Investigation of the MobileNetV2 Optimal Feature Extraction Layer for EEG-Based Dementia Severity Classification: A Comparative Study

Abstract

Diagnosing dementia and recognizing substantial cognitive decline are challenging tasks. Thus, the objective of this study was to classify electroencephalograms (EEGs) recorded during a working memory task in 15 patients with mild cognitive impairment (MCogImp), 5 patients with vascular dementia (VasD), and 15 healthy controls (NC). Before creating spectrogram pictures from the EEG dataset, the data were subjected to preprocessing, which included preprocessing using conventional filters and the discrete wavelet transformation. The convolutional neural network (CNN) MobileNetV2 was employed in our investigation to identify features and assess the severity of dementia. The features were extracted from five layers of the MobileNetV2 CNN architecture—convolutional layers (‘Conv-1’), batch normalization (‘Conv-1-bn’), clipped ReLU (‘out-relu’), 2D Global Average Pooling (‘global-average-pooling2d1’), and fully connected (‘Logits’) layers. This was carried out to find the efficient features layer for dementia severity from EEGs. Feature extraction from MobileNetV2’s five layers was carried out using a decision tree (DT) and k-nearest neighbor (KNN) machine learning (ML) classifier, in conjunction with a MobileNetV2 deep learning (DL) network. The study’s findings show that the DT classifier performed best using features derived from MobileNetV2 with the 2D Global Average Pooling (global-average-pooling2d-1) layer, achieving an accuracy score of 95.9%. Second place went to the characteristics of the fully connected (Logits) layer, which achieved a score of 95.3%. The findings of this study endorse the utilization of deep processing algorithms, offering a viable approach for improving early dementia identification with high precision, hence facilitating the differentiation among NC individuals, VasD patients, and MCogImp patients.

1. Introduction

In terms of prevalence, vascular dementia (VasD) is second only to Alzheimer’s disease (AD) [1]. Roughly 20% of dementia cases globally are vascular dementia [2,3]. The cognitive impairment in memory known as vascular dementia (VasD) is linked to cerebrovascular disease (CVD) [4]. Vascular disease is most often linked to stroke, which is the second most prevalent vascular illness. When compared to motoric dysfunction, the impact of stroke on cognitive performance is frequently underappreciated [5]. The patient’s health difficulties and their reliance on others around them may be worsened by cognitive impairment, according to this theory. A total of 20 to 25% of stroke patients will experience delayed dementia, while 15 to 30% will acquire dementia within three months following the stroke, according to prior research [6]. As people become older, a decline in cognitive function brought on by dementia can have far-reaching consequences, such as a decline in everyday social activities and higher expenses for families, communities, and the government. Consequently, it is crucial to recognise dementia at an early stage to halt its rapid progression [6,7].

Vascular dementia is difficult to diagnose because it requires a clinical diagnosis of cognitive impairment that is severe enough to be dementia, as well as imaging evidence from the brain showing that the dementia is caused by cerebrovascular events, such as a stroke [8]. For the diagnosis of dementia following a stroke, the electroencephalogram (EEG) is one of the biomarkers that could serve as an alternative when researching brain function with dementia [9].

Early detection, severity evaluation, and dementia type differentiation have been the recent foci of EEG research on dementia [10]. The EEG has been the tool of choice for the majority of investigations that use Alzheimer’s (AD) detection or characterization [11,12]. On the other hand, research on dementia following a stroke is limited. It is important to investigate the use of EEG for the diagnosis of post-stroke dementia because the incidence of stroke is rising annually and is followed by the risk of dementia. Normative participants and stroke patients with cognitive impairment had their EEG signals characterized in a recent study. Compared to healthy subjects, those with cognitive impairment following a stroke had distinct power spectrum features [13]. Additional research on the use of EEG for the diagnosis of post-stroke dementia has been investigated by evaluating the signal complexity by applying entropy measures of EEG dynamics as one of the complexity degrees [14]. It is believed that a cognitive function involving a change in signal pattern in response to a stimulus is closely related to the layer of EEG spectral dynamics. Cognitively impaired participants had fewer high relative power layers and an increased number of low relative power layers compared to healthy controls [13].

Researchers have looked into employing several machine learning (ML) algorithms to classify EEGs in people with cognitive problems [13]. However, standard ML methods have a hard time dealing with high-dimensional data since they depend on features that are clearly specified. Deep learning (DL), on the other hand, is a revolutionary new development in ML that fixes these problems. DL can learn and find latent discriminative features directly from raw data using end-to-end learning, which gives it a big edge over traditional approaches [15,16,17].

Because people with dementia typically demonstrate EEG waves that show less non-linear behavioral dynamics and linear connection in the cortex of the brain, this study could provide useful information about the disease. Both the complexity and the functional linkages can be reduced by incorporating these aspects.

Additionally, age and psychopathic illness are known to significant in relation to changing EEG frequency profiles, which are important for explaining brain activity in neurological and cognitive studies. The process of aging is associated with a change in EEG power, including a reduction in alpha (8–13 Hz) and an increase in Delta (1–4 Hz) and Thea (4–8 Hz) activity, reflecting converted cortical excitability and connection [18]. Similarly, psychiatry diseases, such as depression or anxiety, which are normal comorbidities in dementia, can change the EEG profile by increasing the slow wave activity or by reducing beta (13–30 Hz) coherence [19]. Moreover, the coherence of the cerebral EEG and muscular electromyogram (EMG) signal, as well as the extent to which the activity of EEG-EMG can be synchronized, is a significant method of studying sensorimotor integration and its variation during neurological disorders such as dementia. The studies cited by the reviewer form a good foundation to this study. James et al. have demonstrated that EEG-EMG coherence and cumulant changes over time, as well as corticospinal oscillations in the cortex, reduce with aging [20]. Mima and Hallett have provided a detailed review of corticomuscular coherence and its importance in motor control; this could explain the interpretation of EEG patterns [21]. Moreover, the physiological principles and clinical implications of our EEG-EMG frequency analysis have been elucidated by Grosse et al., thus justifying the relevance of using a high frequency range in identifying motor-related aberrations [22]. Liu et al. have provided another evaluation of corticomuscular coherence applications, but focus on their possible use in monitoring neurodegenerative diseases, which is consistent with our aims of dementia classification [23]. Tuncel et al. have explored the concept of time–frequency coherence in the state of tiredness and offer a perspective into the interactions between dynamic signals that can enhance our preprocessing resiliency [24].

Dementia patients’ non-linear EEG dynamics must be investigated if this problem is to be better understood. In the early stages of dementia, the EEG may appear normal and have the same rhythm as healthy individuals of the same age. To assess the influence of dementia on the brain and to understand the progression of the disease, it is necessary to analyse and interpret EEG recordings obtained from individuals with dementia using signal analysis techniques. In this study, MobileNetV2 was employed to extract features from its CNN architecture. Specifically, we looked for the best features to use for EEG-based dementia severity assessment in the following five layers: convolutional (‘Conv-1’), batch normalisation (‘Conv-1-bn’), clipped ReLU (‘out-relu’), 2D Global Average Pooling (‘global-average-pooling2d1’), and fully connected (‘Logits’) layers. Feature extraction from MobileNetV2’s five layers was carried out using a DT and KNN ML classifier, in conjunction with a MobileNetV2 DL network. To better distinguish between NC people, VasD patients, and MCogImp patients, these findings provide credence to the utilisation of deep processing algorithms.

2. Materials and Methods

To help doctors figure out what is wrong with people who have had a stroke, the recorded EEG goes through several steps that separate the signals of dementia patients. Figure 1 demonstrates the process flow diagram of the suggested strategy.

2.1. Participants and EEG Signal Acquisition

There were 35 people in this study, divided into three groups: 15 healthy normal controls (NC; 8F/7M; mean age 60.06 ± 5.21 years), 15 patients with post-stroke mild cognitive impairment (MCogImp; 10F/5M; mean age 60.26 ± 7.77 years), and 5 patients with vascular dementia (VD; 2F/3M; mean age 64.6 ± 4.8 years). Patients with MCogImp and patients with VasD were recruited from Pusat Perubatan Universiti Kebangsaan Malaysia ($PPUKM$). People in the NC group had never had a mental illness before. The $PPUKM$ Human Ethics Committee granted the study protocols with their stamp of approval, and all participants signed a paper saying they understood what they were signing up for. The Nicolet One (V32) system was used to capture EEG activity while participants performed an auditory working memory task [25].

2.2. Preprocessing Steps

• Conventional Filtering and Wavelet Denoising
  In the first step, each channel of EEG-based working memory datasets was tackled with conventional filters, including a notch-filter at 50 Hz and a 0.1–70 Hz band-pass-filter to remove high-frequency noise, such as muscle-related EMG signals typically exceeding 20–50 Hz [26,27]. Additionally, the discrete wavelet transform (DWT) was an additional preprocessing procedure that was used to enhance data quality and to reduce EMG contamination in EEG signals [27].
• Spectrogram Images
  The signals from each of the denoised EEG electrodes were split into 6 segments of 10 s each. Each epoch has $N=2560$ samples because $fs=256$ Hz. Therefore, each problem that needed to be examined was saved on a computer and dealt with separately. The power spectral densities ($PSDs$) were figured out from these segments. A periodogram, which determined the frequency distribution of the EEG data, is a good way to find $PSDs$ [28]. The Fourier transform of the autocorrelation function is a nonparametric evaluation of $PSDs$. For example, Equation (1) shows the periodogram of a signal $x_l$ with a length of L [28], as follows:

  $$PSD\left(f\right)={\displaystyle \frac{T_s}{L}}{\left|\sum _{l=0}^{L-1}{x}_l{e}^{-2\pi ifl}\right|}^2$$

  (1)

  where $-{\displaystyle \frac{1}{2T_s}}<f\le {\displaystyle \frac{1}{2T_s}}$ and ${\displaystyle \frac{1}{2T_s}}$ is the sampling period. The following modified periodogram is obtained by multiplying the input time series by the window function $w_l$, as in Equation (2):

  $$\widehat{PSD}\left(f\right)={\displaystyle \frac{T_s}{L}}{\left|\sum _{l=0}^{L-1}{w}_l{x}_l{e}^{-2\pi ifl}\right|}^2$$

  (2)
  The spectrogram function was used to figure out the $PSDs$ of the $x^{th}$ EEG time series for this study. This is because the modified periodogram reduces the spectral leakage of the standard periodogram and softens the edges of the signal. It has a high resolution and can be used to analyse biomedical signs [29].
  The PSD EEG images were turned into three-dimensional images and were resized into $224\times 224$ pixels to fit the CNN MobileNetV2 input design that is used to automatically extract and sort features.

2.3. Deep Feature Extraction with MobileNetV2 Layers

This study employs MobileNetV2 as a feature extractor network to compare the performance of different MobileNetV2 layers for feature identification in order to determine the most appropriate layer for determining the severity of dementia from an EEG dataset. It should be noted that the optimal feature layer selection can vary depending on the specific characteristics of the EEG dataset and the complexity of the classification task.

This classification necessitates differentiating among three categories: NC individuals, MCogImp, and VasD. The input layer MobileNetV2 decreases the image size from 224 by 224 to 14 by 14, while simultaneously increasing the filter response depth from 3 to 512 [30,31]. The last fully connected layer classes have to be reduced from the standard 1000 to 3 classes in order to meet this study classification structure. Hence, MobileNetV2 transfer learning was utilised in this work [32]. To meet the number of classes needed to discriminate dementia severity, three new layers were added: $fullyconnectedlayer\text{-}New$, $softmaxlayer\text{-}New$ and $classificationlayer\text{-}New$. The three previous layers, ‘Logits’, ‘Logits-softmax,’ and ‘ClassificationLayer-Logits’, were removed and replaced with these three. Layers for classification output, those that were completely connected, and softmax layers augment MobileNetV2’s deep CNN architecture’s several convolutional layers.

Afterwards, five feature sets were obtained from the EEG dataset using MobileNetV2 architecture; these are convolutional layers (‘Conv-1’), batch normalisation (‘Conv-1-bn’), clipped ReLU (‘out-relu’), 2D Global Average Pooling (‘global-average-pooling2d-1’), fully connected (‘Logits’), and Softmax (‘Logits-softmax’) layers that extract features from the MobileNetV2 CNN to be classified by the MobileNetV2 classification output layers (‘ClassificationLayer-Logits’) DL and ML classifier.

2.4. Classification

In total, 30% of the data are used for testing, while the remaining 70% are used for training and checking. The 70:30 ratio is a good balance between training and testing data. It allows the model to learn well while still being tested on enough data. This ratio is commonly recommended in the literature [33,34] since it maintains generalization.

The DL $stochasticgradientdescentwithmomentum$ ($sgdm$) optimizer was used. Additionally, the EEG classification model that relies on MobileNetV2 had its hyperparameters chosen using 70% of the training dataset, which included the NC, MCogImp, and VasD groups. Optimizations were made to the learning rate, batch size, and epoch count, which are critical hyperparameters. A validation frequency of 3, a mini-batch size of 64, and a learning rate of $0.001$ were the final hyperparameters chosen.

The evaluation’s scope is enlarged to include the accuracy and confusion matrix in order to evaluate dementia severity prediction performance [35]. These performance measures are chosen and interpreted based on generally accepted machine learning and deep learning principles [36].

As in medical diagnostics, the confusion matrix is an important tool for analyzing classification results since it provides a granular breakdown of model performance that goes beyond total accuracy. In the context of this EEG-based dementia classification study—differentiating NC, MCogImp, and VasD using MobileNetV2 features—its significance resides in analyzing class-specific performance and is compatible with best practices advised in machine learning and deep learning studies [37,38].

First of all, the features were classified using the DL MobileNetV2 classification output layer. Afterwards, the KNN and DT ML classifiers were used.

2.4.1. Deep MobileNetV2 Classifier

Features from convolutional layers (‘Conv-1’), batch normalisation (‘Conv-1-bn’), clipped ReLU (‘out-relu’), 2D Global Average Pooling (‘global-average-pooling2d-1’), fully connected (‘Logits’), and Softmax (‘Logits-softmax’) layers were classified by MobileNetV2’s classification output layer (‘ClassificationLayer-Logits’) DL network. Figure 2 illustrates the MobileNetV2 architecture.

2.4.2. Machine Learning Classifiers

DT and KNN ML classifiers were implemented to identify the features extracted from MobileNetV2’s five layers. In KNN, the training data are obtained from a data model, making it a lazy learning method. Finding similar patterns in the training data is the main idea for classifying the test data. Euclidean distances were used to calculate similarity among data objects. Before classifying a new test sample, its distance from all samples in the training data must be calculated. The next step is to select the class of the new test sample based on the frequency with which k appears in adjacent (neighboring) training samples [39]. The key feature of KNN is the parameter k. In this study, k = 5 was used and the nearest neighbors were also considered in the calculation of their weights so that the nearest ones would have a greater influence on the final stage than those further away.

IF-THEN-ELSE procedures are implemented by decision trees (DTs), which are classifiers that rely on initial criteria or decision rules. To reach a final decision, these conditions are connected in a tree-based framework [40]. The classifier’s output classes are the “leaves” of the DT. A simple feature check is performed at each node of the tree. One possible outcome of the situation is the presence of children in the node. A DT is constructed during training to find the non-joint partitions of the output classes at each node. Typically, the Gini index and the information received act as a classifier for each node. The feature/value combination that maximizes the Gini index or minimizes the information gained is selected for each node. Furthermore, as a postprocessing step, we construct the trees.

3. Results and Discussion

The results of this research offered a comprehensive comparative analysis of the MobileNetV2 model, delivering insights into the effectiveness of the model through the utilization of evaluation metrics including precision, sensitivity, specificity, F-score, and confusion matrix.

3.1. Results of the Preprocessing

During preprocessing, all of the EEG datasets were denoised using conventional filters and WT techniques. Moreover, spectrogram images were created from the $PSDs$ of the EEG dataset to meet the requirement for the MobileNetV2 inputs with a size of $224\times 224$ pixels.

Once the model’s output layers have been taken away to make the end output layer fit the needs of the task at hand, the model’s weights are added to them.

3.2. Results of Deep MobileNetV2 Feature Extraction

The MobileNetV2 model was already trained using data from EEG spectrogram images. Subsequently, five feature groups were extracted from the EEG dataset using the MobileNetV2 CNN algorithm: ‘Conv-1’ from convolutional layers, ‘Conv-1-bn’ from batch normalization, ‘out-relu’ from clipped ReLU, ‘global-average-collection2d-1’ from 2D Global Average Pooling, ‘Logits’ from full correlation, and ‘Logits-softmax’ from Softmax. The classification output layers of the MobileNetV2 CNN consisting of DL and machine learning classifiers were trained with these feature sets.

3.3. Results of Classification

MobileNetV2, as a DL classifier, was used to classify the features extracted from the last five layers of MobileNetv2.

Table 1. The evaluation metric results from the MobileNet Deep learning classifier.

MobileNetV2	Evaluation Metrics	Control	MCI	VaD
	precision	71.2	76.9	76.1
	sensitivity	64.9	70.2	89.5
	specificity	86.8	89.5	86
	accuracy	74.9	74.9	74.9
	F-measure	67.9	73.4	82.3

shows the evaluation metric results from the MobileNetV2 classifier. Moreover, Figure 3 shows the confusion matrix utilising the characteristics and classifier of MobileNetV2 for cases of VasD, MCogImp, and NC.

KNN, as an ML classifier, was used to classify the features extracted from the last five layers of the MobileNetv2.

Table 2. The evaluation metric results using MobileNetV2 features and DT machine learning classifier.

KNN	Evaluation Metrics	Control	MCI	VaD
‘Conv-1’	precision	89.8	96.2	93.2
	sensitivity	93	89.5	96.5
	specificity	94.7	98.2	96.5
	accuracy	93	93	93
	F-measure	91.4	92.7	94.8
‘Conv-1-bn’	precision	89.8	96.2	93.2
	sensitivity	93	89.5	96.5
	specificity	94.7	98.2	96.5
	accuracy	93	93	93
	F-measure	91.4	92.7	94.8
‘out-relu’	precision	59.6	94.7	98.2
	sensitivity	100	80.6	80
	specificity	100	83.2	97.1
	accuracy	84.2	84.2	84.2
	F-measure	74.7	87	88.2
‘global-average-pooling2d-1’	precision	91.7	98.1	94.8
	sensitivity	96.5	91.2	96.5
	specificity	95.6	99.1	97.4
	accuracy	94.7	94.7	94.7
	F-measure	94	94.5	95.7
‘fc’	precision	94.5	93.1	96.6
	sensitivity	91.2	94.7	98.2
	specificity	97.4	96.5	98.2
	accuracy	94.2	94.2	94.2
	F-measure	92.9	93.9	97.4

shows the evaluation metric results from the KNN machine learning classifiers. Moreover, Figure 4, Figure 5, Figure 6, Figure 7 and Figure 8 show the confusion matrix for VasD, MCogImp, and NC using MobileNetV2 features from the five layers and the KNN classifier.

DT, as an ML classifier, was used to classify the features extracted from the last five layers of MobileNetv2.

Table 3. The evaluation metric results using MobileNetV2 features and KNN machine learning classifier.

DT	Evaluation Metrics	Control	MCI	VaD
‘Conv-1’	precision	93.1	96.4	94.7
	sensitivity	94.7	94.7	94.7
	specificity	96.5	98.2	97.4
	accuracy	94.7	94.7	94.7
	F-measure	93.9	95.6	94.7
‘Conv-1-bn’	precision	93.1	96.4	94.7
	sensitivity	94.7	94.7	94.7
	specificity	96.5	98.2	97.4
	accuracy	94.7	94.7	94.7
	F-measure	93.9	95.6	94.7
‘out-relu’	precision	96.4	88.7	98.1
	sensitivity	93	96.5	93
	specificity	98.2	93.9	99.1
	accuracy	94.2	94.2	94.2
	F-measure	94.6	92.4	95.5
‘global-average-pooling2d-1’	precision	94.8	96.4	96.5
	sensitivity	96.5	94.7	96.5
	specificity	97.4	98.2	98.2
	accuracy	95.9	95.9	95.9
	F-measure	95.7	95.6	96.5
‘fc’	precision	96.4	94.8	96.6
	sensitivity	93	96.5	98.2
	specificity	98.2	97.4	98.2
	accuracy	95.3	95.3	95.3
	F-measure	94.6	95.7	97.4

shows the evaluation metric results from the DT machine learning classifiers. Moreover, Figure 9, Figure 10, Figure 11, Figure 12 and Figure 13 show the confusion matrix utilising the characteristics and classifier of DT for cases of VasD, MCogImp, and NC.

This study investigates the last five convolutional MobileNetV2 network architectures, including convolutional layers (‘Conv-1’), batch normalisation (‘Conv-1-bn’), clipped ReLU (‘out-relu’), 2D Global Average Pooling (‘global-average-pooling2d-1’), fully connected (‘Logits’), and Softmax (‘Logits-softmax’) layers, considering the often time- and space-dependent nature of EEG data. As a result, each layer assumes a distinct layer of abstraction in the input EEG spectrogram data. The Conv-1 layer represents the network’s original convolutional layer, whereas the Conv-bn layer is the depth-separating convolutional layer with batch normalisation. Given the frequently time- and space-dependent nature of EEG data, it is better to employ the convolutional layers (Conv-1-bn) and batch normalisation (Conv-1-bn) as feature layers. These layers capture similar low-layer properties, such as edges, textures, and simple patterns, which may be useful for EEG analysis. Moreover, MobileNetV2’s 2D Global Average Pooling (global-average-pooling2d-1) and fully linked (Logits) layers make excellent feature layers for EEG classification. These layers frequently preserve a lot of spatiotemporal information, which might be useful for capturing meaningful patterns in EEG data. Using these layers as feature layers, you can extract significant features from EEG data and export them to the classification output layers of the MobileNetV2, KNN, and DT classifiers. Using transfer learning with DT and KNN ML classifiers produced an improved classification accuracy of 95.9% and 94.7%, respectively, particularly from the 2D Global Average Pooling (‘global-average-pooling2d-1’) layer, which is considered the optimal feature extraction layer because it captures fine features like edges, textures, and simple patterns that may be suitable for EEG investigation; the MobileNetV2 DL network, as a classifier, obtained the lowest classification accuracy of 74.9%. The research results showed that the DT ML classifier, when used in conjunction with features extracted from different MobileNetV2 CNN layers, produced output from the 2D Global Average Pooling (global-average-pooling2d-1) layer with an accuracy of 95.9%, followed by 95.3% from the fully connected (Logits) layer features. These results provide support for the use of deep processing systems to enhance the differentiation between NC subjects, VasD, and MCogImp patients.

4. Conclusions

This research aimed to classify the EEGs of MCogImp and VasD patients with NC individuals while they performed a working memory task. The EEG data were subjected to preprocessing analysis including noise removal using conventional filters and the DWT, while spectrogram images were made from the EEG dataset. Our study used MobileNetV2 (a CNN) to find relevant features and rate the severity of vascular dementia. Therefore, to find the optimum feature layer for EEG-based vascular dementia severity assessment, the features were specifically extracted from five layers of the MobileNetV2 CNN architecture: convolutional layers (‘Conv-1’), batch normalisation (‘Conv-1-bn’), clipped ReLU (‘out-relu’), 2D Global Average Pooling (‘global-average-pooling2d1’), and fully connected (‘Logits’) layers. As a classifier, a MobileNetV2 DL network and DT and KNN ML classifiers were implemented to identify the features extracted from MobileNetV2’s five layers. This study’s results demonstrate that features obtained from MobileNetV2 with the 2D Global Average Pooling (global-average-pooling2d-1) layer yielded the most accurate results for the DT ML classifier, with an accuracy score of 95.9%. The fully connected (Logits) layer features came in second, with a score of 95.3%. These results provide support for the use of deep processing systems to enhance the differentiation between NC subjects, VasD, and MCogImp patients. The limitations of this study include the relatively small sample size (35 participants). The dependence on a singular working memory task may inadequately represent the varied EEG patterns linked to dementia progression, perhaps constraining the model’s relevance across numerous cognitive circumstances. Future work intends to use XGBoost, SVR, and possibly more models such as Random Forest on identical EEG spectrogram datasets, utilizing the same preprocessing workflow. Moreover, the integration of EEG-EMG coherence metrics will be utilized to improve dementia severity assessment.