Dual-modal Physiological Feature Fusion-based Sleep Recognition Using CFS and RF Algorithm

Difficulties in falling asleep (DFA) are the most common form of insomnia^{[1, 2]} and can seriously impact general health and may be accompanied by other medical conditions such as mental health disorders (e.g., depression)^[3], Alzheimer′s disease (the most common cause of dementia)^[4], and Parkinson′s disease (a progressive neurological condition causing brain issues which get worse over time)^[5]. The prevalence of such medical conditions represents a global healthcare crisis and the control and cure of DFA is an important component in the treatment plans.

Significant research efforts have been made globally to address the healthcare crisis including efforts to develop effective pharmacological and behavioral treatment options for DFA. Currently, mild difficulty in falling asleep (MDFA) has often been considered as a transition phase between normal control subjects (NCs) and insomnia symptoms, especially DFA. MDFA refers to a clinical state in which individuals find difficulty in falling asleep (generally occasionally), however because this state does not significantly influence individuals′ daily lives, it is below the threshold clinical criteria for insomnia.

DFA is experienced by approximately 33% in the general population^[6]. Moreover, due to the prevalence of DFA and the distress caused to individuals, new DFA research results^{[7, 8]} continue to emerge on a regular basis. For example, the research results of Gamaldo et al.^[9] demonstrate that self-reporting of DFA may be a unique predictor of cognitive performance. Several studies (see [10]) have also suggested that DFA is negatively associated with cognition performance. Leigh et al.^[11] suggest that DFA, as measured by the simple four-item Nottingham health profile (NHP) questionnaire, could be useful in clinical or nursing home settings as a simple indicator for mortality. Leigh et al.^[11] also show that women (who were troubled sleepers) are at a higher risk of death than untroubled sleepers. Almeida et al.^[12] have observed that DFA increases the risk of depression in older men.

While there have been a large number of studies investigating DFA, few studies have addressed in detail the transition phase of DFA, i.e., MDFA. Since MDFA is a symptom of DFA, early accurate and effective discrimination of MDFA is crucial to warn potential patients and guide them through appropriate treatment plans to delay (or even prevent) the onset of DFA. Traditionally, clinicians diagnose MDFA by judging whether individuals meet research diagnostic criteria (RDC) of MDFA.

The RDC for MDFA, as formulated by an American Academy of Sleep Medicine (AASM) work group, are as follows^[13–15]:

1) Falling asleep takes more than 30 minutes.

2) Complaints of daytime impairment or distress.

3) Difficulty in falling asleep at least 2 times a week.

4) Difficulty in falling asleep for at least 2 months.

As can be seen from the above, clinicians diagnose MDFA are mainly based on their observation results and patient's subjective reports. Therefore, there may be difficulties in reaching an objective and accurate diagnosis for potential MDFA patients.

Previous related research^[16–18] has shown that human biological signals contain considerable information about sleep, pointing to the possibility of recognizing MDFA by studying these biological signals. Usually, these biological signals are mainly include: electroencephalogram (EEG), electrooculogram (EOG), electromyogram (EMG), electrocardiogram (ECG), etc. However, research has often focused on automatic sleep staging (ASS) using physiological signals. For example, the genetic fuzzy inference system based on expert knowledge for ASS has been developed by Liang et al.^[19], where eight features are used. The features (used as input variables) include temporal signals and a spectrum of physiological signals. Moreover, the fuzzy rules and the fuzzy sets were constructed based on expert knowledge. Tsinalis et al.^[20] present a machine learning methodology which is based on time-frequency analysis and stacked sparse auto-encoders for ASS using a single channel of EEG. Sen et al.^[21] found strong mapping connections between EEG signals and each sleep stage using minimum redundancy maximum relevance and t-test algorithm analysis.

Despite the use of physiological signals to auto-identify sleep stage, research results have sprung up. However, the research on automatic recognition of MDFA is an under-reported area in the literature and remains an open research question. Based on this reason, the purpose of this study is to design a classification method to distinguish MDFA and NCs using physiological data. In this research, we studied the correlations between MDFA and physiological data by calculating heuristic “merit” of different subsets based on fusing physiological features to find the optimal feature subset. We then trained a random forest (RF) classifier based on the optimal feature subset and evaluated the classification performance and statistical significance of our proposed method iteratively using 10-fold cross validation. The experimental result demonstrates that correct recognition rate (CRR) for our method is 96.22%. The performance demonstrates an improvement over alternative popular classification algorithms (including K-nearest neighbors, support vector machine, neural network, Bayesian approaches, and classification trees). There are many technical terms used in this paper, for brevity and to improve readability, we list these technical terms and their abbreviations in Table 1.

This paper is organized as follows: Section 2 provides a systematic description of the proposed MDFA recognition method. In Section 3, we describe our experimental design, experimental results, discussion and analysis in detail. Section 4 concludes the work of this paper with consideration of future directions for research. Before ending this introductory section, the main contributions and novelties of this paper are presented as follows.

1) Fills the gaps in the area of MDFA research.

2) Compared with the traditional fixed order auto regression (AR), our proposed adaptive AR model can extract linear candidate features more accurately.

3) Sleep recognition scheme based on dual-modal physiological feature fusion, EEG features can effectively explore information contained in sleep data, moreover, it relies on EOG features to make up for its inadequacies.

4) An optimal sample subset containing 16 features is obtained to identify MDFA.

5) The study found that slow wave sleep (SWS) stage and rapid eye movement (REM) stages are the most important for the recognition of MDFAs.

6) The experimental results of our method demonstrate improved performance, the highest recognition rate for MDFAs reaches 96.22%.

Raw EEG and EOG data cannot directly be used for research, therefore it is necessary to preprocess the original data.

The flowchart of MDFA recognition method is depicted graphically in Fig. 1. The method includes four principal functions: 1) data preprocessing, 2) candidate feature extraction, 3) dual-modal physiological feature fusion, 4) correlation-based feature selection (CFS) (exploring the optimal feature subset), 5) RF classification. The following is a detailed description of the above five functions.

Figure 1.  Schematic diagram of our proposed method

Raw physiological data represents a mixture of “signal” and “noise”. Data decomposition methods can help separate signal from noise data and disentangle overlapping patterns. Usually, the information of physiological signals with respect to sleep is found at frequencies between 0.5 Hz and 30 Hz^[22]. All signals with frequency above 30 Hz, or below 0.5 Hz need to be removed. Thus, the authors have chosen a Butterworth bandpass filter with 0.5 – 30 Hz passband to eliminate EEG and EOG signal drifting and to remove all invalid frequencies.

Once invalid physiological signals are filtered (removed), linear and nonlinear analysis is performed on the physiological signals for candidate feature extraction.

1) Linear analysis extracts linear candidate features

a) Frequency-domain analysis

In this study, power spectral features are obtained by adaptive auto regression (AR) modeling. The AR model is a parametric model used to describe a stationary time series. Physiological signals are typical non-stationary, thus, we adopted an adaptive AR modeling in this work. AR model used in physiological signals processing is not novel, but previous studies have used fixed order AR modeling in contrast to our adaptive AR modeling^[23].

Traditionally, fixed order AR modeling parameters can be used to determine physiological signals states. Fixed order AR modeling represents the current signal x(t) as the weighted sum of its previous values x(t – i) and the uncorrelated error ε(t) as shown in (1):

where a_i represents the AR coefficients and p is a fixed order of model.

In this study, the authors use Akaikes information criterion (AIC)^[24] to adaptively get the best order p of AR model in each 30 s epoch. This is so that the power spectral ranges of different frequency bands can be divided more accurately. In other words, the range of delta (0.5 – 2 Hz), sawtooth (2 – 4 Hz), theta (4 – 8 Hz), alpha (8 – 12 Hz), spindle (12 – 14 Hz), and beta (14 – 30 Hz) are accurately divided. The calculating process of best order p is shown in Algorithm 1. The implementation is achieved using Matlab R2017a.

Algorithm 1. Process of finding the best order p.

Input: Sample size N; Input signal vector x, i.e., equal interval data column.

for p = [1 : floor(sqrt(N))]

　[a, e] = aryule(x, p) /* e is the variance parameter */

AIC (p) = log(e)+$2 \times\dfrac{p}{N}$

end

[～, best p] = min(AIC) /* bestp is the best order p */

Output: best p

Finally, candidate features are obtained using frequency-domain analysis. The candidate features include: i) the absolute power of alpha, beta, theta, delta, spindle and sawtooth, ii) the relative spectral power of alpha, beta, theta, delta, spindle and sawtooth, iii) the center frequency of alpha, beta, theta, delta, spindle and sawtooth, iv) the maximum power of alpha, beta, theta, delta, spindle and sawtooth, and v) the absolute ratio of beta power to delta power, alpha power to beta power, alpha power to spindle power, theta power to alpha power, delta power to theta power, delta power to alpha power, delta power to spindle power, spindle power to beta power.

b) Time-domain analysis

Time domain analysis is generally used in the analysis of discrete signals. The majority of the commonly used time-domain analyses are grounded in probabilistic analysis of random signals^[25]. Therefore, the authors also adopt this signal analysis method to extract candidate features such as average amplitude, variance, skewness, kurtosis. In this work, time-domain analysis method is also used to extract Hjorth parameters^[26]: activity, mobility, and complexity. One of the original aims of this parameter is to solve sleep problem, such as sleep staging. Therefore, Hjorth parameters are also calculated in this study.

2) Nonlinear dynamics analysis extracts nonlinear candidate features

Linear analysis is the principal method used to extract candidate features. However, to raise the recognition accuracy of MDFAs, nonlinear dynamics analysis is a complementary method of extracting candidate features. Such features reflect the association between sleep activity and physiological signals. In this research, nonlinear candidate features include: C0-complexity, Shannon entropy, spectral entropy, etc.

Following data preprocessing, feature subset₁ is obtained from dual-channel EEG, a total of 90 features (including 78 linear candidate features and 12 nonlinear candidate features) are identified. Meanwhile, feature subset₂ is obtained from one channel EOG, a total of 45 features (including 39 linear candidate features and 6 nonlinear candidate features) are identified. Next, all extracted candidate features are integrated into a composite feature set, which is the fusion of subset₁ and subset₂. Meanwhile using a feature search method called bests first, a ranked feature list (RFL) $\left\{ {{f_1},{f_2},{f_3}, \cdots ,{f_{144}},{f_{145}}} \right\}$ is generated based on correlations between individual features and classes (i.e., MDFAs class or NCs class) from high to low. Every sample is organized in a table, where each sample is described as 135 feature attributes, and 1 class attribute.

Generally, there are a large number of EEG and EOG features, some of which have a negative effect on the evaluation. Using a classification algorithm to select a subset of features that best reflect different class attributes (i.e., MDFAs class or NCs class) requires the testing of all possible combination of features^{[27, 28]}. However, this process is a very computationally expensive and time-consuming task, e.g., in this study to find the best combinations of 145 features it would need to try 2¹⁴⁵ combinations. To avoid combinatorial explosion, the authors propose a correlation-based feature selection (CFS) approach to find optimal feature subset.

CFS^[29] is a simple filtering algorithm that ranks feature subsets according to a correlation-based heuristic evaluation function. The evaluation function bias is toward subsets that contain features that are highly correlated with the class and uncorrelated with each other. Irrelevant features should be ignored because they will have low correlation with the class. Redundant features should be screened out as they will be highly correlated with one or more of the remaining features. In this study, an optimal feature subset of 145 features is identified using the following steps (see Fig. 1) which is a schematic diagram illustrating the proposed CFS method.

Step 1. Calculate r_cf: the correlation of every feature-class, and r_ff: the inter-correlation of feature-feature.

Step 2. The feature subsets $\left\{ {{f_1}} \right\}\!,\;\left\{ {{f_1},{f_2}} \right\}\!,\;\left\{ {{f_1},{f_2},{f_3}} \right\}, $$\cdots\!,\;\left\{ {{f_1},{f_2}, \cdots ,{f_{145}}} \right\}$ are then tested by (2):

where M_S is the heuristic “merit” of a feature subset that is containing k features, k$ =1,2,\cdots,145$. ${{\bar r}_{cf}}$ is the mean feature-class correlation. ${{\bar r}_{ff}}$ is the average feature-feature inter-correlation.

Step 3. From the results obtained in the previous step, the feature subset with the highest M_S is selected. Selected feature subsets and corresponding class attributes constitute an optimal feature subset.

The optimal feature subset will be used to reduce the feature dimension. Furthermore, it will reduce the number of features in MDFA′s recognition process.

Following the identification of the optimal feature subset, an inference mechanism is performed in this subset to recognize MDFAs. Generally, to ensure the accuracy, efficiency and simplicity of the algorithm, the classical algorithm and development kit should be chosen wherever possible. Based on these design considerations, the authors chose random forest (RF) algorithm^[30], to recognize MDFAs, and the motivation of this choice can be summarized as follows:

1) The RF can be run large datasets and shows a strong robustness.

2) The RF has an effective method for estimating missing data and maintains accuracy when a part of the data is missing.

3) RF is one of the preeminent inductive inference algorithms with the rules usually expressed by an appropriate IF-THEN logic statement.

4) Many free open-source tool packages exist such as Matlab2Weka toolbox^[31].

In this study, the Matlab2Weka toolbox is used to implement the RF algorithm with 10-fold cross validation. Additionally, to obtain a statistically meaningful result, the number of iterations is set to 10 (see Fig. 1 data partitioning). This means that the RF classifier is called 100 times to train and test the optimal feature subset.

The RF algorithm is an ensemble of a multitude of classification decision trees and it has been widely used in many fields including sleep research^[32]. It is a type of inductive statistical classification model and measures the relationship between a categorical dependent variable and one or more independent variables using an IF-THEN logic inference. In our method, we used RF to explore the best inference model to describe the relationship between MDFAs individuals and the optimal feature subset. By putting the input optimal feature subset down to each decision tree in RF, each tree will give an inference result. The final result is the most popular class by voting from all trees. In this research, the RF algorithm is executed in three steps as follows:

Step 1. Ten bootstrap samples are extracted from the optimal feature subset using sampling with replacement.

Step 2. Create a classification decision tree model for each bootstrap sample.

Start from the root node of each tree, and split bootstrap sample into different subsets according to the node attribute. The selection of node attribute and splitting criterion are based on information gain (IG) of the node attribute. The IG of splitting one data set S into subsets S_i can be defined as

In the above (3), c denotes the number of classes (here c = 2 possible test state: NCs and MDFAs) and E(S_i) is the information entropy of the subset S_i. It is calculated as

where p_i is the proportion of test state i in the subset S_i.

Each attribute is applied to calculate the IG, and the attribute with the highest IG is selected as the root node. This process is recursively repeated at each branch node until either every attribute is selected, or this process reaches a leaf node that is a test state output.

Step 3. Output the results: Step 2 is repeated to build 10 trees, and the final classification output is the most frequently occurring test state out of the 10 output trees.

Fig. 2 shows a simplified example of partial RF generating process. The example has omitted a large number of information. However, it provides a clear overview and understanding of the inference rules. A rule, with its IF-THEN structure, defines a basic fact about subject′s current test state. For example, an inference rule generation process shown as the bold in Fig. 2 can be described as the following IF-THEN structure:

Figure 2.  A simplified example for partial RF generating process based on rules

[IF “Fpz-Cz_Skew” < 0.09

THEN IF Fpz-Cz_Center_Frequency_Sawthooth” ≥ 2.04

THEN Subject = <MDFAs>]

The population (the subjects used in this research) were selected from the Sleep-EDF database [EXPANDED]^{[33, 34]} which contains 61 data recordings taken from 42 Caucasian subjects. This database includes dual-channel EEG [Fpz-Cz and Pz-Oz] and one horizontal EOG with all samples being at 100 Hz. In addition, samples also contain other physiological signals and peripheral data. Data used in this research includes 39 recordings from 20 NCs, and 22 recordings from 22 MDFAs. Among them, the NCs age ranged from 25 to 34 years old (male/female=10/10), and the MDFAs age ranged from 18 to 79 (male/female=7/15). For all the sleep physiological signals, EEG is the most important, because it originates from the central nervous system of the brain and can objectively assess the relationship between brain activity and sleep. Meanwhile, different sleep stages are directly reflected by different EEG bands. For example, slow wave sleep (SWS) is dominated by delta wave, non-rapid eye movement (NREM) sleep stage 1 (NREM1) is dominated by theta wave. Except EEG, the secondary is EOG. Therefore, this study only uses EEG and EOG.

Typically, the sleep data sample period is divided into 20 s or 30 s^{[35, 36]}, in this study the sample period is fixed at 30 s. To ensure the rationality of experimental data, the sample numbers of MDFAs and NCs were almost equal. Each data record of MDFAs includes 540 samples while NCs includes 300 samples, i.e., the total number of MDFAs and NCs were 22×540 and 39×300, respectively. Statistical results are shown as Table 2.

Table 3 shows the results of the optimal feature subset, including importance level, salient channel, and feature. From Table 3, it can be seen that:

1) The optimal feature subset contains 16 features, and the most effective feature to recognize MDFAs was derived from linear analysis methods, especially frequency-domain analysis. This conclusion indicates that linear analysis method was useful for understanding the potential sleep mechanisms of MDFAs.

2) There was a relationship between 16 features and different frequency waves as shown in Fig. 3. As shown, the features associated with delta account for 25%, while the features associated with sawtooth, beta, alpha and theta account for 56%. It is known by R&K rules^[37] or new guidelines developed by the American Academy of Sleep Medicine (AASM)^[38] that slow wave sleep (SWS) was determined by delta, while rapid eye movement (REM) are determined by sawtooth, beta, alpha and theta. Thus, the authors concluded that SWS and REM stages were the most important for the recognition of MDFAs.

Figure 3.  Relationship between 16 features and different frequency waves

3) Some of the most effective features are derived mainly from Fpz-Cz, it indicates that the frontal area was related to MDFAs recognition.

In this section, a comparative analysis was conducted between the CFS+RF mechanism and other alternative strategies. Table 4 shows the results of performance comparison of different strategies for recognizing MDFAs with 10-fold cross validation. As can be seen from Table 4, to explore the optimal feature subset, we chose four typical feature selection algorithms: CFS, relief^[39], information gain, and gain ratio^[40].

Fig. 4 shows the results of feature numbers generated by feature selection methods. From the results, we can conclude that:

Figure 4.  Feature numbers generated by feature selection methods

1) CFS generates fewer features, but attains better CRR than other alternative feature selection methods. Especially CFS+RF strategy can achieve the best result.

2) The number of features generated by Information gain and Gain ratio are very close, additionally, features are basically the same. In our opinion, this result may be due to the same of basic principle of its feature production.

To find the highest CRR, we conducted a comparative analysis between our proposed approach and alternative popular classification algorithms: a) RF, b) K-nearest neighbor (KNN)^[41], c) support vector machine (SVM)^[42], d) neural network (NN)^[43], e) Bayes^[44], and f) classification tree (CR)^[45]. The results can be seen in Table 4 which shows the optimal feature subset [according to CFS and CRR] in the first column. Matlab2Weka toolbox is used to accomplish these algorithms, the main parameter settings are shown in Table 5. To better evaluate the superiority/inferiority of each classification mechanism, we also computed the following statistical measures: a) sensitivity, also called the true positive rate (TPR) or recall in other fields such as information retrieval, b) 1-specificity, also called the true negative rate (TNR), c) Precision, d) F-measure, and e) the Kappa statistic. Space restricts a detailed discussion on the five statistical measures, however a detailed introduction may be found in [46, 47]. The specific experimental comparative results are shown in Fig. 5.

Figure 5.  Specific experimental result of five classifiers base on the optimal feature subset

Usually for five statistical metrics (see Fig. 5), we expect that better recognition performance is indicated by higher values for TPR, precision, F-measure, and Kappa statistic, and a lower value for TNR. The Kappa statistic (which measures the agreement of prediction with the true class attribute) is expected to have a value close to 1.0. As anticipated, our reported experimental result supports the prediction that it achieves a higher TPR, precision, F-measure, and Kappa statistic while using RF algorithm, it also has a lower TNR than other six classification algorithms. The value of Kappa statistic when using RF is 0.92 which is the most closer to 1.0 than the alternative method.

It is hard to accurately compare the performance differences of various recognition methods due to the differences of datasets used in various methods. However, by analyzing and comparing various methods using different datasets, it can indirectly reflect the advantages and disadvantages of different methods. Table 6 lists a comparison of three aspects of some existing methods. In [48, 49], EEG data was used to perform sleep recognition based on deep learning and deep neural networks, respectively. As you can see in Table 6, the accuracy of the above two schemes is lower than that of our proposed method. Moreover, the Kappa statistics in the literature^[48] is also lower than that of our proposed method. We think that this result may be because the feature fusion strategy of EEG and EOG is adopted in this study. EOG features further explore the information contained in sleep data and make up for the inadequacies in EEG features.

As discussed above, the CRR derived from the dual-modal optimal feature subset was high and encouraging, especially the accuracy of the RF algorithm which can reach 96.22%. The study has achieved a number of research goals related to recognition of MDFA in a laboratory environment. The next logical stage in our research is the implementation of MDFA in “real-world” environments. Future research will address the development of an MDFA recognition system capable of implementation in the “real-world” to meet the requirements of “real-time” high CRR.

In terms of the real-time requirements, there is basically one factor which must be taken into consideration. To realize real-time physiological data recording and processing, the complexity of all aspects of the data processing must be kept to a minimum. If data collection and transmission are too time-consuming, the time delays will not conform to the real-time requirements, this will cause the real-time system to be essentially meaningless. Hence, this system needs to adopt a wearable and non-invasive data collection device, and real-time data transmission. We will improve our three recording electrodes belt^[50] to meet this requirement. This belt (see Fig. 6) is developed in our previous study, and its original main function is to collect and send raw EEG data to a computer.

Figure 6.  Prototype of three recording electrodes belt

Given the developments in the power and memory capacity of mobile devices, such processing (at least pre-processing) may be implemented in body area networks with a high powered mobile device acting as a local server and communications device operating over Wi-Fi and mobile systems (e.g., 3G, 4G, and in future 5G) to maintain connectivity in robust systems incorporating redundancy.

Our goal in the future is to develop an automatic identification system of MDFA, physiological data being collected from a dual-channel EEG and one channel EOG using improved three recording electrodes belt. Furthermore, dual-modal optimal feature subset and RF algorithm will be configured in a system designed to recognize MDFAs.

The purpose of this study is to explore the correlations between physiological signals and MDFA, the goal being automatic recognition of MDFAs from NCs. In this research, an optimal feature subset containing 16 features is identified. The research outcomes are encouraging as our proposed method obtained a high CRR which reached 96.22%, a TPR of 95.31%, a TNR of 2.90%, a precision of 97.10%, a F-Measure of 96.20%, and a Kappa statistic of 0.92.

From the experimental results, we have found that a certain relationship exists between an optimal feature subset and different frequency waves (see Fig. 3). The authors therefore suggest that the SWS stage and REM stage are the most important for the recognition of MDFAs. In future research, we aim to develop a real-time decision support system (DSS) based on current research to assist clinicians in MDFA diagnosis and recognition. We propose that such a DSS will be potentially extremely useful in recognizing MDFA in “real-world” local and remote diagnostic situations.

This work has been supported by National Natural Science Foundation of China (Nos. 61761027 and 61461025), the Yong Scholar Fund of Lanzhou Jiaotong University (No. 2016004) and the Teaching Reform Project of Lanzhou Jiaotong University (No. JGY201841).