Dual-Layer Ranking Feature Selection Method Based on Statistical Formula for Driver Fatigue Detection of EMG Signals

The basic framework of the study is shown in Figure 1. The collected data will be pre-processed to remove unwanted noise before performing feature extractions. In the feature extraction step, filtered data will be applied to twenty-five time-domain features equations to acquire the feature values. Meanwhile, for the frequency-domain feature values, the filtered data is first transformed to the frequency domain before applying the nine frequency-domain feature equations. Then, in the first layer FS the features will be ranked using 21 FS filters selected in three groups; time-domain feature group, frequency-domain feature group and combined time and frequency-domain feature group. In the second layer, a new rank based on statistical formula was calculated from the 21 ranks. Then, the features will be classified using all the ranks calculated. The classification performance of all available ranks in first and second layer FS will be compared to select the best subset of features based on the highest classification accuracy and minimum number of features.

2.1 Experimental setup

The data was collected from a driving simulator that was set in the lab. The simulator was comprised of a monitor, a computer with an installed driving game (Need for Speed), a steering wheel with pedal and EMG sensors that would be attached to the biceps brachii on the right arm of the participants. Figure 2 shows the experimental setup of the driving simulator. The participants were required to sleep for seven hours on the night before the experiment and avoid caffeinated drinks within 24 hours before the experiment. A total of 15 participants were recruited (seven males and eight females) with an average age of 23 ± 3 years. All participants filled out the Fatigue Assessment Scale (FAS) [19], which was composed of ten questions to evaluate the participants’ fatigue levels before the driving simulation started. The FAS score was later used to classify the drivers into two classes—fatigue and non-fatigue—in the feature selection and classification stage. Seven out of fifteen samples were labelled fatigue and eight samples were labelled as non-fatigue Personal information such as name, age, driving experience and health were also collected. The experiment was conducted at 2 pm to induce natural drowsiness [4] and the participants drove for two hours on the driving simulator. The collected data in the form of EMG signals were kept as a database for analysis in MATLAB 2020b®. The procedure of this study has been approved with ethics number IREC 2020-070 from International Islamic University Malaysia’s ethics committee.

1.png

Figure 1. The research framework of dual-layer ranking feature selection

2.png

Figure 2. Experimental setup of the driving simulator

2.2 Signal pre-processing

Approximately three million points were extracted from the Shimmer EMG sensor input. The raw EMG data is first sampled to regenerate the signal in MATLAB. The sampling rate of 512 Hz was chosen in accordance with the hardware requirements of the Shimmer^TM EMG sensor. Then, these data were filtered using a Butterworth bandpass filter between 10 and 20 Hz to remove the unnecessary noise. Figure 3 shows an example of the raw and filtered data. The filtered signal, will be processed further for feature extraction; where for the time-domain features, the filtered signal can be directly applied to the feature calculations. While, for the frequency-domain features, filtered signal must be transformed using FFT.

2.3 Feature extraction

In general, EMG signal characteristics can be classified into three categories: time domain, frequency domain and time-frequency domain. The time-frequency features are mathematically complex, highly dimensional and require pre-condition transformation to extract the features [20]. This study explored the time-domain and frequency-domain features since they are the most commonly used in EMG signal processing due to their robustness and high performance [17]. A modified version of MATLAB 2021(b) code was used for feature extraction [21] and feature selection [22] methods.

2.3.1 Time-domain features

The time-domain features are the features calculated from the raw EMG time-series signals. They are fast and straightforward to be implemented because no signal transformations are involved in the feature extractions [23, 24]. These features are extensively utilised in medical and technical research. The time-domain features treat the signal as stationary, which overcomes the drawback of EMG signals that have non-stationary characteristics [25]. Although these features may be affected by the dynamic movement noise, especially features that involve energy characteristics [26], the high classification performance in low noise environments and low computational complexity made it the extensive features for EMG signal analysis.

In this study, twenty-five time-domain features, as listed in Table 1, were included in the analysis of long-duration EMG signal processing. These proposed features have been widely used in EMG signal processing for hand movement detection [27]. Out of these 25 features, four features had more than one feature due to their essential parameters: absolute temporal moment (TM) has four features, auto-regressive coefficients (AR) have four features, cepstral coefficients (CC) has two features, and v-order (V) has two features. The total time domain for each sample was extracted resulting in 525 features acquired.

3.png

Figure 3. Raw and filtered EMG data for sample no. 8

Table 1. Time-domain features

2.3.2 Frequency-domain features

Frequency-domain features are the raw EMG signals that are transformed into the frequency domain, which have the properties of the spectrum domain [36]. The main frequency-domain feature was the Power spectral density (PSD), which was calculated by applying the Fourier transform of the autocorrelation function to the pre-processed EMG signal. By applying statistical properties to PSD, more frequency-domain features were generated from previous studies, as listed in Table 2. For each sample, a set of nine frequency-domain features were extracted. The total features became twelve due to the additional features from the spectral moment (SM), which had four values. A total of 180 features were acquired.

Table 2. Frequency-domain features

2.4 Feature selection methods

As discussed previously, there was a total of 47 features that were analysed in this paper, which were composed of 35 and 12 features from the time and frequency domains respectively. Feature selection or feature reduction is a method in acquiring the best set of features of the EMG signals. In this study, a total of 705 features were extracted from 15 samples with 47 features. Feature selection aimed to choose the best subset of features with a minimal number of features.

2.4.1 First layer feature selection

Feature selection algorithm can be performed in three ways: filter method, wrapper method and embedded method. This paper adopted the first method, in which filter algorithms were calculated to rank the features and it’s the best classification performance will be selected as the best group of features. Twenty-one filter methods were used to calculate the rankings of features. Table 3 lists down all filter algorithms used in this study. The first 15 ranking methods are the supervised learning ranking algorithm while the last six methods are the unsupervised learning ranking algorithm.

Table 3. Filter feature selection methods

Infinite Latent Feature Selection (ILFS) is a probabilistic algorithm that performs the ranking step by considering all possible subsets of features and bypassing the combinatorial problem [59]. Infinite Feature Selection (InFS) has two steps. The first step is to rank the features without a label, where the score is based on the connecting path of the features on the affinity graph. Then the best features are selected based on the performance of the cross-validated classification model [39]. Similar to InFS, Features Selection via Eigenvector Centrality (ECFS) uses the affinity graph based on the features mutual information, Fisher’s score and maximum standard deviation. The rank is then calculated based on the eigenvector centrality [40]. Minimum redundancy, maximum relevancy (MRMR) is a criterion-driven algorithm that determines the maximum relevance and minimum redundancy of each feature based on the determined criteria [41]. ReliefF is an algorithm that iteratively calculates the rank of random labelled features with the goal of finding the distinction between each neighbouring feature [42]. Mutual Information Feature Selection (MIFS) identifies the mutual information predetermined by the criteria that are established between the value distribution of a feature in its class to rank the feature [43]. Feature Selection via Concave Minimization (FSCM) ranks features based on linear Support Vector Machine (SVM) as the training sets [44]. Neighbourhood Component Analysis Feature Selection (NCA) ranks features based on the highest positive weight calculated for each feature [60]. The Fisher score method ranks the features by calculating the ratio between interclass separation and interclass variance [46]. The Pearson filter is the simplest form of ranking that is performed by computing the correlation of every pair of features for all features [47]. ANOVA calculates the parametric statistical hypothesis score, F-value by comparing two or more samples of data (typically three samples) [48]. Chi-square scores calculate the dependencies between features by calculating the error between the observed and expected values [49]. Spearman’s score is the correlation between a pair of features [50]. T-test score is the statistical difference of features between two labels [51]. By using nearest neighbour graphs, the Laplacian method ranks feature relevance based on location maintenance [52].

The unsupervised FS (UFS) algorithm MCFS, which used the eigen decomposition of a similarity matrix of the features is used for multi cluster data. The best feature subset is chosen using a L1-norm regularizer to approximate the eigenvectors produced from the data's spectral embedding, which induces sparsity [53]. Unsupervised Discriminative Feature Selection (UDFS) identifies the strong discriminant structure by computing maximum inter-class divergence as well as minimizes intraclass divergences and L2,1 norm of the linear classifier coefficient matrix [54]. Learning for Local Learning-based Clustering (LLCFS) adaptively learns the feature structure through clustering by iteratively updating the Laplacian graph to find the feature relevance [55]. Dependence-guided Unsupervised Feature Selection (DGUFS) is a hybrid of FS and clustering, where L2,0 norm equality constraint is used as FS and two guided terms are used to rank features that have the highest dependence raw data, cluster labels and features [57].

A set of 21 feature rankings were generated from the filter methods listed in Table 3. These sets of rankings were calculated for three groups of features: 1) rankings for time-domain features (RTDF), 2) rankings for frequency-domain features (RFDF) and 3) rankings for combined time-domain and frequency-domain features (RCF). These groups of rankings were classified so as to compare their performance to determine the best subset of features. The performance of these grouped rankings was also compared to the performance of the second layer feature selection method.

2.4.2 Second layer feature selection method (statistical rank)

A new set of statistical ranks (SR) was calculated based on the pool of all 21 rankings for all 47 features. Four statistical formulas were considered in this study:

1. Average statistical rank (ASR), where the average rank for each sample was recalculated as the new rank.
2. Median statistical rank (MedSR), where the rank for each sample was sorted and the middle value of the sorted rank was selected as the new rank for the sample.
3. Mode statistical rank (MSR), where the most repeated number that appeared for each sample rank was selected as the new rank for the sample.
4. Variance statistical rank (VSR), where the variance of the rank across each sample was calculated as the new score for the statistical reranking.

To calculate the SR, consider the ranking matrix (RM) in Eq. (1), where m is the number of features and n is the number of rankings. The pseudo-code for calculating ASR and VSR is presented in Algorithm 1 is the pseudocode for calculating the new statistical ranking.

$R M=\left[\begin{array}{cccc}r_{11} & r_{12} & \cdots & r_{1 n} \\ r_{21} & r_{22} & \cdots & r_{2 n} \\ \vdots & \vdots & \cdots & \vdots \\ r_{m 1} & r_{m 2} & \cdots & r_{m n}\end{array}\right]$        (1)

2.5 Classification

All calculated rankings were classified to assess the performance of the selected features. Initially, the classification was performed on one feature according to the rankings and performance and was then recorded. Subsequently, features were added one by one to the classifiers until all features were classified according to their ranking. All performances were recorded and the highest accuracy from the classification was selected as the best subset of features. Seven classifiers were used in this work in search of the best performance of the selected features.

The k-Nearest Neighbor (kNN) is a slow classifier that classifies an input according to the class of the majority of its labelled neighbours' point k. In this study, k was selected to be three [61]. Linear Discriminant Analysis (LDA) identifies and develops a linear separator between two classes, which assigns all labelled data to their classes in such a way that when it is plotted on its labelled axis, the samples are separated according to their classes [62]. Naïve Bayes (NB) is based on the Bayes probability, in which the attribute of each sample is independent of the attributes of other samples. The likelihood of a sample belonging to a particular class is determined by the label assigned to the sample [63]. Decision Tree (DT) generates a decision tree for the provided features based on the labels that have been assigned to each feature, starting with the most significant feature and progressing down to the least significant feature [64]. Support Vector Machine (SVM) is similar in concept to LDA in that both create a hyperplane separating the labelled data according to their classes. The difference is that the LDA hyperplane maximises the classes mean, whereas the SVM hyperplane maximises the margin between the two classes boundaries with the least amount of error [65]. In this study, two kernels were used, namely linear SVM-L and radial basis function SVM-RBF. Random Forest (RF) trains all features using a decision tree and then classifies the samples to minimize the variance [66].

Due to the limited number of samples, the k-fold cross-validation was employed on each classifier model to ensure that all features were thoroughly utilised as training and testing data. k=10 is selected by dividing the training samples to 90% and 10% samples as the test samples for each k-fold. The classification process was repeated ten times and the average accuracy for each classifier was recorded. All data were efficiently used as both test and training data with k-fold cross-validation.

In this paper, twenty-one filter methods for feature selection were applied to the EMG raw data for driver fatigue detection and a new statistical rank was proposed. It has been shown that the new statistical rank ASR delivered the best performance as compared to the single filter rank in terms of the least number of the best subset of features. The best subset of features was a combination of frequency-domain and time-domain features. Concurrently, the best single filter method for the EMG signals of driver fatigue, performed with the same classification accuracy compared to ASR, but with more features. The number of features was reduced significantly from 47 to 12 (75% fewer features), which meant less computational time. The proposed model is the potential to the development of a more robust feature filter selection method. In the future, multi validation procedure is desirable for a more accurate drowsiness level validation. Based on the results of the single ranking method, combinations of two or three filter methods might be a promising new algorithm for feature selection.