A real-time wireless wearable electroencephalography system based on Support Vector Machine for encephalopathy daily monitoring

Abstract

Wearable electroencephalography systems of out-of-hospital can both provide complementary recordings and offer several benefits over long-term monitoring. However, several limitations were present in these new-born systems, for example, uncomfortable for wearing, inconvenient for retrieving the recordings by patients themselves, unable to timely provide accurate classification, and early warning information. Therefore, we proposed a wireless wearable electroencephalography system for encephalopathy daily monitoring, named as Brain-Health, which focused on the following three points: (a) the monitoring device integrated with electroencephalography acquisition sensors, signal processing chip, and Bluetooth, attached to a sport hat or elastic headband; (b) the mobile terminal with dedicated application, which is not only for continuous recording and displaying electroencephalography signal but also for early warning in real time; and (c) the encephalopathy’s classification algorithm based on intelligent Support Vector Machine, which is used in a new application of wearable electroencephalography for encephalopathy daily monitoring. The results showed a high mean accuracy of 91.79% and 93.89% in two types of classification for encephalopathy. In conclusion, good performance of our Brain-Health system indicated the feasibility and effectiveness for encephalopathy daily monitoring and patients’ health self-management.

Keywords

Wearable system wireless electroencephalography encephalopathy Support Vector Machine daily monitoring

Introduction

Encephalopathy is proved to be one of the major diseases threatening human health in recent years.¹ Electroencephalography (EEG), near-infrared spectroscopy (NIRS), magnetic resonance imaging (MRI), functional magnetic resonance imaging (fMRI), positron emission tomography (PET), magnetoencephalography (MEG), and other advanced modalities are utilized separately or jointly for encephalopathy monitoring, medical diagnosis, and functional rehabilitation.^2–5 The EEG is the only non-invasive measure for neuronal function of the brain, and kinds of encephalopathy can be assessed by using EEG.⁶ Generally, current monitoring by EEG devices is inpatient due to the huge volume.^7–9 Therefore, when patients are out of hospital, their rehabilitation conditions cannot be tracked real-timely. Fortunately, wearable solutions were proposed by several researchers.^10,11 Patients can undergo ambulatory monitoring anywhere with a wearable device.^12,13 Moreover, the continuous recordings obtained from these outpatient wearable devices are beneficial for monitoring and rehabilitation after diagnosis.¹ These devices will become a potential diagnostic tool in the near future. However, it still cannot meet the requirements of daily monitoring and patients’ health self-management depending only on such a device. The main reasons can be concluded as follows: (a) lacking of the life-oriented device which is easy to be accepted in public; (b) the patients cannot understand the meanings of the recordings without the help of doctors and indicators, for example, a yes/no outcomes for warning, which are easily understood and what they need deeply; and (c) for healthcare big data, it is better to collect and store the recordings in a cloud terminal. As a result, in this article, we concentrate on developing a wireless wearable EEG system, called Brain-Health, to solve these problems.

The main contributions of our Brain-Health system are as follows: (a) the monitoring device for EEG data collecting, including acquisition sensors, signal processing chip, and Bluetooth, attached to a sport hat or elastic headband; (b) the mobile terminal with dedicated application (APP) for EEG continuous recording, displaying, and real-time monitoring; and (c) the classification algorithm applied to encephalopathy in clinical for early warning based on intelligent Support Vector Machine (SVM). The monitoring is crucial in many clinical applications of encephalopathy.¹⁴

In order to evaluate the performance of the Brain-Health system, we collect three groups of EEG data: comatose patients of hepatic encephalopathy (HE, Group 1), rehabilitated patients of HE (Group 2), and normal controls (Group 3). The results prove the fact that HE patients are correlated with abnormal EEG and their basic rhythms gradually become slow.^15–17 Furthermore, our Brain-Health system achieved the high accuracy of about 91.79% and 93.89% corresponding to the classification of Group 1–Group 2 and Group 1–Group 3 based on SVM. In conclusion, the good performance and high accuracy indicated the feasibility and effectiveness of our Brain-Health system for encephalopathy daily monitoring, health self-management, and clinical medical research.

Methods

There are two key sections in this part: participants selecting (section “Participants”) and Brain-Health monitoring system developing (section “Brain-Health system”), including the overview of the system (section “Overview of the system”), EEG acquisition device (section “EEG acquisition device”), mobile APP (section “Mobile application (APP) for daily monitoring”), and classification algorithm based on SVM (section “Classification algorithm based on SVM”).

Participants

In total, 15 participants from the Second Hospital of Nanjing are selected in this study: five HE comatose patients (Glasgow Coma Scale (GCS) score: 3–8 points, mean age: 41.8 ± 11.5), five HE rehabilitated patients (GCS score: 14–15 points, mean age: 42.8 ± 14.3), and five normal controls (GCS score: 15 points, mean age: 38.2 ± 12.1, no physical or mental problem), where the GCS score is a standard generally applied for evaluating the state of cognition in clinical, and the higher the score, the better the state of cognition (full score: 15 points). This research is approved by the Institutional Review Board (IRB) and Ethical Committee of the Second Hospital of Nanjing. Written informed consents are obtained from all participants.

Brain-Health system

Overview of the system

The overview of Brain-Health monitoring system is shown in Figure 1. The system consists of seven major modules: EEG acquisition, EEG pre-processing, microcontroller, power supply, communication and positioning, mobile terminals, and cloud storage. Briefly, these modules can be categorized into two parts: hardware part (modules a–e) and software part (modules f and g). In addition, an effective classification algorithm is proposed and can be embedded in software part. The key contributions of Brain-Health system are concluded in the “EEG acquisition device,”“Mobile application (APP) for daily monitoring,” and “Classification algorithm based on SVM” sections.

Figure 1.

General architecture of system: (a) EEG acquisition module, (b) EEG pre-processing module, (c) microcontroller module, (d) power supply module, (e) communication and positioning module, (f) mobile terminals module, and (g) cloud storage module.

EEG acquisition device

Two kinds of daily wearing appearance are designed in our EEG device: the sport hat (Figure 2(a) and (c)) and the elastic headband (Figure 2(b) and (d)). Life-oriented appearances both are suitable for indoor and outdoor. We can change circumference flexibly according to the different sizes of the patient’s head in both designs. From Figure 3, the details of the main parts of schematic circuit diagram are introduced as follows:

EEG acquisition: Dry electrodes (coated with Ag/AgCl, located in FP₁, FP₂, and A₁ according to 10–20 system international standard^14,18) are chosen in the system instead of traditional wet electrodes requiring conductive paste and good skin condition.^19,20 Here, the 10–20 system is an internationally recognized method to describe the location of scalp electrodes for an EEG experiment, where the “10” and “20” refer to the fact that the actual distances between adjacent electrodes are either 10% or 20% of the skull.

EEG processing: ThinkGear EEG-processing chip with the sampling frequency of 512 Hz (medical grade accuracy chip produced by NeuroSky Company, San Jose, CA, USA) is adopted for EEG pre-processing, time-frequency transformation, and digitized power spectra calculating. Dry electrodes are known to have micromotion artifact and noisy, and ThinkGear chip will deal with the problem using self-contained noise filtering algorithm.²¹ Besides, the dry electrodes are recommended in guidance manual of ThinkGear. Microcontroller, as a signal control center, is responsible for the EEG re-processing and coordinates the ordered work among modules.

Communication: CC2540 Bluetooth chip (produced by Texas Instruments Company, Dallas, TX) is selected for communication due to its characteristics of small size and ultra low power consumption especially suitable for wearable products.

Power supply: The 3.6 V polymer lithium-ion battery provides power of the whole system, which is commonly used for wearable products because of its small volume, large capacity, and good discharge performance.

Figure 2.

Two kinds of appearances of wearable EEG device: (a) and (b) schematic diagram of appearances, (c) and (d) pictures of appearances.

Figure 3.

Main parts of schematic circuit diagram.

Hardware circuit has the small volume of 4 cm ×2 cm × 1 cm and low current consumption of 44 mA. The above parameters meet design requirements of wearable EEG devices.

Mobile application (APP) for daily monitoring

A dedicated APP for encephalopathy daily monitoring is conducive to record EEG signal timely and help to understand the meanings of the recordings easily by patients themselves (shown in Figure 4). The stream data are sent to the APP via Bluetooth, and the APP interface will show the EEG recording in milliseconds. Meanwhile, the data will be passed to the cloud though the mobile terminal by network. The cloud will run program and return the results to APP, and thus, the warning algorithm and daily monitoring are realized. Many practical functions are developed in main user interface: EEG data recording, history searching, and encephalopathy knowledge. In addition, Emergency Call module is presented for patients calling to relatives and hospital when they suffer from sudden encephalopathy attack.

Figure 4.

The mobile application (APP) interfaces: (a) login interface, (b) main interface, (c) data recording interface, and (d) history searching interface.

Classification algorithm based on SVM

It will be dangerous when a patient falls into encephalopathy attack suddenly. Therefore, an effective classification for timely early warning is very important for encephalopathy daily monitoring. As a result, taking HE as an example, we propose a useful classification algorithm based on SVM. The main reason for choosing the SVM is that this method often provides considerably better classification performance than other algorithms on small data set.

HE

HE is a common encephalopathy; the patient’s clinical manifestation is verified that EEG wave basic rhythm gradually slows down.^15,16 The normalized spectral power of α₁ (8–10 Hz), α₂ (10–13 Hz), β₁ (13–17 Hz), and β₂ (17–30 Hz) will all decrease while δ wave (1–3 Hz) increases. The ratio of α₁/δ, α₂/δ, β₁/δ, and β₂/δ will become even smaller than before. Based on this fact, we propose an early warning algorithm based on SVM by selecting above nine features, including five relative spectral powers and four spectral power ratios: δ, α₁, α₂, β₁, β₂, α₁/δ, α₂/δ, β₁/δ, and β₂/δ.

Main idea behind SVM

Generally speaking, given n points x_i in feature space, each belonging to either of two classes y_i∈{−1, 1}. The main idea behind SVMs is to construct an optimal hyperplane as the decision surface, so as to maximize the separation between examples in the two classes. This can be transformed into the following convex optimization problem

\begin{matrix} \underset{α}{Minimize} \frac{1}{2} α^{T} Q α - e^{T} α \\ Subject to y^{T} α = 0 0 \leq α_{i} \leq C, i = 1, . . ., n \end{matrix}

(1)

where $e = [1, \dots, 1]^{T}$ is the vector of all ones, $C > 0$ is the upper bound on the error, and Q is n × n positive semi-definite matrix, defined as Q_ij = y_iy_j K(x_i, x_j), where K(x_i, x_j) is a kernel function. In our experiment, a Gaussian radial basis function is used as kernel, where γ is shape parameter

K (x_{i}, x_{j}) = \exp (- γ {‖ x_{i} - x_{j} ‖}^{2})

(2)

Criteria of performance evaluation

Matthews correlation coefficient (MCC) is an important criterion, which is used commonly to assess a classifier’s performance. It can be calculated as

MCC = \frac{P_{11} P_{00} - P_{10} P_{01}}{\sqrt{(P_{11} + P_{01}) (P_{11} + P_{10}) (P_{00} + P_{10}) (P_{00} + P_{01})}}

(3)

It can be seen from formula (3) that the MCC values will fall into [−1,+1]. Note that the higher the value, the better the classifier’s performance. MCC = 0 implies the predictive ability of a classifier equivalent to random guess, where the P₀₀, P₀₁, P₁₀, and P₁₁ are defined by a confusion matrix as shown in Table 1.

Table 1.

A confusion matrix.

	Actually true	Actually false
Predicted true	P₁₁	P₁₀
Predicted false	P₀₁	P₀₀

During receiver operating characteristic (ROC) curve analysis, we can find the relationship of sensitivity (or true positive rate) and false positive rate (or 1-specificity). Generally, the ROC curve of random classifier is used as a baseline. The further the ROC curve from the baseline, the classifier shows the better the prediction performance. A good classification model should be as close as possible to the upper left corner or coordinate (0, 1). Here, the random classifier adopts the strategy of random guessing, which is randomly classified that half of the samples are positive and the other half are negative.

The area under the curve (AUC) indicates the area under the ROC curve and has been accepted as the standard measure for assessing the accuracy of a classification model. The values of AUC fall into [0, +1]; a larger value indicates better performance of classifier. Typically, the value is between 0.5 and +1.0, where 0.5 corresponds to the random classifier and +1.0 to the perfect classifier with 100% accuracy.

Details of classification algorithm

In our work, the details for our classification algorithm for early warning based on SVM are concluded as follows:

Sampling frequency is 512 Hz of our wearable device and every feature can be calculated per second based on the corresponding 512 EEG recordings. Due to the EEG is a kind of non-stationary signal, each sampling is identical independent with each other. Hence, in order to expand the database, one sample can be selected randomly from every continuous 10 samples as a database (Uniform Distribution is adopted, for that every value is chosen with the same possibility).

Due to the imbalance of the data size in different groups, we randomly select 100 samples from each volunteer and divide them into five sets, where the four sets are for training and the rest one for testing. Note that a total of 15 volunteers and 100 × 15 = 1500 samples are contained in the database from all volunteers including 1200 samples as training and the remaining 300 as testing.

Fivefold cross-validation is used in the training phase to make the results more reliable. The cost value c and the parameter g in SVM are both selected from the set {2⁻¹⁰, 2⁻⁹, 2⁻⁸, …, 2⁸, 2⁹, 2¹⁰}.

Six important criteria are calculated for performance evaluation in testing group: accuracy, sensitivity, specificity, ROC curve, AUC, and MCC.

The main pseudocode of classification algorithm based on the SVM is presented in Algorithm 1.

Algorithm 1. Pseudocode of classification algorithm based on SVM.
Algorithm Support Vector Machine (SVM) Input: (1) D_feature ={ (α₁₁, α₁₂,…, α₁₉), (α₂₁, α₂₂,…, α₂₉),…, (α_n1, α_n2,…, α_n9)}// feature matrix space (2)Label_real = (C_r1, C_r2,…, C_{r n}) // Real label corresponding to Feature matrix // $C_{ri} \in {- 1, + 1}$ (3) K // divide the samples into K sets Output: Label_predict = (C_p1, C_p2,…, C_p(n/K)) // predict label $C_{pi} \in {- 1, + 1}$ Procedure: (1) Build training and testing set selected from D_feature: D_train, D_test (2) Train SVM model Initialization: c(0,0) =–10; g(0,0) =–10; // c and g are parameters of SVM bestacc = 0; bestc = 1; bestg = 0.1; N = 5 // 5-fold cross-validation Seek best parameters of c and g: for i = 1,2,…, m; // m is the size of c for j = 1,2,…, n; // n is the size of g cg(i,j) =svmtrain(Label_train, D_train, cmd); //SVM training if (cg(i,j) > bestacc) and (the termination condition of cross-validation is satisfied) {bestacc = cg(i,j); bestc = 2^c(i,j); bestg = 2^g(i,j);} end if end for end for Update current model =svmtrain(Label_train, D_train, bestc, bestg); (3) Simulation using testing set Label_predict =libsvmpredict(Label_test, D_test, model); Return Label_predict = (C_p1, C_p2,…, C_p(n/K)) //total of n/k samples as testing

Algorithm 1. Pseudocode of classification algorithm based on SVM.

Algorithm Support Vector Machine (SVM)
Input: (1) D_feature ={ (α₁₁, α₁₂,…, α₁₉),
(α₂₁, α₂₂,…, α₂₉),…,
(α_n1, α_n2,…, α_n9)}// feature matrix space
(2)Label_real = (C_r1, C_r2,…, C_{r n}) // Real label corresponding to Feature matrix //

C_{ri} \in {- 1, + 1}

(3) K // divide the samples into K sets
Output: Label_predict = (C_p1, C_p2,…, C_p(n/K)) // predict label

C_{pi} \in {- 1, + 1}

Procedure:
(1) Build training and testing set selected from D_feature: D_train, D_test
(2) Train SVM model
Initialization:
c(0,0) =–10; g(0,0) =–10; // c and g are parameters of SVM
bestacc = 0; bestc = 1; bestg = 0.1;
N = 5 // 5-fold cross-validation
Seek best parameters of c and g:
for i = 1,2,…, m; // m is the size of c
for j = 1,2,…, n; // n is the size of g
cg(i,j) =svmtrain(Label_train, D_train, cmd); //SVM training
if (cg(i,j) > bestacc) and (the termination condition of
cross-validation is satisfied)
{bestacc = cg(i,j); bestc = 2^c(i,j); bestg = 2^g(i,j);}
end if
end for
end for
Update current model =svmtrain(Label_train, D_train, bestc, bestg);
(3) Simulation using testing set
Label_predict =libsvmpredict(Label_test, D_test, model);
Return Label_predict = (C_p1, C_p2,…, C_p(n/K)) //total of n/k samples as testing

Procedure

In total, 15 volunteers are required to close their eyes without tranquilizer in a quiet state equipped on our Brain-Health device. And the EEGs are simultaneously recorded in the connected smart phone. Then, the nine features mentioned above constitute a feature space. Next, the classification in groups can be derived using SVM algorithm. Finally, six important criteria for performance evaluation are calculated compared with linear discriminant analysis (LDA) algorithm. In brief, the fundamental principle of LDA is that to find a linear decision boundary, which can minimize the intra-class variance and maximize the between-class variance, where the LDA is a common tool used as a reference of classification and pattern recognition.

We do two-tailed t-test on the EEG normalized power to evaluate whether there are significant differences between groups. Significance is defined as p < 0.05 for all tests.

Results and discussion

The results show that significant differences of normalized spectral power are found in HE comatose group compared with rehabilitated group and normal controls (Figure 5). Particularly, the normalized spectral power of low frequency wave (δ) significantly increases in HE comatose group (p < 0.05), while intermediate frequency waves (α₁, α₂, β₁, β₂) all decrease (p < 0.05). This conclusion is consistent with the existing research.^22–24 Meanwhile, no significant difference is found between the rehabilitated group and normal controls (p > 0.1), which implies that as HE patients convalesce after the treatment, EEGs tend to be in normal level. Based on these facts, two types of classification experiments based on SVM are performed subsequently. The one is between HE comatose group and rehabilitated group, which is conductive to assess patient’s rehabilitation condition for timely early warning. The other is between HE comatose group and normal controls in order to research the differences between patients and healthy persons for encephalopathy’s prevention.

Figure 5.

Normalized spectral power distribution in different EEG frequency bands.

In the first classification (HE comatose group and rehabilitated group), the details of the performance provided by the SVM compared with LDA algorithm are tabulated in Table 2. Regarding the overall classification results, the SVM provides a better performance of mean accuracy of 91.79%, sensitivity of 87.33%, specificity of 96.04%, AUC of 0.92, and MCC of 0.84, while the evaluation standards fell to mean accuracy of 74.23%, sensitivity of 86.95%, specificity of 61.54%, AUC of 0.74, and MCC of 0.51 based on LDA. From the results, we can see that the proposed algorithm results in an improvement with accuracy of 17.56%, sensitivity of 0.38%, specificity of 34.5%, AUC of 0.18, and MCC of 0.33, which indicates that the classification performance based on SVM is better than LDA. Note that the sensitivity and specificity should be analyzed conjointly rather than separately. Although the sensitivity based on LDA is slightly better than SVM using some features, absolutely poor performance of corresponding specificity is presented as well. Similarly, Table 3 summarizes the evaluation criteria of the second classification (HE comatose group and normal controls) based on SVM and LDA, respectively. The SVM algorithm achieves an improvement with accuracy of 17.81%, specificity of 38.04%, AUC of 0.18, and MCC of 0.32 compared with LDA algorithm.

Table 2.

Performance of SVM compared with LDA algorithm in Classification 1 (mean values).

Classification 1: classify comatose group and rehabilitated group
Algorithm	Feature(s)	Accuracy (%)	Sensitivity (%)	Specificity (%)	AUC	MCC
LDA	δ	74.99	71.25	78.84	0.75	0.50
	α₁	67.94	79.89	56.30	0.68	0.37
	α₂	73.80	85.31	62.54	0.74	0.49
	β₁	74.88	86.75	63.08	0.75	0.51
	β₂	73.81	84.06	63.52	0.74	0.49
	α₁/δ	70.34	92.34	49.08	0.71	0.46
	α₂/δ	74.83	94.35	54.77	0.75	0.54
	β₁/δ	75.27	95.10	54.80	0.75	0.55
	β₂/δ	74.01	92.86	54.99	0.74	0.52
	δ ∼ β₂/δ	82.42	87.55	77.46	0.83	0.65
	Average	74.23	86.95	61.54	0.74	0.51
SVM	δ ∼ β₂/δ	91.79	87.33	96.04	0.92	0.84

SVM: Support Vector Machine; LDA: linear discriminant analysis; AUC: area under the curve; MCC: Matthews correlation coefficient.

Table 3.

Performance of SVM compared with LDA algorithm in Classification 2 (mean values).

Classification 2: classify comatose group and normal controls
Algorithm	Feature(s)	Accuracy (%)	Sensitivity (%)	Specificity (%)	AUC	MCC
LDA	δ	81.38	76.01	86.58	0.81	0.62
	α₁	73.61	84.83	62.08	0.73	0.48
	α₂	76.23	88.71	63.13	0.76	0.54
	β₁	75.01	84.96	65.49	0.75	0.51
	β₂	78.39	90.17	66.37	0.78	0.58
	α₁/δ	72.35	97.59	46.94	0.72	0.52
	α₂/δ	70.11	98.97	41.72	0.70	0.49
	β₁/δ	72.51	96.92	48.61	0.73	0.52
	β₂/δ	72.03	98.50	46.03	0.72	0.52
	δ ∼ β₂/δ	89.22	88.79	89.64	0.89	0.78
	Average	76.08	90.55	61.66	0.76	0.56
SVM	δ ∼ β₂/δ	93.89	88.08	99.70	0.94	0.88

SVM: Support Vector Machine; LDA: linear discriminant analysis; AUC: area under the curve; MCC: Matthews correlation coefficient.

The ROC curves of above-mentioned two classifications are shown in Figure 6. For a comparison, the ROC curve of random classifier is used as a baseline which sensitivity identically equal to 1 – specificity (AUC = 0.5). In general, the ROC curve of SVM is closer to coordinate (0, 1) and further away from the baseline than LDA both in two classifications, which confirm the superior classification performance of the SVM.

Figure 6.

ROC curves of two classifications using SVM compared with LDA algorithm. Feature vector = [δ, α1, α2, β1, β2, α1/δ, α2/δ, β1/δ, β2/δ]. (a) Classification 1: classify comatose group and rehabilitated group (AUC of SVM = 0.93, AUC of LDA = 0.83) and (b) Classification 2: classify comatose group and normal controls (AUC of SVM = 0.94, AUC of LDA = 0.88).

In conclusion, the results have shown the significant difference with normalized spectral power between the comatose group and the other two groups, which is verified that the abnormal performance of EEG can be collected by our system. And the results indicate that the EEG has a characteristic pattern according to the consciousness level. Meanwhile, the results of above two classifications have shown that the performances of SVM are better than the LDA, which is verified by the feasibility of SVM algorithm for HE classification.

Conclusion and future work

In this article, we propose a real-time wireless wearable EEG system based on SVM for encephalopathy daily monitoring, named as Brain-Health. An innovative technology, which takes advantage of (a) the wearable EEG acquisition device, (b) the mobile terminal with the dedicated application, and (c) the intelligent classification algorithm based on SVM, is introduced in our Brain-Health system. The results demonstrate that the good performance of accuracy, sensitivity, specificity, AUC, MCC, and ROC curve is achieved in classifications based on SVM, which indicate the feasibility and effectiveness of our Brain-Health encephalopathy monitoring system.

In the future, we will focus on the following: (a) collect EEG data from more volunteers and optimize algorithm to improve the performance of classifier; (b) enrich the function and improve real-time response speed of mobile APP—meanwhile, strengthen the construction of the cloud for analysis and storage of the EEG recordings; and (c) extract and analyze different features applied to another encephalopathy, such as epilepsy, attention-deficit/hyperactivity disorder (ADHD), and Alzheimer’s disease (AD).

Footnotes

Acknowledgements

Qing Zhang and Pingping Wang contributed equally to this work and should be considered co-first authors.

Handling Editor: Shinsuke Hara

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This study was supported partly by the National key research and development program (2017YFC0114300, 2017YFB1103600, 2017YFB1103602), Key R&D (Research and Development, R&D) Program of Jiangsu Province (BE2016613, BE2016010, BE2016010-3, BE2016010-4, BE2017675, BE2017664), Key R&D Program of Zhejiang Province (2017C031SAB00881), Research and Development Program of Scientific Research Instrument and Equipment of the Chinese Academy of Sciences (YJKYYQ20170050), Science and Technology program of Jiangsu Province (BK20170387), Science and Technology program of Suzhou City (SYG201707, SYS201656, SYG201706), National Natural Science Foundation of China (61501452), Postdoctoral Science Foundation of Jiangsu Province (1501089C).

References

Alotaiby

Samie

Alshebeili

et al . A review of channel selection algorithms for EEG signal processing. EURASIP J Adv Sig Pr 2015; 2015: 66.

Wintermark

Hansen

Warfield

et al . Near-infrared spectroscopy versus magnetic resonance imaging to study brain perfusion in newborns with hypoxic-ischemic encephalopathy treated with hypothermia. Neuroimage 2013; 85: 287–293.

Jung

. An improved technique to consider mismatches between fMRI and EEG/MEG sources for fMRI constrained EEG/MEG source imaging. Biomed Eng Lett 2011; 1: 32–41.

Lee

Park

Jang

et al . Decoding brain states using functional magnetic resonance imaging. Biomed Eng Lett 2011; 1: 82–88.

Nishimura

Aoki

Inagawa

et al . Brain activities of visual thinkers and verbal thinkers: a MEG study. Neurosci Lett 2015; 594: 155–160.

Lovelace

Witt

Beyette

et al . Modular, Bluetooth enabled, wireless electroencephalograph (EEG) platform. In: International conference of the IEEE Engineering in Medicine and Biology Society, Osaka, Japan, 3–7 July 2013, pp.6361–6364. New York: IEEE.

Lin

Chang

et al . Review of wireless and wearable electroencephalogram systems and brain-computer interfaces—a mini-review. Gerontology 2009; 56: 112–119.

Casson

Smith

Duncan

et al . Wearable EEG: what is it, why is it needed and what does it entail? In: International conference of the IEEE Engineering in Medicine and Biology Society, Vancouver, BC, Canada, 21–24 August 2008, pp.5867–5870. New York: IEEE.

Mihajlović

Grundlehner

Vullers

et al . Wearable, wireless EEG solutions in daily life applications: what are we missing?IEEE J Biomed Health 2014; 19: 6–21.

10.

Wang

. A smart device enabled system for autonomous fall detection and alert. Int J Distrib Sens N. Epub ahead of print 17 February 2016. DOI: 10.1155/2016/2308183.

11.

Gargiulo

Bifulco

Cesarelli

et al . Wearable dry sensors with Bluetooth connection for use in remote patient monitoring systems. Stud Health Technol Inform 2010; 161: 57–65.

12.

Zander

Lehne

Ihme

et al . A dry EEG-system for scientific research and brain-computer interfaces. Front Neurosci 2011; 5: 53.

13.

Lin

Chiou

et al . Noninvasive neural prostheses using mobile and wireless EEG. P IEEE 2008; 96: 1167–1183.

14.

Mahajan

Pacareu

Abusaud

et al . Ambulatory EEG NeuroMonitor platform for engagement studies of children with development delays. Smart Biomed Physiol Sens Tech X 2013; 8719: 237–250.

15.

Olesen

Graversen

Hansen

et al . Spectral and dynamic electroencephalogram abnormalities are correlated to psychometric test performance in hepatic encephalopathy. Scand J Gastroenterol 2011; 46: 988–996.

16.

Amodio

Marchetti

Piccolo

et al . Spectral versus visual EEG analysis in mild hepatic encephalopathy. Clin Neurophysiol 1999; 110: 1334–1344.

17.

Montagnese

Rui

Schiff

et al . Prognostic benefit of the addition of a quantitative index of hepatic encephalopathy to the MELD score: the MELD-EEG. Liver Int 2015; 35: 58–64.

18.

Jurcak

Tsuzuki

Dan

. 10/20, 10/10, and 10/5 systems revisited: their validity as relative head-surface-based positioning systems. Neuroimage 2007; 34: 1600–1611.

19.

Chi

Jung

Cauwenberghs

. Dry-contact and noncontact biopotential electrodes: methodological review. IEEE Rev Biomed Eng 2010; 3: 106–119.

20.

Mullen

Kothe

Chi

et al . Real-time neuroimaging and cognitive monitoring using wearable dry EEG. IEEE T Biomed Eng 2015; 62: 2553–2567.

21.

Zhu

Chen

Zhang

et al . Design of portable multi-channel EEG signal acquisition system. In: International conference on biomedical engineering and informatics, Tianjin, China, 17–19 October 2009, pp.1–4. New York: IEEE.

22.

Knyazev

. Motivation, emotion, and their inhibitory control mirrored in brain oscillations. Neurosci Biobehav Rev 2007; 31: 377–395.

23.

Prichep

John

Ferris

et al . Prediction of longitudinal cognitive decline in normal elderly with subjective complaints using electrophysiological imaging. Neurobiol Aging 2006; 27: 471–481.

24.

Stappung

Fernández

Becerra

et al . Healthy aging: relationship between quantitative electroencephalogram and cognition. Neurosci Lett 2012; 510: 115–120.