1016/j.pneurobio.2020.101878
Jeong, J. S. (2004). EEG dynamics in patients with Alzheimer's disease. Clinical Neurophysiology, 115(7), 1490-1505. https://doi.org/10.1016/j.clinph.2004.01.001
Khanna, A., Pascual-Leone, A., Michel, C. M., & Farzan, F. (2015). Microstates in resting-state EEG: Current status and future directions. Neuroscience and Biobehavioral Reviews, 49, 105-113. https://doi.org/10.1016/j.neubiorev.2014.12.010
Langa, K. M., & Levine, D. A. (2014). The Diagnosis and Management of Mild Cognitive Impairment A Clinical Review. Jama-Journal of the American Medical Association, 312(23), 2551-2561. https://doi.org/10.1001/jama.2014.13806
Lempel, A., & Ziv, J. (1976). COMPLEXITY OF FINITE SEQUENCES. Ieee Transactions on Information Theory, 22(1), 75-81. https://doi.org/10.1109/tit.1976.1055501
Morris, J. C., Storandt, M., Miller, J. P., McKeel, D. W., Price, J. L., Rubin, E. H., & Berg, L. (2001). Mild cognitive impairment represents early-stage Alzheimer disease. Archives of Neurology, 58(3), 397-405. https://doi.org/10.1001/archneur.58.3.397
Nasreddine, Z. S., Phillips, N. A., Bedirian, V., Charbonneau, S., Whitehead, V., Collin, I., Cummings, J. L., & Chertkow, H. (2005). The montreal cognitive assessment, MoCA: A brief screening tool for mild cognitive impairment. Journal of the American Geriatrics Society, 53(4), 695-699. https://doi.org/10.1111/j.1532-5415.2005.53221.x
Palop, J. J., & Mucke, L. (2009). Epilepsy and Cognitive Impairments in Alzheimer Disease. Archives of Neurology, 66(4), 435-440. https://doi.org/10.1001/archneurol.2009.15
Petersen, R. C. (2004). Mild cognitive impairment as a diagnostic entity. Journal of Internal Medicine, 256(3), 183-194. https://doi.org/10.1111/j.1365-2796.2004.01388.x
Petersen, R. C., Smith, G. E., Waring, S. C., Ivnik, R. J., Tangalos, E. G., & Kokmen, E. (1999). Mild cognitive impairment - Clinical characterization and outcome. Archives of Neurology, 56(3), 303-308. https://doi.org/10.1001/archneur.56.3.303
Petersen, S. E., & Posner, M. I. (2012). The Attention System of the Human Brain: 20 Years After. In S. E. Hyman (Ed.), Annual Review of Neuroscience, Vol 35 (Vol. 35, pp. 73-89). https://doi.org/10.1146/annurev-neuro-062111-150525
Rao, S. M., Leo, G. J., Bernardin, L., & Unverzagt, F. (1991). COGNITIVE DYSFUNCTION IN MULTIPLE-SCLEROSIS .1. FREQUENCY, PATTERNS, AND PREDICTION. Neurology, 41(5), 685-691. https://doi.org/10.1212/wnl.41.5.685
Robnik-Sikonja, M., & Kononenko, I. (2003). Theoretical and empirical analysis of ReliefF and RReliefF. Machine Learning, 53(1-2), 23-69. https://doi.org/10.1023/a:1025667309714
Rubinov, M., & Sporns, O. (2010). Complex network measures of brain connectivity: Uses and interpretations. Neuroimage, 52(3), 1059-1069. https://doi.org/10.1016/j.neuroimage.2009.10.003
Schapire, R. E. (1990). THE STRENGTH OF WEAK LEARNABILITY. Machine Learning, 5(2), 197-227. https://doi.org/10.1007/bf00116037
Sun, J. N., He, J., & Gao, X. R. (2021). Neurofeedback Training of the Control Network Improves Children's Performance with an SSVEP-based BCI. Neuroscience, 478, 24-38. https://doi.org/10.1016/j.neuroscience.2021.08.010
Sun, Y. K., Shen, A. R., Sun, J. N., Du, C. L., Chen, X. G., Wang, Y. J., Pei, W. H., & Gao, X. R. (2022). Minimally Invasive Local-Skull Electrophysiological Modification With Piezoelectric Drill. Ieee Transactions on Neural Systems and Rehabilitation Engineering, 30, 2042-2051. https://doi.org/10.1109/tnsre.2022.3192543
Thapa, N., Park, H. J., Yang, J. G., Son, H., Jang, M., Lee, J., Kang, S. W., Park, K. W., & Park, H. (2020). The Effect of a Virtual Reality-Based Intervention Program on Cognition in Older Adults with Mild Cognitive Impairment: A Randomized Control Trial. Journal of Clinical Medicine, 9(5), Article 1283. https://doi.org/10.3390/jcm9051283
Tombaugh, T. N., & McIntyre, N. J. (1992). THE MINI-MENTAL-STATE-EXAMINATION - A COMPREHENSIVE REVIEW. Journal of the American Geriatrics Society, 40(9), 922-935. https://doi.org/10.1111/j.1532-5415.1992.tb01992.x
West, R. L. (1996). An application of prefrontal cortex function theory to cognitive aging. Psychological Bulletin, 120(2), 272-292. https://doi.org/10.1037/0033-2909.120.2.272
Wolz, R., Julkunen, V., Koikkalainen, J., Niskanen, E., Zhang, D. P., Rueckert, D., Soininen, H., Lotjonen, J., & Alzheimers Dis, N. (2011). Multi-Method Analysis of MRI Images in Early Diagnostics of Alzheimer's Disease. Plos One, 6(10), Article e25446. https://doi.org/10.1371/journal.pone.0025446

(Not answered)

Supplementaryfigure.docx

Download PDF

Journal Publication

published 02 Jan, 2024

Read the published version in npj Aging →

Version 1

posted

You are reading this latest preprint version

A Cross-Device Ensemble Learning Model for Continuous Cognition Assessment Based on Resting-State EEG

Status:

Journal Publication

Version 1

Abstract

Figures

Introduction

Results

Discussion

METHODOLOGY

Participants

EEG data acquisition and pre-processing

Cognitive scale tests

Time domain feature extraction (fractal features)

Time domain feature extraction (signal complexity)

Time domain feature extraction (microstate feature)

Frequency domain feature extraction

Spatial domain feature

Feature selection and model training

Declarations

References

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1

Feature	Description	Calculation
occurrences	Number of occurrences of each state	\(\sum {S}_{i}\)
Status transformation	Number of interconversions between any two types of states	\({\sum }_{i=1}^{20}{\sum }_{j=1}^{20}{S}_{i}{\to S}_{j}\)
Conversion distance	Distance between two mutually transitional states	\(\sqrt{{\sum }_{1}^{n}{({S}_{in}-{S}_{jn})}^{2}}\)
Distribution entropy	Distribution properties of sequences	hctsa toolbox(Fulcher & Jones, 2017)
Approximate entropy	Degree of self-similarity of sequences	hctsa toolbox(Fulcher & Jones, 2017)
auto-mutual information change	Changes in the auto-mutual information with the addition of noise.	hctsa toolbox(Fulcher & Jones, 2017)
generalized linear self-correlation	The generalized linear self-correlation function of a time series.	hctsa toolbox(Fulcher & Jones, 2017)
AR model	Fits an AR model	hctsa toolbox(Fulcher & Jones, 2017)
GARCH model	Fits an GARCH model	hctsa toolbox(Fulcher & Jones, 2017)
KPSS stationarity	The KPSS stationarity test	hctsa toolbox(Fulcher & Jones, 2017)
PP stationarity	Phillips-Peron unit root test	hctsa toolbox(Fulcher & Jones, 2017)
Lyapunov index	Dispersion rate of chaotic systems	hctsa toolbox(Fulcher & Jones, 2017)
Poincare section	Parametric description of non-linear systems	hctsa toolbox(Fulcher & Jones, 2017)
binary statistics	Statistics on a binary symbolization of the time series.	hctsa toolbox(Fulcher & Jones, 2017)

Feature	Description	Calculation
Fast Fourier Transform	Frequency domain power distribution of signals	\(fft\left(X\right)\)
Center of gravity frequency	The point at which the sum of the power equals half the total power	0.5*\(\sum fft\left(X\right)\)
RMS frequency	root mean square of the spectrum	\(\sqrt{\stackrel{-}{{X*fft\left(X\right)}^{2}}}\)
variance frequency	frequency of variance	\(std\left(fft\left(X\right)\right)\)
Bispectral frequencies	Signal symmetry	Method in paper(Glass et al., 1997)