Hybrid Optimized Techniques for EEG Based Mild Cognitive Impairment Detection Using Time Domain Feature Extraction

MCI is a state where there is a noticeable decline in cognitive abilities greater than what would typically be anticipated for an individual's age, yet it does not lead to significant impairment in their day-to-day activities. Early detection of MCI is crucial for timely intervention and management of cognitive decline and if not identified earlier may lead to Dementia or Alzheimer’s disease [1, 2]. Electroencephalography (EEG) is becoming increasingly recognized as a valuable tool in the detection of MCI due to its nonintrusive characteristics and ability to capture neural activity in real-time [3].

The removal of Electrooculogram (EOG) artifacts in diverse EEG signals is achieved using a novel approach that combines Independent Component Analysis (ICA) and Ensemble Empirical Mode Decomposition (EEMD) is introduced to demonstrate improved performance in EOG artifact rejection, making it a promising method for EEG signal processing and analysis [4]. A comprehensive artifact removal methods like Principal Component Analysis (PCA), Canonical correlation analysis (CCA), Empirical Mode Decomposition (EMD), and Filtering methods for EEG signals were discussed for effectively removing artifacts from EEG signals [5].

For detecting cognitive impairments and neurological disorders, the application of Permutation Entropy (PE) and Statistical Complexity (SC) in the case of MEG recordings from patients detected with MCI and Alzheimer's disease (AD) examines broadband signals and breaks them down into frequency bands to identify changes in each band linked to MCI and AD [6]. The feature extraction scheme based on discrete wavelet transform demonstrates the potential of relative wavelet energy features to categorize EEG signals obtained during intricate cognitive tasks and rest conditions. The performance evaluation with different classifiers highlights the ability of this approach to achieve high accuracy rates in EEG signal classification [7].

The effectiveness of various machine learning algorithms by feature extraction of input EEG samples belonging to motor movements is conducted and shown that the Medium-ANN model emerges as the top performer, suggesting the applicability of the approach in scenarios like brain-computer interfaces and neural prostheses [8]. Regarding neural decoding and brain computer interfaces, Saeidi et al. [9] illustrate the latest developments in the decoding and categorization of EEG signals made possible by supervised ML and DL models. This study explores the increasing importance of deep learning in the analysis of EEG data through a range of tasks, such as finding emotions, mental workload assessment, detection of seizures, program-based potential identification, and sleep scoring [10]. Furthermore, the research delves into the input formulations employed for deep network training and investigates distinct network architectures tailored to various task categories. CNN and RNN consistently demonstrate superior classification accuracy compared to alternative models.

Existing methods for processing EEG signals and classifying MCI are helpful but still have many issues that need fixing. Techniques like PCA, CCA, EMD, and filtering often have trouble separating noise from real brain signals, especially when the data is messy or comes from different sources. Even though ICA and EEMD perform better, they are too complicated and slow to work well in real-time situations.

For feature extraction, methods like Permutation Entropy (PE) and Statistical Complexity (SC) are promising but aren’t fully designed to work well with EEG data. They miss out on important patterns in the signal. Wavelet-based methods also depend on choosing the right settings, which might not work for all tasks. Machine learning and deep learning models like CNNs and RNNs often do well but can struggle with overfitting, require a lot of labeled data, and might not work as well on new datasets. Most studies focus on simpler tasks like detecting movements or emotions, so they don’t work well for complex conditions like MCI. There’s also not enough research comparing these models across bigger and more varied datasets to figure out which ones work best.

The proposed model tries to solve these problems by improving artifact removal, using features from time-domain analysis, and comparing advanced models like 1D CNN, 1D CRNN, 1D PBCNN, and 1D PBCRNN on different EEG datasets. The goal is to make it easier to detect MCI early and create better tools to understand and diagnose this condition.

3.1 Preprocessing techniques

In recent times, there has been a rising interest in the examination of EEG data for the detection of MCI. However, the raw EEG signals are often noisy and contain various artifacts that can hinder accurate interpretation. To address these challenges, preprocessing techniques, such as filtering and epoch segmentation, are applied to enhance the signal quality and facilitate subsequent analysis [18]. The EEG data is preprocessed to enhance its quality and prepare it for further analysis. Firstly, the EEG data is read from EDF files using the MNE library. Then the data undergoes re-referencing with a common average reference to reduce the presence of noise. A bandpass filter is applied to retain frequencies between 0.5 Hz and 45 Hz, eliminating unwanted artifacts. Subsequently, the data is segmented into fixed-length epochs of 5 seconds with a 1-second overlap, facilitating easier analysis. Finally, the preprocessed EEG data is organized into epochs and can be utilized for various tasks like feature extraction or classification. Figure 1 displays a one-minute duration sample EEG signal.

1.png

Figure 1. Sample EEG signal shown for 1 minute

3.2 Feature extraction techniques

In addition to preprocessing, feature extraction plays a vital role in capturing appropriate data from EEG signals. Time domain features, including Kurtosis and Hjorth parameters and Zero Crossing Rate (ZCR) have shown promise in quantifying signal characteristics associated with cognitive impairment. These features provide insights into the amplitude, variability, and distribution of the EEG signals, which can be indicative of underlying cognitive dysfunction.

3.2.1 Time domain features

A time domain feature is a statistical measure that characterizes the amplitude or waveform of an EEG signal at a specific point in time [19].

a) Kurtosis:

Kurtosis is a valuable time domain feature for EEG signal analysis, as it provides insights into the shape and distribution of EEG data. Extracting kurtosis features allow researchers and clinicians to capture and analyze important characteristics of EEG signals related to brain activity. Kurtosis helps us to identify the peakedness or flatness of the distribution of brain activity [20]. If kurtosis is high, it indicates that the brain signal has high spikes. These sharp peaks can happen during very strong brain activity or when something is wrong, like seizures, memory problems, or diseases like Alzheimer’s. Low kurtosis shows that brain activity is normal.

By quantifying the degree of these peaks and tails, kurtosis allows the detection of irregularities in neural activity, which are often associated with conditions like MCI or Alzheimer’s. The kurtosis formula is shown in Eq. (1).

$Kurtosis=\frac{1}{n} \sum_{i=1}^n\left(\frac{x_i-\bar{x}}{s}\right)^4-3$   (1)

where, $n$ stands for the total number of data points within the dataset, $i$ takes on values $1,2,3, \ldots, n, x_i$ represents each specific data point in your dataset, $\bar{x}$ represents the sample mean or average of your dataset.

This is computed by adding together all the data points $\sum x_i$ and then dividing the sum by the total number of data points $n$. $s$ represents the sample standard deviation of your dataset. $\left(\frac{x_i-\bar{x}}{s}\right)^4$ calculates the fourth $s$ power of the standardized deviation of each data point. It measures how each data point deviates from the mean, scaled by the standard deviation. The final step in the formula is subtracting 3. This is done to make the kurtosis of a normal distribution equal to 0. In other words, if the calculated Kurtosis is greater than 0, it indicates heavier tails than a normal distribution, and if it's less than 0, it indicates lighter tails.

The input provided for the kurtosis_features function is a 3-dimensional array with dimensions (number of epochs, number of channels, number of data points per channel). This array contains EEG data for two groups: the control group and the patient group.

The kurtosis_features function is designed to calculate the kurtosis value for each epoch and each channel separately. As the function iterates through each epoch and channel, it calculates the kurtosis value for the data recorded in that specific epoch and channel using the "kurtosis" function from the scipy. stats module. The resulting kurtosis values are stored in a new array called "kurtosis_features." After the function has processed all the epochs and channels, it returns the "kurtosis_features" array. The shape of this array is (num_epochs, num_channels), where "num_epochs" initializes the number of epochs, and "num_channels" signifies the count of electrodes.In the input data, there are 14757 epochs in MCI group and 13645 epochs in HC group, and each epoch contains 19 channels. Therefore, the resulting "Patient kurtosis features" and "Control kurtosis features" arrays have shapes (14757, 19) and (13645, 19), respectively.

b) Hjorth:

Hjorth features are a group of time-domain attributes utilized for the characterization of signals, particularly in the context of EEG signals. They were introduced by Dr. Bjorn Hjorth to quantify the activity, mobility and complexity of signals. In EEG signal processing, Hjorth features provide valuable information about the signal's mobility and complexity. Mobility represents the rate of change or activity in the signal, while complexity measures how irregular or diverse the signal's behavior is over time.

Hjorth Mobility shows us how often the EEG signal varies over time. It shows whether the signal is constant or fluctuating. The mobility with a higher score means that the signal is varying at a faster rate and this can occur during increased brain activity or even brain pathology. A low mobility number indicates that the signal is much more constant and smoother. This is useful since it allows one to determine changes in brain activity for instance when the brain is in function or if there are any functional abnormalities. The mobility evaluation is shown in Eq. (2).

$Mobility=\sqrt{\frac{\operatorname{var}\left(\frac{d x(t)}{d t}\right)}{\operatorname{var}(x(t))}}$    (2)

where, $\left(\frac{d x(t)}{d t}\right)$ represents the first derivative of the function $x(t)$ over time $\mathrm{t}, \operatorname{var}\left(\frac{d x(t)}{d t}\right)$ calculates the variance of $\left(\frac{d x(t)}{d t}\right) \cdot \operatorname{var}(x(t))$ calculates the variance of the position values $x(t)$ over time. $\sqrt{\frac{\operatorname{var}\left(\frac{d x(t)}{d t}\right)}{\operatorname{var}(x(t))}}$ calculates the square root of the ratio between the variance of the signal's derivative and the variance of the signal.

Hjorth Complexity measures the signal’s irregularity or variability, capturing how unpredictable or difficult the signal's behavior is higher complexity means the brain activity is more complicated and has lots of different patterns. This can happen when the brain is solving problems or when something is wrong. Lower complexity means the brain activity is simpler, more repetitive, or not as busy. The calculation of hjorth complexity is shown in Eq. (3).

$Complexity=\sqrt{\frac{\operatorname{Mobility}\left(\frac{d y(t)}{d t}\right)}{\operatorname{Mobility}(y(t))}}$    (3)

where, $Mobility\left(\frac{d y(t)}{d t}\right)$ calculates the first derivative of Hjorth Mobility and $\operatorname{Mobility}(y(t))$ calculates the Hjorth Mobility of the original signal, measuring its overall mobility [21].

The input to the hjorth_features function is a 3D array called data_array, which contains the EEG signal data. To calculate the Hjorth mobility and complexity features, the function iterates over each epoch and channel in the data_array. For each epoch and channel, It computes the signal's first derivative to determine mobility, and then further computes the first derivative of the first derivative to calculate complexity.

In total, the function extracts two sets of features for each epoch and channel: the Hjorth mobility features and the Hjorth complexity features. Since there are num_epochs epochs and num_channels channels, the resulting hjorth_mobility_features and hjorth_complexity_features arrays have the shape (num_epochs, num_channels). To combine the mobility and complexity features into a single feature vector, the function uses np.column_stack to stack the two feature arrays horizontally. As a result, the final Hjorth features arrays, control_hjorth_features,and patient_hjorth_features, have the shape (num_epochs, 38). Each row in these arrays represents a different epoch, and the 38 columns contain the combined mobility and complexity features for each channel of the EEG signal.

c) Zero Crossing Rate (ZCR):

ZCR counts how many times the EEG signal crosses the zero line. This shows how much the signal goes up and down. A high ZCR means the signal changes quickly, like during bursts of brain activity. A low ZCR means the signal is smoother and changes less, like when the brain is resting or calm. This is important for finding problems with thinking or memory because the brain’s patterns might look different from normal.

These features help us understand EEG signals better by looking at how they work. They make it easier to find small but important changes in the signals that might show problems with thinking or memory. This helps with studying and finding out what might be wrong. It helps identify sharp waveforms, bursts of activity, or the onset and offset of specific events in the EEG signal [22]. The ZCR evaluation is shown in Eq. (4).

$Z C R=\frac{1}{N-1} \sum_{i=1}^{N-1} \begin{cases}1 & \text { if } x[i] \cdot x[i-1]<0 \\ 0 & \text { otherwise }\end{cases}$     (4)

where, $x[i]$ is the sample at index i in the signal, $x[i-1]$ is the sample immediately preceding $x[i] . x[i] . x[i-1]<0$ checks if the product of $x[i]$ and $x[i-1]$ is less than zero, indicating a sign change (crossing). The summation runs over all pairs of consecutive samples, and 1 is added to the sum if a sign change is detected; otherwise, 0 is added.

The input to the zcr_features function is a 3D array representing EEG data, comprising the number of epochs, channels (electrodes), and data points per channel. The function computes the ZCR value for each epoch and channel separately, Tallying the occurrences when the EEG signal intersects the zero baseline. and dividing it by the total length of the epoch.

As a result, the function extracts a ZCR feature for each epoch and channel in the EEG data. Ultimately, the ZCR features matrix provides valuable information about the rate of zero crossings, with each row representing an epoch and each column representing a specific EEG channel. For the control group, there are 13645 epochs, each containing 19 channels resulting in "Control ZCR features" array has a shape of (13645, 19) and for the patient group, there are 14757 epochs, each containing 19 channels resulting in "Patient ZCR features" array has a shape of (14757, 19).

3.2.2 Deep learning techniques

To improve the effectiveness of MCI detection, more advanced machine learning techniques have to be utilized. Specifically, 1D deep learning models like CNN and CRNNs, along with hybrid models such as Population Based CNN (PBCNN) and Population Based CRNN (PBCRNN), have shown remarkable capabilities in capturing spatial and temporal dependencies within EEG data.

a) 1D Convolutional Neural Network:

1D CNN is a popular neural-network architecture which is used for various signal processing tasks, including EEG signal analysis [23]. The temporal aspect of EEG data is crucial for understanding brain activity, and 1D CNN is well-suited for handling this dimension. Furthermore, the network is able to extract both local and global features of the EEG signal, thus making it an effective and robust tool for EEG signal analysis.

2.png

Figure 2. 1D CNN architecture

Figure 2 shows the architecture of 1D CNN which is used for EEG signal analysis. The 1D Convolutional Neural Network is used to study EEG signals and find patterns in the data. The input to the model is like a 3D tensor with parts that include the batch size set to 32 which is how many examples the model looks at one time, epochs set to 30 and features set to 76. The first layer of the model is a 1D convolutional layer that uses 32 filters to capture important details, kernel size of 3 to find small patterns in the EEG signals and the ReLU activation function, which helps the model learn better and faster. After this, the data goes through a 1D max pooling layer with a pool size of 2, which makes the data smaller by keeping only the most important parts, so the model can work more efficiently.

Next, the output from the pooling layer is flattened into a single line, which is sent to the fully connected layers. The first dense layer has 64 units and uses ReLU activation, helping it find bigger patterns in the data. Finally, the model gives one result using a sigmoid activation function, which gives a value between 0 and 1. A value close to 0 means the data is from the healthy control group, and a value close to 1 means it’s from the mild cognitive impairment group. The model learns using binary cross-entropy loss, which is used for binary classification, and the Adam optimizer, which helps the model learn quickly and accurately. This design makes the CNN good at studying EEG data and classifying it correctly.

b) 1D Convolutional Recurrent Neural Network (1D CRNN)

The CRNN architecture is useful for tasks that involve sequential data with long-term dependencies, such as speech, signal, and language processing [24]. The convolutional layers can identify spatial patterns present in the input data, including the frequency features within a signal., while the recurrent layers can capture the temporal patterns. The 1D CRNN architecture is shown in Figure 3.

The model starts with a normal 1D convolutional layer which has 64 filters, kernel size of 3 in order to capture the basic and frequency characteristics of the EEG signals and a ReLU activation function which assist in the detection of complex patterns. Then max pooling to applied to convolutional layer with pool size 2 through which the data size can be reduced while keeping the most important information from the previous layer. Then the model has two LSTM layers each with 64 units each. These layers help in comprehending the variations in the data at a certain interval and identifying trends in the data over a long period. Following the LSTM layers the data is supplied to a dense layer with 64 units with ReLU activation function. This layer assists in enhancing the features identified in the prior layers. Last, the model contains one output layer with one unit and uses the sigmoid activation function. This layer outputs a value between 0 with and binary 1 cross for entropy the loss classification function of and the optimizer signal Adam as is control used or which patient. helps the in-model training is the trained model faster and more accurately.

3.png

Figure 3. 1D CRNN architecture

3.2.3 Hybrid deep optimized techniques

a) Population Based Training (PBT)

PBT is an optimization technique within machine learning and deep learning that aims to optimize hyperparameter tuning efficiency and elevate the performance of trained models. It utilizes the principles of genetic algorithms to iteratively evolve a population of models over multiple training epochs. The PBT process typically involves the steps like- Population Initialization, Model Evaluation, Exploitation and Exploration, Hyperparameter Exploration, Population Update. By dynamically allocating more resources (e.g., computation time, memory) to promising models and discarding fewer promising ones early in the training process, PBT converges to better hyperparameter settings more efficiently than traditional hyperparameter optimization methods like grid or random search. This makes PBT particularly advantageous when training resource-intensive deep learning models. PBT has gained popularity for its ability to efficiently explore the hyperparameter space and find optimal configurations, enabling faster and more effective training of machine learning models [25].

Population Based Training (PBT) algorithm

(1) Start with a population of 5 models and choose the hyperparameters of each model at random. Filters are selected as any number between 32 and 128, the kernel size is chosen from the set of 2 to 5 and the number of dense units can also be between 32 and 128. Make the weights of each model to be any random value.

(2) For each model in the population, fine tune each model for every epoch using the training data set.

(3) To determine the validation accuracy of each model, test each model on the validation dataset.

(4) Sort the models in the decreasing order of validation accuracy and retain the best models.

(5) Select hyperparameters 20% of the population. Modify hyperparameters by changing kernel size, filter size and dense units for the selected models within the specified ranges.

(6) To transfer the hyperparameters and weights from the top performing models to the low performing models in order to enhance the whole population.

(7) Repeat the training, evaluating and updating the models for a total of 10 generations.

(8) Choose and return the model with the best validation accuracy from the last generation.

b) 1D PBCNN:

In the hybrid model, the 1D CNN serves as the core architecture for processing one-dimensional sequential data. The PBT algorithm is integrated with 1D CNN model in the training process to optimize its hyperparameters effectively. During the training process, PBT dynamically adjusts hyperparameters depending on the evaluation of different models in the population, allowing the models to explore the hyperparameter space efficiently. This integration of PBT with the 1D CNN model results in faster convergence to optimal hyperparameter settings, ultimately leading to a more accurate and efficient model and its architecture is shown in Figure 4.

4.png

Figure 4. Architecture of 1D PBCNN

c) 1D PBCRNN (Population Based Convolutional Recurrent Neural Network)

In the hybrid model, the1D CRNN plays a central role in processing sequential data, enabling it to capture both local and long-range patterns present in the input. To efficiently optimize its performance, the Population-Based Training (PBT) algorithm is seamlessly integrated into the training process of the 1D CRNN model. PBT effectively manages a diverse population of 1D CRNN models, each with distinct hyperparameter configurations. During training, these models undergo evaluation on a validation set, and the best-performing models serve as a foundation for exploitation. Through this exploitation process, their hyperparameters are dynamically updated to potentially improve their performance. This cycle of evaluation, exploitation, and exploration continues over multiple epochs. By dynamically updating the hyperparameters based on the models' performance within the population, PBT allows the hybrid model to explore the hyperparameter space more efficiently. Consequently, this optimization approach empowers the hybrid 1D CRNN model to discover more effective hyperparameter configurations, ultimately leading to a more accurate and efficient model for processing sequential data. The architecture of 1D PBCRNN is shown in Figure 5.

5.png

Figure 5. Architecture of 1D PBCRNN