Deep learning for healthcare applications based on physiological signals:

a review

Oliver Fausta,∗, Yuki Hagiwarab , Tan Jen Hongb , Oh Shu Lihb , U Rajendra Acharyab,c,d
a
Department of Engineering and Mathematics, Sheffield Hallam University, United Kingdom
b
Department of Electronic & Computer Engineering, Ngee Ann Polytechnic, Singapore
c
Department of Biomedical Engineering, School of Science and Technology, SIM University, Singapore
d
Department of Biomedical Imaging, Faculty of Medicine, University of Malaya, Kuala Lumpur, Malaysia

Abstract
Background and Objective: We have cast the net into the ocean of knowledge to retrieve the
latest scientific research on deep learning methods for physiological signals. We found 53 research
papers on this topic, published from 01.01.2008 to 31.12.2017.
Methods: An initial bibliometric analysis shows that the reviewed papers focused on Electromyogram
(EMG), Electroencephalogram (EEG), Electrocardiogram (ECG), and Electrooculogram (EOG).
These four categories were used to structure the subsequent content review.
Results: During the content review, we understood that deep learning performs better for big
and varied datasets than classic analysis and machine classification methods. Deep learning
algorithms try to develop the model by using all the available input.
Conclusions: This review paper depicts the application of various deep learning algorithms used
till recently, but in future it will be used for more healthcare areas to improve the quality of
diagnosis.
Keywords: Deep learning, Physiological signals, Electrocardiogram, Electroencephalogram,
Electromyogram, Electrooculogram

1. Introduction

Physiological signals are an invaluable data source which assists in disease detection, reha-
bilitation, and treatment [1]. The signals come from sensors implanted or placed on the skin
[2]. The application areas dictate where the sensors are to be placed [3, 4]. In turn, the sensor
location determines the characteristics of the physiological signal. Relevant information must
be extracted from the physiological signal in order to support a specific healthcare application
[5, 6]. It is difficult to establish what constitutes relevant information, because various medical
ontologies can be used for data interpretation [7]. It is unavoidable that these ontologies con-
tain conflicts and inconsistencies [8]. Another fundamental problem is that physiological signal
interpretation suffers from intra-individual variability [9]. For example, both electrode position
and noise influence the signal waveform [10]. Therefore, human interpretation requires years of
training to acquire specialized knowledge. Even with expert knowledge, manual physiological
signal interpretation suffers from intra- and inter-operator variability [11]. Furthermore, physi-
ological signal interpretation is a tiresome process where human errors can be caused by fatigue
[12, 13]. Computer supported signal analysis is not affected by fatigue related mistakes and it
can also eliminate both intra- as well as inter-observer variability. In addition, most computer
based analysis interpretation can be done quicker and more cost effective when compared to

∗
Corresponding author

Preprint submitted to Computer Methods and Programs in Biomedicine April 9, 2018

human interpretation [14]. Having established the need for computer based physiological sig-
nal analysis, we have to find the most appropriate processing methods for a given healthcare
application.
Nowadays, performance measures, such as classification accuracies [15, 16], govern computer
based analysis and classification of physiological signals for healthcare applications [17, 18, 19].
The goal is to design algorithms that outperform the established methods [20]. The design
process is based on the idea of having an offline and online system, as shown in Figure 1 [21, 22].
The offline system is used to design the required algorithm structure based on labeled data.
That algorithm structure is used in the online system to process measurement data. In most of
the cases, the system consists of three sequential processing steps: 1) preprocessing [23, 24], 2)
feature extraction [25, 26] 3) classification [27]. The first two steps establish the analysis system
which extracts information from the physiological signal. This approach is very target directed,
since the design process is governed by the feature performance. However, real competition
is difficult to establish, because the underlying testing data are rarely comparable. Another
problem is that, there is no way of knowing which feature extraction algorithm is suitable for
a given problem. The feature quality can only be established aposter priory with statistical
methods [28, 29]. This situation is dis-satisfactory for two reasons. Various feature extraction
algorithms must be evaluated on the physiological signals before selecting the best performing
feature extraction method. Even the increased effort fails to ensure that the most relevant
information are extracted. The second reason comes from the fact that only a small number of
features are used for decision making, because the performance of decision making algorithms
deteriorates for higher dimensional inputs [30]. Hence, the main design goal of current decision
support systems is to restrict the amount of information that underpins the decision making [31].
Having to restrict the amount of information for the decision-making process is a big problem,
because the decision quality will suffer. The negative effects become more prominent for large
and diverse physiological signal datasets that hold the information to tackle more challenging
application areas, such as Brain Computer Interface (BCI), cardiovascular diseases, sleep apnea,
sleep stages and muscular abnormalities.

Offline system Online system

Physiological signals Physiological signals

Feature extraction
and Feature extraction
assessment

Classification and assessment Decision support

Figure 1: Blcokdiagram for the design of a traditional decision support system based on physiological signals.

Current decision-making systems under-perform for large and varied datasets, because they
fail to process information that is sufficiently diverse to cover all scenarios. More depth is re-
quired to represent all the information in a dataset. Fundamentally, we require implicit knowl-
edge to make good decisions based on the information extracted from physiological signals. In
the past, that implicit knowledge was exclusively found in highly skilled medical practitioners.
Deep learning was designed to overcome these shortcomings by taking into consideration all in-
formation a training dataset has to offer. The method promises to establish implicit knowledge
in a machine. In this review, we investigate the validity of this claim by analysing the ways in

2
which deep learning has been applied to physiological signals. We found promising research in
the areas of Electrocardiogram (ECG), Electroencephalogram (EEG), Electromyogram (EMG)
and Electrooculogram (EOG) based healthcare applications. However, the amount of research
work does not reflect the diversity of healthcare application that can benefit from physiological
signal interpretation. Therefore, we adopt the position that the diversification of the research
is needed. The depth and specialization must come from training the deep learning algorithms
with larger and more varied data sets.
To support our claim that deep learning is a major step towards understanding physiological
signals, we have organized the remainder of the paper as follows. The next section establishes
the importance of implicit knowledge for physiological signal interpretation and it provides
background on deep learning. Section 3 presents bibliometric and content analysis results.
Based on these results, we put forward a range of research gaps in the discussion section. In this
section, we have also discussed limitations and future work. The paper concludes with Section
5.

2. Background
Physiological signals reflect the electrical activity of a specific body part [6]. As such, the
electrical activity provides information about the physiological condition. Traditionally, this
information can be used by medical practitioners for decision making [32, 33]. These decisions
have far reaching consequences on diagnosis, treatment monitoring, drug efficacy tests, and
quality of life. From a machine learning perspective, a medical practitioner is an intelligent
agent who makes good decisions. In order to reach these decisions some sort of knowledge is
required. This knowledge shapes the process which generates actions from both state and input.
For example, a clinician analyses an ECG signal. As that analysis process continues, there is a
growing suspicion that the signal was taken from a patient with coronary artery disease. More
and more suspicious waveform segments emerge until a positive diagnosis is reached. From a
processing perspective, the ECG signal is the input, the reading clinician is the system and the
system state is reflected in the growing suspicion.
In its simplest form, knowledge can be explicitly expressed as mathematical rules and physical
facts. Engineers make use of this explicit knowledge to create expert systems [34]. However, these
systems are limited to a specific domain and they operate in highly controlled environments.
Only for these environments, we are reasonably sure about the physical facts and mathematical
rules which shape the correct behaviour. Furthermore, expert systems don’t reflect the concept
of suspicion and common sense. Overcoming these limitations requires implicit knowledge which
can be found in the complex wiring and the synaptic setup of biological brains as well as the
mechanical and sensory properties of biological bodies. The only way for computers to mimic
implicit knowledge is to learn from examples. The idea is to find a way to learn general features
in order to make sense of new data. This description highlights the central role of data for
establishing implicit knowledge. The amount of data must be sufficiently large to provide many
training examples from which a large set of parameters can be extracted. Only a large number
of parameters give rise to the richness of class functions which model the implicit knowledge.
Deep learning is a change in basic assumptions of artificial intelligence algorithm design [35].
This change percolates through to all application areas of machine learning, such as computer
vision, speech recognition, natural language processing and indeed diagnosis support [36]. Cen-
tral to deep learning are the ideas of parallel processing and networked entities. The next section
details these ideas by discussing the deep learning methods.

2.1. Deep Learning Methods

Deep learning belongs to the class of machine learning methods. It is a special form of
representation-based learning, where a network learns and constructs inherent features from

3
each successive hidden layer of neurons [31]. The term “deep” is derived from the numerous
hidden layers in the Artificial Neural Network (ANN) structure.
The ANN algorithm models the functionality of a biological brain [37]. The model is real-
ized with a structure that is made up of input-, hidden-, and output-layers, as shown in Figure
2. Every neuron or node (nerve cell) is connected to each neuron in the next layer through a
connection link. A nerve cell is made up of axon (output), dendrites (input), a node (soma), nu-
cleus (activation function), and synapses (weights) [38]. The activation function in the artificial
neuron acts as the nucleus in a biological neuron whereas the input signals and its respective
weights model the dendrites and synapses respectively. Figure 3 illustrates the neuron structure.

Figure 2: Traditional ANN

Unfortunately, the ANN structure is receptive to translation and shift deviation, which
may adversely affect the classification performance [39]. To eliminate these shortcomings, an
extended version of ANN, the Convolutional Neural Network (CNN), was developed [40]. The
CNN architecture ensures translation and shift invariance [41]. Figure 4 illustrates a generic
CNN network structure. It is a feed-forward network, which comprises of: convolution, pooling,
and fully-connected, layers [41]. They are briefly explained below.
1. Convolution layer: The input sample is convolved with a kernel (weights) in this layer. The
output of this layer is the feature map. The stride controls how much the kernel convolves
with the input sample. The convolution operation acts as a feature extractor by learning
from the diverse input signals. The extracted features can be used for classification in
subsequent layers. Figure 5 illustrates a convolution operation between f (input) and g
(kernel), giving an output c. Equations 1 and 2 provide an example of the convolution

4
 Figure 3: Neuron structure

calculation.
c(1) = 1 × 3 + 2 × 4 + 3 × 1 = 14
c(2) = 2 × 3 + 3 × 4 + 4 × 1 = 22 (1)
c(3) = 3 × 3 + 4 × 4 + 5 × 1 = 30

Output1 = 22 × W1 + 38 × W3 + 54 × W5
(2)
Output2 = 22 × W2 + 38 × W4 + 54 × W6
2. Pooling layer: This is a down-sampling layer where the pooling operation is employed to
reduce the spatial dimension of the input sample, while retaining the significant informa-
tion. The pooling operation can be average, max, or sum. The max-pooling operation is
typically employed. An example of a max-pooling operation, with a stride of 2, is shown
Figure 6. In this example, the number of input samples are halved by retaining only the
maximum value within a selected stride. In a stride that contains 14 and 22, the value 22
is retained, and, 14 is discarded.
3. Fully-connected layer: Fully-connected signifies that each neuron in the previous layer is
connected to all the neurons in the current layer. The total number of fully-connected
neurons in the final layer determines the number of classes. Figure 7 provides a graphical
representation of a fully-connected layer. The neurons are all connected and each connec-
tion has a specific weight. This layer establishes a weighted sum of all the outputs from
the previous layer to determine a specific target output.
A leaky rectifier linear unit [42] is used as an activation function after the convolution layer.
The purpose is to map the output to the input set and introduce non-linearity as well as sparsity
to the network. The CNN is trained with backpropagation [43] and the hyperparameters may
be tuned for optimal training performance.
In addition to CNN, there are other deep learning architectures, such as autoencoder [44],
deep generative models [45] [46] and Recurrent Neural Network (RNN), that can be used to
monitor the physiological signals [47].

5
 Figure 4: CNN network structure

Figure 5: Convolution layer

The autoencoder is an unsupervised neural network that is trained to imitate the input
to its output. The dimension of the input is the same as the dimension of the output. The
encoders are stacked together to form a deep autoencoder network. The autoencoder takes
unlabelled inputs, encodes these inputs, and subsequently it reconstructs the inputs as precisely
as possible. Hence, the network structure must determine which data features are significant.
An autoencoder consists of three layers: input, output, and hidden. Encoding and decoding
are the two main steps in the autoencoder algorithm. During encoding and decoding, the same
weights are employed to encode the feature and reconstruct an output sample in the output.
Autoencoders are trained with a backpropagation algorithm that employs a metric known as the
loss function [43]. That function computes the amount of information which is lost during input
reconstruction. Thus, a network structure with a small loss value will produce a reconstruction
that is nearly identical to the input sample [37].
Deep Belief Network (DBN) [45] and Restricted Boltzmann Machine (RBM) [48] are common
forms of the deep generative model, where DBNs are stacked layers of RBM. The RBM is made
up of a two-layer neural net with one visible and one hidden layer. Each node in the layer learns
a single feature from the input data by making random decisions on whether to transmit the
input or not. The nodes in the first (visible) layer are connected to every node in the second
(hidden) layer. The RBM takes the inputs and translates them into a set of numbers that encode

6
 Figure 6: Max pooling layer

Figure 7: Fully connected layer

the inputs. Then, a backward translation is implemented to reconstruct the inputs. The RBM
network is trained to reconstruct through forward and backward passes [49, Chapter 6].
The DBN is a probabilistic model with several hidden layers [50]. It can be efficiently
trained by using a technique known as greedy layer-wise pre-training [51]. Typically, the DBNs
are stacked layers of RBM. The first layer of the RBM is trained to reconstruct its input. After
which, the first hidden layer is considered as the visible layer and it is trained using the outputs
from the input layer. This process is iterated until all layers in the structure are trained [45].
In contrast to the feed-forward network, the RNN employs a recursive approach (recurrent
network) whereby the network performs a routine task with the output being dependent on the
previous computation. This functionality is created with inbuilt memory. The most common
type of RNN is the Long Short-Term Memory (LSTM) network [52]. It has the capability to
learn long-term dependencies. The LSTM algorithm incorporates a memory block with three
gates: the input, output, and forget gate. These gates control the cell state and decide which
information to add or remove from the network. This process repeats for every input.
1. Input gate: decides what new information is to be stored and updated in the cell state.
2. Output gate: judges what information is used based on the cell state.
3. Forget gate: evaluates what information is redundant and discards it from the cell state.
These deep learning architectures have demonstrated their potential by surpassing the perfor-
mance of traditional machine learning techniques [31]. Furthermore, deep learning algorithms
minimize the need for feature engineering. The next section, focuses on scientific work that
applied deep learning to physiological signal interpretation for health care applications.

3. Review

In this review, we consider 53 articles focusing on deep learning methods applied to phys-
iological signals for healthcare applications. These papers were published in the period from
01.01.2008 to 31.12.2017. Figure 8 shows the yearly distribution of 53 articles together with a
trend line [53]. Apart from the data distribution, the figure also features a trend line which was

7
generated with a linear regression approach [54]. The large positive gradient of the trend line
indicates more papers, on this topic, published in recent years.

30
30 Trend line gradient: 2.46
25
Number of articles

20
15
10 9 8
5 2 2
0 0 0 1 1
0
2008 2010 2012 2014 2016
Year

Figure 8: Number of articles on deep learning, published each year from 01.01.2008 to 31.12.2017

Our review of the 53 articles is based on two analysis methods. The first method is a
bibliometric analysis, which helps us to detect the structure that governs relationship between
the paper topics. We use this information to arrange the second analysis method, which deals
with the content of the reviewed papers. The paper content is dissected in terms of common
parameters that enable us to compare the studies across different healthcare applications.

3.1. Bibliometric analysis

The bibliometric analysis establishes the topic structure of the reviewed papers. This un-
locks the paper content. We have used the well-known VOSviewer1 software to establish a
co-occurrence network [55, 56]. Before feeding the data to the VOSviewer algorithm, we created
a thesaurus file to group keywords, having the same meaning, into broader topics. Table A.8
shows that mapping. Figure 9 shows the co-occurrence network and the topic clusters. Table
1 reveals the topic structure by aligning the topics to clusters. These clusters are centred on a
specific physiological signal. To be specific, the largest topic cluster2 includes EMG. The physi-
ological signals are widely used in healthcare areas, such as speech, prosthesis and rehabilitation
robotics. CNN is the deep leaning method of choice in this cluster. The second cluster is related
to the EEG which is used for BCI. the cluster shows that noise is a problem for EEG processing.
Cluster 3 focuses on ECG, which is linked to Cardiovascular Disease (CVD), arrhythmia and
Automated External Defibrillator (AED). The last cluster deals with EOG which can be used
to detect sleep problems. The deep learning methods of DBN and RNN are used to the detect
of Obstructive Sleep Apnea (OSA)using different physiological signals. The clustering helps us
to structure the content review, which is presented in the next section.

3.2. Content review

The bibliometric research shows four distinct clusters focused on four physiological signals
namely: ECG, EEG, EMG and EOG. They are briefly discussed in the following sections.

1
Web page (Last accessed 11.12.2017): http://www.vosviewer.com/
2
In terms of the number of papers that include a keyword that was mapped to the cluster topics.

8
 Figure 9: Network visualization for the author supplied keywords

Table 1: Topic cluster summary, where C is the Cluster number. Colour indicates the cluster colour as shown in
Figure 9.
C Topics Colour
1 emg, cnn, cognitive states, dnn, fnirs, hand gesture, infrared spec- Red
troscopy, ml, myoelectric interfaces, prostheses, real-time, recognition,
rehabilitation robotics, speech, wearable, word generation
2 eeg, bci, learning, motor imagery, multifractal attribute, neuromorphic Blue
computing, noise, wavelet leader
3 ecg, aed, arrhythmia, contaminant mitigation, cvd, motion artifact Yellow
4 eog, dbn, heartbeat, osa, rnn, signal, sleep, svm Green

9
 Table 2: Summary of works carried out using DL algorithms
with EMG signals.

Author Application DL algo- Data Results

rithm
Xia et al., 2017 [60] Limb movement estima- RNN Measurements from The RNN outperforms other
tion eight healthy subjects methods for estimating a 3D
trajectory
Zhai et al., 2017 [61] Neuroprosthesis control CNN NinaPro Database accuracy: 83%
(DB)2&3 [62]
Park et al., 2016 [63] Movement intention de- CNN Kinematic and EMG accuracy: > 90%
coding data NinaPro DB [64]
Atzori et al., 2016 [65] Hand movement classi- CNN NinaPro DB [62] accuracy: 66.59%
fication
Geng et al., 2016 [66] Gesture recognition CNN 18 subjects perform- accuracy: 99.5%
ing 16 gestures each
recorded 10 times
10

Wand and Schmidhu- Speech recognition Deep Neu- EMG-UKA Corpus [68] DNN outperforms a Gaussian
ber, 2016 [67] ral Network Mixture Model (GMM) front
(DNN) end
Allard et al., 2016 [69] Robotic arm guidance CNN 18 subjects performing accuracy: ≈ 97.9%
7 gestures
Wand and Schultz, 2014 Speech recognition DNN 25 sessions from 25 sub- Visualization of hidden nodes
[70] jects activity
3.2.1. Deep learning methods applied to EMG
The EMG measures the electrical activity of skeletal muscles [57]. The electrical sensors,
known as electrodes, are placed on the skin above the muscles of interest [58]. The EMG signals
indicate: muscle activation, force produced by the muscle, and muscle state [59]. All these factors
are superimposed, i.e. the EMG signal measurement picks up the contributions from different
sources and sums them up. Hence, it is difficult to gain information about one particular aspect
by observing the EMG signal. Table 2 details the research work which describe the deep learning
methods used to analyze the EMG signal. Individual columns healthcare application area, Deep
Learning (DL) algorithm, the data used for the study, and the study results. As such, the DL
algorithms were introduced in Section 2.1.

3.2.2. Deep learning methods applied to EEG

The EEG is measured by placing electrodes on the skull of the subject, where they pick up
the electrical activity of the brain [71, 72]. As such, that activity results from firing neurons
emitting action potentials. At any given time, the electrode sums up many charges from different
sources. The resulting EEG signal has a noise like characteristic, which makes the interpretation
difficult. Furthermore, the signal is affected by EOG and EMG artifacts [73, 74, 75]. It takes the
trained eye of a practitioner to spot the features that indicate a specific mental state [76]. One of
the main drivers behind the application of DL algorithms applied to EEG is BCI [77]. The real-
time nature of this application makes human signal interpretation impossible [78]. Hence, BCI
requires automated decision making. Table 3 lists the reviewed research works conducted using
deep learning with EEG signals. The list contains one exception, Huve et al. used functional
near-infrared spectroscopy [79] to track neural dynamics of the brain [80].

3.2.3. Deep learning methods applied to ECG

The ECG is measured by placing electrodes on the chest [81]. These electrodes record the
electrical activity of the human heart [82]. The signal reflects the functioning of the heart and it
has well distinguishable features, even in the time domain. The difficulty in ECG interpretation
is to spot the morphological changes which indicate a particular cardiac problem [83, 84] or
diabetes [85, 86]. These abnormalities may be minute and very often they may be transients or
present all the time [87]. The research work, listed in Table 4, discusses the DL algorithms used
for the automated detection of various cardiac abnormalities. The scientific work, described
in 13 out of 14 reviewed papers, relies on measurements which were taken from public DBs.
Therefore, the results can be compared with state-of-the-art methods. For example, Acharya et
al. [88] reported the performance measures obtained through the deep learning system alongside
the performance figures of the traditional diagnosis support systems. Their system outperformed
traditional approaches and it has the added benefit of not having to perform feature extraction
and de-noising. Table 4 lists the works reported on deep learning applied to ECG signals.

3.2.4. Deep learning methods applied to EOG and a combination of signals

The EOG measures the corneo-retinal standing potential which occurs between back and
front of the human eye. Two electrodes are placed either left and right of the eye or above
and below the eye. The signal can be used to detect eye movements [89]. It is useful for
ophthalmological diagnosis [90]. However, the EOG is affected by noise and artifacts, that makes
the signal interpretation difficult [91]. Table 5 details research work which aims to overcome
these difficulties with deep learning algorithms. Only Zhu et al. applied the deep leaning
algorithm exclusively to EOG signals. Table 5 presents a summary of works carried out using
DL algorithms with EOG and other physiological signals.

11
 Table 3: Summary of works carried out using DL algorithms
with EEG signals.

Author Application DL algo- Data Results

rithm
Fraiwan et al., 2017 [93] Sleep state identifica- Autoencoders Measurements [94] 80.4% accuracy
tion
Schirrmeister et al., EEG decoding and visu- CNN BCI competition IV dataset Up 89.8% accuracy
2017 [95] alization 2a [96] and measurement data
Hosseini et al., 2017 [97] Epileptogenicity local- CNN EEG and rs-fMRI measure- Normal p-value 1.85e-14, p-
ization ments from the ECoG dataset Seizure value 4.64e-27
[98]
Schirrmeister et al., Decoding excited move- CNN Not reported Accuracy comparable to stan-
2017 [99] ments dard methods
van Putten et al., 2017 Butcome prediction CNN EEGs from 287 patients at 12 Sensitivity of 58% at a speci-
[100] for patients with a h after cardiac arrest and 399 ficity of 100% for the predic-
postanoxic coma after patients at 24 h after cardiac tion of poor outcome
12

cardiac arrest arrest

Spampinato et al., 2017 Discriminate brain ac- CNN Six subject were shown 2000 Accuracy 86.9%
[101] tivity images in 4 sessions
Kiral-Kornek et al., BCI CNN 6 subjects up to 1000 individ- Power comparisons of various
2017 [102] ual hand squeezes processing platforms.
Acharya et al., 2017 Seizure detection CNN Freiburg EEG DB [104, 105, accuracy: 88.67%
[103] 106, 107, 108]
Lu et al., 2017 [109] Motor imagery classifi- RBM BCI competition IV data set Accuracy has been improved
cation 2b [110, 111, 112] about 5% compared with
other methods.
Huve et al., 2017 [80] Tracking of neural dy- CNN, DNN 1 subject 180 trials DNN outperforms CNN
namics
Hajinoroozi et al., 2016 Prediction of driver’s CNN 37 subjects, 70 sessions Area under the receiver oper-
[113] cognitive performance ating characteristic: 86.08
Nurse et al., 2016 [114] BCI CNN 1 subject 30 min of data accuracy 81%
 Jingwei et al., 2015 Response representa- CNN EEG motor activity data set accuracy: 100%
[115] tion [116]
Hajinoroozi et al., 2015 Prediction of driver’s CNN 37 subjects, 80 sessions Area under the receiver oper-
[117] cognitive performance ating characteristic: 82.78
Jirayucharoensak et al., Based Emotion Recog- Deep learn- DEAP Dataset [119] valence accuracy: 49.52%,
2014 [118] nition ing network arousal accuracy 46.03%
An et al., 2014 [120] Based Emotion Recog- Deep learn- 4 subjects 60 trials valence accuracy: 49.52%,
nition ing network arousal accuracy 46.03%
Zheng et al., 2014 [121] Emotion classification DBN 6 subjects 2 trials each accuracy: 87.62%
Li and Cichocki, 2014 Motor imagery DBN 3 subjects performing motor Accuracy: 96%
[122] imagery tasks in four sessions.
Each session had 15 trails.
Jia et al., 2014 [123] Affective state recogni- RBM DEAP Dataset [119] RBM-based model to extract
tion representative features and to
reduce the data dimensional-
ity
13

Ren and Wu, 2014 [124] Feature extraction CNN Dataset III in 2003 BCIC II, Accuracy: 87.33%
Dataset Iva in 2005 BCIC III,
Dataset III in 2005 BCIC III
[110, 111, 112]
Ahmed et al., 2013 [125] Detecting target images DBN Not specified DBN outperforms SVM.
Mirowski et al., 2008 Epileptic seizure predic- CNN Freiburg EEG DB [104, 105, zero-false-alarm seizure pre-
[126] tion 106, 107, 108] diction on 20 patients out of
21
Cecotti and Graeser, Motor imagery classifi- CNN 2 subjects 5 trails of about 3 Recognition accuracy: 53.47%
2008 [127] cation minutes
 Table 4: Summary of works carried out using DL algorithms
with ECG signals.

Author Application DL algo- Data Results

rithm
Acharya et al., 2017 Arrhythmia detection CNN MIT-BIH arrhythmia Accuracy of 92.50%, sensitiv-
[128] using different intervals database [129] ity of 98.09%, specificity of
of tachycardia ECG 93.13%
segments
Tan et al., 2017 [130] Coronary artery disease LSTM, Physionet databases: Fan- Accuracy of 99.85%.
signal identification CNN tasia (for Normal) and St.-
Petersburg Institute of Car-
diology Technics (for CAD)
[129]
Acharya et al., 2017 Coronary artery disease CNN Physionet databases: Fan- Accuracy of 94.95% and
[131] detection tasia (for Normal) and St.- 95.11% are obtained for
Petersburg Institute of Car- two and five seconds ECG
14

diology Technics (for CAD) segments respectively.

[129]
Pourbabaee et al., 2017 Features for Screening CNN The PAF prediction challenge Precision=93.6%
[132] Paroxysmal Atrial Fib- database [133]
rillatio
Zheng et al., 2017 [134] ECG identification DNN MIT arrhythmia database 94.39% recognition rate
[135] and self-collected data
Majumdar et al., 2017 Arrhythmia classifica- Robust MIT arrhythmia database 97.0% recognition rate
[136] tion deep dic- [135]
tionary
learning
Shashikumar et al., Monitoring and detect- CNN 98 subjects, 45 with atrial fib- Up to 91.8% accuracy
2017 [137] ing atrial fibrillation rillation and 53 with other
rhythms.
 Acharya et al., 2017 [88] Detection of myocardial CNN lead II from the ECG DB accuracy: 93.18%
infarction (Physikalisch-Technische
Bundesanstalt diagnostic
ECG DB) [129]
Lou et al., 2017 [138] Heartbeat classification DNN Lead II from the MIT-BIH ar- Accuracy: 97.5%
rhythmia DB [135]
Acharya et al., 2017 Identification of ventric- CNN MIT-BIH arrhythmia DB accuracy: 92.50%
[139] ular arrhythmias (MITDB) [129, 135], MIT-
BIH malignant ventricular
arrhythmia DB (VFDB)
[140], Creighton University
ventricular tachyarrhythmia
DB (CUDB) [141]
Acharya et al., 2017 Heart beat classification CNN MIT-BIH arrhythmia DB accuracy: 94.03%
[142] [129]
Cheng et al., 2017 [143] Sleep apnea detection RNN ECG sleep apnea DB [144] accuracy: ≈ 90%
15

Taji et al., 2017 [145] Signal quality classifica- RBM MIT-BIH arrhythmia DB accuracy: 99.5%
tion [129]
Lou et al., 2017 [138] Heartbeat classification RBM MIT-BIH arrhythmia DB RBM-based model to extract
[135] representative features and to
reduce the data dimensional-
ity
Muduli et al., 2017 [146] Fetal-ECG signal recon- Stacked De- Abdominal non-invasive accuracy: 99.5%
struction noising Au- FECG DB [129]
toencoder
Kiranyaz et al., 2016 Heart beat classification CNN MIT/BIH arrhythmia DB accuracy: 99%
[147] [129]
Zheng et al., 2014 [148] Congestive heart failure CNN BIDMC Congestive Heart accuracy: 94.67%
detection Failure data set [129]
 Table 5: Summary of works carried out using DL algorithms
with EOG and other signals.

Author Application Signal DL algo- Data Results

rithm
Du et al., 2017 [149] Driving fatigue detec- EOG and Autoencoder Measurements from 21 Correlation Coefficient
tion EEG subjects 0.85, Root Mean Square
Error 0.09
Zhang et al., 2017 [150] Momentary mental EOG, EEG, CNN 6 subjects and 2 ses- Accuracy: 93.8%
workload classification ECG sions each
Xia et al., 2017 [151] Sleep stage classifica- EOG, EEG DBN sleep EDF DB [Ex- Accuracy: 83.3%
tion panded] in Physionet
[129, 152]
Zhu et al., 2014 [153] Drowsiness detection EOG CNN 22 subjects and 22 ses- mean correlation coeffi-
sions each cient of 0.73
Längkvist et al., 2012 Sleep Stage Classifica- EOG, EEG, DBN 25 acquisitions from 72.2%
[154] tion EMG PhysioNet [129] for
16

training and test-

ing and Home Sleep
Dataset 60 hours from
normal one subject for
validation
4. Discussion
Even though the deep learning is still in its infancy, published studies have shown that it
possesses the capability of faster and more reliable diagnoses in physiological signals. That
potency may well trigger a shift away from the currently used decision support methods, such
as Support Vector Machine (SVM) and K-Nearest Neighbour (K-NN), towards deep learning
[92]. It can noted from Table 4 that the employment of deep learning in ECG signals yielded
promising diagnostic performances. This is because the developed model is able to capture the
distinctive features from the ECG signals. Hence, the network can be trained from these learned
features even without big data, achieving desirable diagnostic performance. On the other hand,
the automated analysis of EMG and EEG signals is more challenging as these signals are more
chaotic in nature. Therefore, it is more complicated for the network to learn from the hidden
and subtle information present in these signals.
During the content review, we have studied the DL algorithm types used for their research
works. Table 6 shows the number of studies that used a particular DL algorithm to analyze a
specific physiological signals. CNN was used in 5, 15, 10 and 2 research works on EMG, EEG,
and ECG respectively. Overall, CNN was used 32 times, that makes it by far the most popular
DL algorithm. At the bottom of the table is LSTM which was only used for ECG processing.

Table 6: Summary of various DL algorithms applied to different physiological signals.

DL algorithm EMG EEG ECG EOG Total
CNN 5 15 10 2 32
DNN 2 1 2 0 5
DBN 0 3 0 2 5
RBM 0 2 2 0 4
Autoencoder 0 1 1 1 3
RNN 1 0 1 0 2
Deep learning network 0 2 0 0 2
Robust deep leanring 0 0 1 0 1
LSTM 0 0 1 0 1

Apart from the algorithms, we have also studied the type of data used to conduct these
studies. Table 7 indicates a summary of the data type used to implement the DL algorithm.
Row one indicates that 28 studies were conducted using the public databases. Only 20 studies
used private databases and 3 studies have used both databases. 16 out of 17 studies using ECG
and 8 out of 23 studies on EEG have used public data. Indeed, 12 studies have used private
EEG data to implement the deep learning algorithms.

Table 7: Summary of data type used to implement the DL algorithm.

EMG EEG ECG EOG Total
Publicly available databases 3 8 16 1 28
Private databases 4 12 1 3 20
Both 1 1 0 1 3
Not reported 0 2 0 0 2
Number of papers 8 23 17 5 53

Conventional machine learning approaches require lots of time and effort for feature selection.
These features must extract relevant information from huge and diverse data in order to produce
the best diagnostic performance. Furthermore, the best algorithm is unknown and thus, a lot
of trial and error is necessary to select the best feature extraction algorithms and classification
methods to develop a robust and reliable decision support systems for the physiological signals.

17
On the other hand, deep learning eliminates the need for feature extraction and feature selection.
The decision-making algorithm can consider all available evidence [155].
The training algorithms for deep leaning systems have a high computational complexity.
Hence, this results in a high run-time complexity which translates into a long training time
[156, 157, 158]. This will be a problem during the design phase, because during this phase a
designer must decide what deep leaning architecture to use. Once the architecture is chosen, the
tuning parameters must be adjusted. Both the structure selection and parameter adjustment
will basically influence the model. Hence, it is necessary to have many test runs. Shortening the
training phase of deep leaning models is an active area of research [159]. The challenge is speeding
up the training process in a parallel distributed processing system [160]. The network between
the individual processors becomes the bottle neck [161]. Graphics Processing Units (GPUs) can
be used to reduce the network latency [162].
For this state of the art technology, the training time may be an issue, because it can influence
the model selection strategies. To be specific, none of the reviewed papers approached the model
selection process with statistical methods, such as cross validation [163]. We can only assume
that the deep learning architectures, used in the reviewed scientific work, were selected based
on single run trial. Similarly, we have to assume that the hyperparameters were also optimized
by training the network once. This is a serious shortcoming, because these statistical validation
methods reduce the sample selection bias [164, 165].
In addition, the use of deep learning can also be extended from one-dimensional physiological
signals to two-dimensional medical images. Researchers are also exploring the benefits utilizing
deep learning in medical image analyses [166]. It was reported that the automated staging and
detection of Alzheimers disease, breast cancer, and lung cancer has shown optimistic diagnostic
performances.

5. Conclusion

In this paper, we reviewed 53 papers on deep learning methods applied to healthcare ap-
plications based on physiological signals. The analysis was carried out in two distinct steps.
The first step was bibliometric keyword analysis based on the co-occurrence map. This analysis
step reveals the connection between the topics covered in the reviewed papers. We found four
distinct clusters, one for each physiological signal. That result helped us to structure the second
analysis step, which focuses on the paper content. As such, the paper content was established
by extracting the specific application area, the deep learning algorithm, system performance,
and the types of dataset used to develop the system.
The fact that deep learning algorithms performs well with large and diverse datasets which
has two consequences. First, the dataset becomes critically important for the system design.
Therefore, we focused our efforts on this criterion during our analysis. We found that, the scien-
tific work, documented in 31 of the reviewed papers, was based on one or more freely available
datasets. Therefore, we predict that the importance of these freely available public datasets may
increase. The other consequence is that deep learning algorithms will perform well in practical
settings, because clinical routine produces lots of data with large variations. However, none
of the reviewed papers verified this in a practical setting. Another fundamental point is that,
our literature survey yielded only 53 papers. The small number of studies imply that there is
scope for future work. To be specific, 53 papers do not reflect the comprehensive healthcare
applications based on the physiological signals. In future, there may be more advanced deep
learning algorithms focused on the early detection of diseases using physiological signals.

Acknowledgements

Funding: No external funding sources supported this review.

18
 Competing interests: The authors declare that they have no conflict of interest.
Statements of ethical approval: For this review article, the authors did not undertake work
that involved human participants or animals.

Acronyms

AED Automated External Defibrillator

ANN Artificial Neural Network
BCI Brain Computer Interface
CNN Convolutional Neural Network
CVD Cardiovascular Disease
DB Database
DBN Deep Belief Network
DL Deep Learning
DNN Deep Neural Network
ECG Electrocardiogram
EEG Electroencephalogram
EMG Electromyogram
EOG Electrooculogram
GMM Gaussian Mixture Model
GPU Graphics Processing Unit
K-NN K-Nearest Neighbour
LSTM Long Short-Term Memory
OSA Obstructive Sleep Apnea
RBM Restricted Boltzmann Machine
RNN Recurrent Neural Network
SVM Support Vector Machine

References

[1] O. Faust, M. G. Bairy, Nonlinear analysis of physiological signals: a review, Journal of

Mechanics in Medicine and Biology 12 (2012) 1240015.

[2] D. Wang, Y. Shang, Modeling physiological data with deep belief networks, International
journal of information and education technology (IJIET) 3 (2013) 505.

[3] H. Kantz, J. Kurths, G. Mayer-Kress, Nonlinear analysis of physiological data, Springer

Science & Business Media, 2012.

[4] W. Van Drongelen, Signal processing for neuroscientists: an introduction to the analysis
of physiological signals, Elsevier, 2006.

[5] L. Sörnmo, P. Laguna, Bioelectrical signal processing in cardiac and neurological applica-
tions, volume 8, Academic Press, 2005.

[6] S. R. Devasahayam, Signals and systems in biomedical engineering: signal processing and
physiological systems modeling, Springer Science & Business Media, 2012.

[7] R. Miotto, F. Wang, S. Wang, X. Jiang, J. T. Dudley, Deep learning for healthcare:
review, opportunities and challenges, Briefings in Bioinformatics (2017) bbx044.

[8] K. Gödel, On formally undecidable propositions of Principia Mathematica and related

systems, Courier Corporation, 1992.

19
 [9] L. K. Morrell, F. Morrell, Evoked potentials and reaction times: a study of intra-individual
variability, Electroencephalography and Clinical Neurophysiology 20 (1966) 567–575.

[10] B. Schijvenaars, Intra-individual Variability of the Electrocardiogram: Assessment and

exploitation in computerized ECG analysis, 2000.

[11] O. Faust, E. Y. Ng, Computer aided diagnosis for cardiovascular diseases based on ecg
signals: A survey, Journal of Mechanics in Medicine and Biology 16 (2016) 1640001.

[12] E. A. Krupinski, K. S. Berbaum, R. T. Caldwell, K. M. Schartz, J. Kim, Long radiology

workdays reduce detection and accommodation accuracy, Journal of the American College
of Radiology 7 (2010) 698–704.

[13] T. Vertinsky, B. Forster, Prevalence of eye strain among radiologists: influence of viewing
variables on symptoms, American Journal of Roentgenology 184 (2005) 681–686.

[14] U. R. Acharya, O. Faust, S. V. Sree, D. N. Ghista, S. Dua, P. Joseph, V. T. Ahamed,

N. Janarthanan, T. Tamura, An integrated diabetic index using heart rate variability signal
features for diagnosis of diabetes, Computer methods in biomechanics and biomedical
engineering 16 (2013) 222–234.

[15] A. Miller, B. Blott, et al., Review of neural network applications in medical imaging and
signal processing, Medical and Biological Engineering and Computing 30 (1992) 449–464.

[16] O. Faust, W. Yu, N. A. Kadri, Computer-based identification of normal and alcoholic eeg
signals using wavelet packets and energy measures, Journal of Mechanics in Medicine and
Biology 13 (2013) 1350033.

[17] O. Faust, U. R. Acharya, T. Tamura, Formal design methods for reliable computer-aided
diagnosis: a review, IEEE reviews in biomedical engineering 5 (2012) 15–28.

[18] K. Y. Zhi, O. Faust, W. Yu, Wavelet based machine learning techniques for electrocar-
diogram signal analysis, Journal of Medical Imaging and Health Informatics 4 (2014)
737–742.

[19] U. R. Acharya, E. C.-P. Chua, O. Faust, T.-C. Lim, L. F. B. Lim, Automated detection
of sleep apnea from electrocardiogram signals using nonlinear parameters, Physiological
measurement 32 (2011) 287.

[20] O. Faust, U. R. Acharya, E. Ng, H. Fujita, A review of ecg-based diagnosis support

systems for obstructive sleep apnea, Journal of Mechanics in Medicine and Biology 16
(2016) 1640004.

[21] O. Faust, U. R. Acharya, H. Adeli, A. Adeli, Wavelet-based eeg processing for computer-
aided seizure detection and epilepsy diagnosis, Seizure 26 (2015) 56–64.

[22] O. Faust, U. R. Acharya, V. K. Sudarshan, R. San Tan, C. H. Yeong, F. Molinari, K. H. Ng,

Computer aided diagnosis of coronary artery disease, myocardial infarction and carotid
atherosclerosis using ultrasound images: A review, Physica Medica 33 (2017) 1–15.

[23] R. Rao, R. Derakhshani, A comparison of eeg preprocessing methods using time delay
neural networks, in: Neural Engineering, 2005. Conference Proceedings. 2nd International
IEEE EMBS Conference on, IEEE, pp. 262–264.

[24] T. Kalayci, O. Ozdamar, Wavelet preprocessing for automated neural network detection
of eeg spikes, IEEE engineering in medicine and biology magazine 14 (1995) 160–166.

20
[25] R. Kohavi, G. H. John, Wrappers for feature subset selection, Artificial intelligence 97
(1997) 273–324.

[26] M. A. Hall, Correlation-based feature selection for machine learning, 1999.

[27] O. Faust, U. R. Acharya, E. Ng, T. J. Hong, W. Yu, Application of infrared thermography

in computer aided diagnosis, Infrared Physics & Technology 66 (2014) 160–175.

[28] H. Yoon, K. Yang, C. Shahabi, Feature subset selection and feature ranking for mul-
tivariate time series, IEEE transactions on knowledge and data engineering 17 (2005)
1186–1198.

[29] H. Liu, H. Motoda, Computational methods of feature selection, CRC Press, 2007.

[30] Y. Bengio, A. C. Courville, P. Vincent, Unsupervised feature learning and deep learning:
A review and new perspectives, CoRR, abs/1206.5538 1 (2012).

[31] Y. LeCun, Y. Bengio, G. Hinton, Deep learning, Nature 521 (2015) 436–444.

[32] N. Z. N. Jenny, O. Faust, W. Yu, Automated classification of normal and premature

ventricular contractions in electrocardiogram signals, Journal of Medical Imaging and
Health Informatics 4 (2014) 886–892.

[33] O. Faust, L. M. Yi, L. M. Hua, Heart rate variability analysis for different age and gender,
Journal of Medical Imaging and Health Informatics 3 (2013) 395–400.

[34] P. D. McAndrew, D. L. Potash, B. Higgins, J. Wayand, J. Held, Expert system for pro-
viding interactive assistance in solving problems such as health care management, 1996.
US Patent 5,517,405.

[35] M. Längkvist, L. Karlsson, A. Loutfi, A review of unsupervised feature learning and deep
learning for time-series modeling, Pattern Recognition Letters 42 (2014) 11–24.

[36] Y. Bengio, O. Delalleau, On the expressive power of deep architectures, in: Algorithmic
Learning Theory, Springer, pp. 18–36.

[37] I. Goodfellow, Y. Bengio, A. Courville, Deep learning, MIT press, 2016.

[38] L. Squire, D. Berg, F. E. Bloom, S. Du Lac, A. Ghosh, N. C. Spitzer, Fundamental

neuroscience, Academic Press, 2012.

[39] K. Fukushima, S. Miyake, Neocognitron: A self-organizing neural network model for a

mechanism of visual pattern recognition, in: Competition and cooperation in neural nets,
Springer, 1982, pp. 267–285.

[40] Y. LeCun, L. Bottou, Y. Bengio, P. Haffner, Gradient-based learning applied to document

recognition, Proceedings of the IEEE 86 (1998) 2278–2324.

[41] Y. LeCun, Y. Bengio, et al., Convolutional networks for images, speech, and time series,
The handbook of brain theory and neural networks 3361 (1995) 1995.

[42] K. He, X. Zhang, S. Ren, J. Sun, Delving deep into rectifiers: Surpassing human-level per-
formance on imagenet classification, in: Proceedings of the IEEE international conference
on computer vision, pp. 1026–1034.

[43] J. Bouvrie, Notes on convolutional neural networks, 2006.

21
[44] G. E. Hinton, R. R. Salakhutdinov, Reducing the dimensionality of data with neural
networks, science 313 (2006) 504–507.

[45] G. E. Hinton, S. Osindero, Y.-W. Teh, A fast learning algorithm for deep belief nets,
Neural computation 18 (2006) 1527–1554.

[46] D. P. Kingma, S. Mohamed, D. J. Rezende, M. Welling, Semi-supervised learning with

deep generative models, in: Advances in Neural Information Processing Systems, 2014,
pp. 3581–3589.

[47] J. J. Hopfield, Neural networks and physical systems with emergent collective compu-
tational abilities, in: Spin Glass Theory and Beyond: An Introduction to the Replica
Method and Its Applications, World Scientific, 1987, pp. 411–415.

[48] H. Larochelle, M. Mandel, R. Pascanu, Y. Bengio, Learning algorithms for the classification
restricted boltzmann machine, Journal of Machine Learning Research 13 (2012) 643–669.

[49] P. Smolensky, Information processing in dynamical systems: Foundations of harmony

theory, Technical Report, COLORADO UNIV AT BOULDER DEPT OF COMPUTER
SCIENCE, 1986.

[50] R. Salakhutdinov, Learning deep generative models, University of Toronto, 2009.

[51] Y. Bengio, P. Lamblin, D. Popovici, H. Larochelle, Greedy layer-wise training of deep

networks, in: Advances in neural information processing systems, pp. 153–160.

[52] S. Hochreiter, J. Schmidhuber, Long short-term memory, Neural computation 9 (1997)

1735–1780.

[53] H. Anton, I. Bivens, S. Davis, Calculus, volume 2, Wiley Hoboken, 2002.

[54] M. H. Kutner, C. Nachtsheim, J. Neter, Applied linear regression models, McGraw-

Hill/Irwin, 2004.

[55] N. J. van Eck, L. Waltman, Visualizing bibliometric networks, in: Measuring scholarly
impact, Springer, 2014, pp. 285–320.

[56] L. Waltman, N. J. van Eck, E. C. Noyons, A unified approach to mapping and clustering
of bibliometric networks, Journal of Informetrics 4 (2010) 629–635.

[57] J. V. Basmajian, C. De Luca, Muscles alive, Muscles alive: their functions revealed by
electromyography 278 (1985) 126.

[58] C. J. De Luca, The use of surface electromyography in biomechanics, Journal of applied

biomechanics 13 (1997) 135–163.

[59] C. De Luca, Myoelectric manifestation of localized muscular fatigue in humans, in: Ieee
Transactions on Biomedical Engineering, volume 30, IEEE-INST ELECTRICAL ELEC-
TRONICS ENGINEERS INC 445 HOES LANE, PISCATAWAY, NJ 08855 USA, pp.
531–531.

[60] P. Xia, J. Hu, Y. Peng, Emg-based estimation of limb movement using deep learning with
recurrent convolutional neural networks, Artificial organs (2017).

[61] X. Zhai, B. Jelfs, R. H. Chan, C. Tin, Self-recalibrating surface emg pattern recognition for
neuroprosthesis control based on convolutional neural network, Frontiers in neuroscience
11 (2017) 379.

22
[62] M. Atzori, A. Gijsberts, C. Castellini, B. Caputo, A.-G. M. Hager, S. Elsig, G. Giat-
sidis, F. Bassetto, H. Müller, Electromyography data for non-invasive naturally-controlled
robotic hand prostheses, Scientific data 1 (2014) 140053.
[63] K.-H. Park, S.-W. Lee, Movement intention decoding based on deep learning for multiuser
myoelectric interfaces, in: Brain-Computer Interface (BCI), 2016 4th International Winter
Conference on, IEEE, pp. 1–2.
[64] M. Atzori, A. Gijsberts, I. Kuzborskij, S. Elsig, A.-G. M. Hager, O. Deriaz, C. Castellini,
H. Müller, B. Caputo, Characterization of a benchmark database for myoelectric movement
classification, IEEE Transactions on Neural Systems and Rehabilitation Engineering 23
(2015) 73–83.
[65] M. Atzori, M. Cognolato, H. Müller, Deep learning with convolutional neural networks
applied to electromyography data: A resource for the classification of movements for pros-
thetic hands, Frontiers in neurorobotics 10 (2016).
[66] W. Geng, Y. Du, W. Jin, W. Wei, Y. Hu, J. Li, Gesture recognition by instantaneous
surface emg images, Scientific reports 6 (2016) 36571.
[67] M. Wand, J. Schmidhuber, Deep neural network frontend for continuous emg-based speech
recognition., in: INTERSPEECH, pp. 3032–3036.
[68] M. Wand, M. Janke, T. Schultz, The emg-uka corpus for electromyographic speech pro-
cessing, in: Fifteenth Annual Conference of the International Speech Communication
Association, pp. 1593–1597.
[69] U. C. Allard, F. Nougarou, C. L. Fall, P. Giguère, C. Gosselin, F. Laviolette, B. Gosselin,
A convolutional neural network for robotic arm guidance using semg based frequency-
features, in: Intelligent Robots and Systems (IROS), 2016 IEEE/RSJ International Con-
ference on, IEEE, pp. 2464–2470.
[70] M. Wand, T. Schultz, Pattern learning with deep neural networks in emg-based speech
recognition, in: Engineering in Medicine and Biology Society (EMBC), 2014 36th Annual
International Conference of the IEEE, IEEE, pp. 4200–4203.
[71] Q. Liu, K. Chen, Q. Ai, S. Q. Xie, recent development of signal processing algorithms for
ssvep-based brain computer interfaces, Journal of Medical and Biological Engineering 34
(2014) 299–309.
[72] Y.-F. Chen, K. Atal, S. Xie, Q. Liu, A new multivariate empirical mode decomposition
method for improving the performance of ssvep-based brain computer interface, Journal
of Neural Engineering (2017).
[73] T.-P. Jung, S. Makeig, M. Westerfield, J. Townsend, E. Courchesne, T. J. Sejnowski,
Removal of eye activity artifacts from visual event-related potentials in normal and clinical
subjects, Clinical Neurophysiology 111 (2000) 1745–1758.
[74] A. Schlögl, C. Keinrath, D. Zimmermann, R. Scherer, R. Leeb, G. Pfurtscheller, A fully
automated correction method of eog artifacts in eeg recordings, Clinical neurophysiology
118 (2007) 98–104.
[75] D. Moretti, F. Babiloni, F. Carducci, F. Cincotti, E. Remondini, P. Rossini, S. Salinari,
C. Babiloni, Computerized processing of eeg–eog–emg artifacts for multi-centric studies
in eeg oscillations and event-related potentials, International Journal of Psychophysiology
47 (2003) 199–216.

23
[76] S. Xing, R. McCardle, S. Xie, The development of eeg-based brain computer interfaces:
potential and challenges, International Journal of Computer Applications in Technology
50 (2014) 84–98.

[77] J. R. Wolpaw, D. J. McFarland, G. W. Neat, C. A. Forneris, An eeg-based brain-computer

interface for cursor control, Electroencephalography and clinical neurophysiology 78 (1991)
252–259.

[78] C. Guger, H. Ramoser, G. Pfurtscheller, Real-time eeg analysis with subject-specific

spatial patterns for a brain-computer interface (bci), IEEE transactions on rehabilitation
engineering 8 (2000) 447–456.

[79] S. C. Bunce, M. Izzetoglu, K. Izzetoglu, B. Onaral, K. Pourrezaei, Functional near-infrared

spectroscopy, IEEE engineering in medicine and biology magazine 25 (2006) 54–62.

[80] G. Huve, K. Takahashi, M. Hashimoto, Brain activity recognition with a wearable fnirs us-
ing neural networks, in: Mechatronics and Automation (ICMA), 2017 IEEE International
Conference on, IEEE, pp. 1573–1578.

[81] P. R. Jeffries, S. Woolf, B. Linde, Technology-based vs. traditional instruction: A compar-

ison of two methodsfor teaching the skill of performing a 12-lead ecg, Nursing education
perspectives 24 (2003) 70–74.

[82] A. D. Waller, A demonstration on man of electromotive changes accompanying the heart’s

beat, The Journal of physiology 8 (1887) 229–234.

[83] U. R. Acharya, O. Faust, V. Sree, G. Swapna, R. J. Martis, N. A. Kadri, J. S. Suri, Linear

and nonlinear analysis of normal and cad-affected heart rate signals, Computer methods
and programs in biomedicine 113 (2014) 55–68.

[84] U. R. Acharya, D. N. Ghista, Z. KuanYi, L. C. Min, E. Ng, S. V. Sree, O. Faust, L. Wei-

dong, A. Alvin, Integrated index for cardiac arrythmias diagnosis using entropies as
features of heart rate variability signal, in: Biomedical Engineering (MECBME), 2011 1st
Middle East Conference on, IEEE, pp. 371–374.

[85] U. R. Acharya, O. Faust, N. A. Kadri, J. S. Suri, W. Yu, Automated identification of

normal and diabetes heart rate signals using nonlinear measures, Computers in biology
and medicine 43 (2013) 1523–1529.

[86] O. Faust, V. R. Prasad, G. Swapna, S. Chattopadhyay, T.-C. Lim, Comprehensive analysis

of normal and diabetic heart rate signals: A review, Journal of Mechanics in Medicine
and Biology 12 (2012) 1240033.

[87] P. De Chazal, M. O’Dwyer, R. B. Reilly, Automatic classification of heartbeats using ecg

morphology and heartbeat interval features, IEEE Transactions on Biomedical Engineering
51 (2004) 1196–1206.

[88] U. R. Acharya, H. Fujita, S. L. Oh, Y. Hagiwara, J. H. Tan, M. Adam, Application of

deep convolutional neural network for automated detection of myocardial infarction using
ecg signals, Information Sciences 415 (2017) 190–198.

[89] A. Bulling, J. A. Ward, H. Gellersen, G. Troster, Eye movement analysis for activity
recognition using electrooculography, IEEE transactions on pattern analysis and machine
intelligence 33 (2011) 741–753.

24
 [90] R. Barea, L. Boquete, M. Mazo, E. López, System for assisted mobility using eye move-
ments based on electrooculography, IEEE transactions on neural systems and rehabilita-
tion engineering 10 (2002) 209–218.

[91] A. R. Kherlopian, J. P. Gerrein, M. Yue, K. E. Kim, J. W. Kim, M. Sukumaran, P. Sajda,

Electrooculogram based system for computer control using a multiple feature classification
model, in: Engineering in Medicine and Biology Society, 2006. EMBS’06. 28th Annual
International Conference of the IEEE, IEEE, pp. 1295–1298.

[92] O. Faust, Documenting and predicting topic changes in computers in biology and medicine:
A bibliometric keyword analysis from 1990 to 2017, Informatics in Medicine Unlocked 11
(2018) 15 – 27.

[93] L. Fraiwan, K. Lweesy, Neonatal sleep state identification using deep learning autoen-
coders, in: Signal Processing & its Applications (CSPA), 2017 IEEE 13th International
Colloquium on, IEEE, pp. 228–231.

[94] A. Piryatinska, G. Terdik, W. A. Woyczynski, K. A. Loparo, M. S. Scher, A. Zlotnik,

Automated detection of neonate eeg sleep stages, Computer methods and programs in
biomedicine 95 (2009) 31–46.

[95] R. T. Schirrmeister, J. T. Springenberg, L. D. J. Fiederer, M. Glasstetter, K. Eggensperger,

M. Tangermann, F. Hutter, W. Burgard, T. Ball, Deep learning with convolutional neural
networks for eeg decoding and visualization, Human brain mapping 38 (2017) 5391–5420.

[96] C. Brunner, R. Leeb, G. Müller-Putz, A. Schlögl, G. Pfurtscheller, Bci competition 2008–

graz data set a, Institute for Knowledge Discovery (Laboratory of Brain-Computer Inter-
faces), Graz University of Technology 16 (2008).

[97] M.-P. Hosseini, T. X. Tran, D. Pompili, K. Elisevich, H. Soltanian-Zadeh, Deep learning

with edge computing for localization of epileptogenicity using multimodal rs-fmri and eeg
big data, in: Autonomic Computing (ICAC), 2017 IEEE International Conference on,
IEEE, pp. 83–92.

[98] M. Stead, M. Bower, B. H. Brinkmann, K. Lee, W. R. Marsh, F. B. Meyer, B. Litt,

J. Van Gompel, G. A. Worrell, Microseizures and the spatiotemporal scales of human
partial epilepsy, Brain 133 (2010) 2789–2797.

[99] R. T. Schirrmeister, L. D. J. Fiederer, J. T. Springenberg, M. Glasstetter, K. Eggensperger,

M. Tangermann, F. Hutter, W. Burgard, T. Ball, Designing and understanding convo-
lutional networks for decoding executed movements from eeg, in: The First Biannual
Neuroadaptive Technology Conference, p. 143.

[100] M. J. van Putten, J. Hofmeijer, B. J. Ruijter, M. C. Tjepkema-Cloostermans, Deep

learning for outcome prediction of postanoxic coma, in: EMBEC & NBC 2017, Springer,
2017, pp. 506–509.

[101] C. Spampinato, S. Palazzo, I. Kavasidis, D. Giordano, N. Souly, M. Shah, Deep learning

human mind for automated visual classification, in: 2017 IEEE Conference on Computer
Vision and Pattern Recognition (CVPR), pp. 4503–4511.

[102] I. Kiral-Kornek, D. Mendis, E. S. Nurse, B. S. Mashford, D. R. Freestone, D. B. Gray-

den, S. Harrer, Truenorth-enabled real-time classification of eeg data for brain-computer
interfacing, in: Engineering in Medicine and Biology Society (EMBC), 2017 39th Annual
International Conference of the IEEE, IEEE, pp. 1648–1651.

25
[103] U. R. Acharya, S. L. Oh, Y. Hagiwara, J. H. Tan, H. Adeli, Deep convolutional neural
network for the automated detection and diagnosis of seizure using eeg signals, Computers
in Biology and Medicine (2017).

[104] M. Winterhalder, T. Maiwald, H. Voss, R. Aschenbrenner-Scheibe, J. Timmer, A. Schulze-

Bonhage, The seizure prediction characteristic: a general framework to assess and compare
seizure prediction methods, Epilepsy & Behavior 4 (2003) 318–325.

[105] R. Aschenbrenner-Scheibe, T. Maiwald, M. Winterhalder, H. Voss, J. Timmer, A. Schulze-

Bonhage, How well can epileptic seizures be predicted? an evaluation of a nonlinear
method, Brain 126 (2003) 2616–2626.

[106] T. Maiwald, M. Winterhalder, R. Aschenbrenner-Scheibe, H. U. Voss, A. Schulze-Bonhage,

J. Timmer, Comparison of three nonlinear seizure prediction methods by means of the
seizure prediction characteristic, Physica D: nonlinear phenomena 194 (2004) 357–368.

[107] B. Schelter, M. Winterhalder, T. Maiwald, A. Brandt, A. Schad, A. Schulze-Bonhage,

J. Timmer, Testing statistical significance of multivariate time series analysis techniques
for epileptic seizure prediction, Chaos: An Interdisciplinary Journal of Nonlinear Science
16 (2006) 013108.

[108] B. Schelter, M. Winterhalder, T. Maiwald, A. Brandt, A. Schad, J. Timmer, A. Schulze-

Bonhage, Do false predictions of seizures depend on the state of vigilance? a report from
two seizure-prediction methods and proposed remedies, Epilepsia 47 (2006) 2058–2070.

[109] N. Lu, T. Li, X. Ren, H. Miao, A deep learning scheme for motor imagery classification
based on restricted boltzmann machines, IEEE Transactions on Neural Systems and
Rehabilitation Engineering 25 (2017) 566–576.

[110] R. Leeb, F. Lee, C. Keinrath, R. Scherer, H. Bischof, G. Pfurtscheller, Brain–computer

communication: motivation, aim, and impact of exploring a virtual apartment, IEEE
Transactions on Neural Systems and Rehabilitation Engineering 15 (2007) 473–482.

[111] P. Sajda, A. Gerson, K.-R. Muller, B. Blankertz, L. Parra, A data analysis competi-
tion to evaluate machine learning algorithms for use in brain-computer interfaces, IEEE
Transactions on neural systems and rehabilitation engineering 11 (2003) 184–185.

[112] B. Blankertz, K.-R. Muller, G. Curio, T. M. Vaughan, G. Schalk, J. R. Wolpaw, A. Schlogl,

C. Neuper, G. Pfurtscheller, T. Hinterberger, et al., The bci competition 2003: progress
and perspectives in detection and discrimination of eeg single trials, IEEE transactions
on biomedical engineering 51 (2004) 1044–1051.

[113] M. Hajinoroozi, Z. Mao, T.-P. Jung, C.-T. Lin, Y. Huang, Eeg-based prediction of driver’s
cognitive performance by deep convolutional neural network, Signal Processing: Image
Communication 47 (2016) 549–555.

[114] E. Nurse, B. S. Mashford, A. J. Yepes, I. Kiral-Kornek, S. Harrer, D. R. Freestone, De-

coding eeg and lfp signals using deep learning: heading truenorth, in: Proceedings of the
ACM International Conference on Computing Frontiers, ACM, pp. 259–266.

[115] L. Jingwei, C. Yin, Z. Weidong, Deep learning eeg response representation for brain
computer interface, in: Control Conference (CCC), 2015 34th Chinese, IEEE, pp. 3518–
3523.

26
[116] H. Piroska, S. Janos, Specific movement detection in eeg signal using time-frequency
analysis, in: Complexity and Intelligence of the Artificial and Natural Complex Systems,
Medical Applications of the Complex Systems, Biomedical Computing, 2008. CANS’08.
First International Conference on, IEEE, pp. 209–215.

[117] M. Hajinoroozi, Z. Mao, Y. Huang, Prediction of driver’s drowsy and alert states from
eeg signals with deep learning, in: Computational Advances in Multi-Sensor Adaptive
Processing (CAMSAP), 2015 IEEE 6th International Workshop on, IEEE, pp. 493–496.

[118] S. Jirayucharoensak, S. Pan-Ngum, P. Israsena, Eeg-based emotion recognition using deep

learning network with principal component based covariate shift adaptation, The Scientific
World Journal 2014 (2014).

[119] S. Koelstra, C. Muhl, M. Soleymani, J.-S. Lee, A. Yazdani, T. Ebrahimi, T. Pun, A. Ni-
jholt, I. Patras, Deap: A database for emotion analysis; using physiological signals, IEEE
Transactions on Affective Computing 3 (2012) 18–31.

[120] X. An, D. Kuang, X. Guo, Y. Zhao, L. He, A deep learning method for classification of
eeg data based on motor imagery, in: International Conference on Intelligent Computing,
Springer, pp. 203–210.

[121] W.-L. Zheng, J.-Y. Zhu, Y. Peng, B.-L. Lu, Eeg-based emotion classification using deep
belief networks, in: Multimedia and Expo (ICME), 2014 IEEE International Conference
on, IEEE, pp. 1–6.

[122] J. Li, A. Cichocki, Deep learning of multifractal attributes from motor imagery induced
eeg, in: International Conference on Neural Information Processing, Springer, pp. 503–510.

[123] X. Jia, K. Li, X. Li, A. Zhang, A novel semi-supervised deep learning framework for
affective state recognition on eeg signals, in: Bioinformatics and Bioengineering (BIBE),
2014 IEEE International Conference on, IEEE, pp. 30–37.

[124] Y. Ren, Y. Wu, Convolutional deep belief networks for feature extraction of eeg signal, in:
Neural Networks (IJCNN), 2014 International Joint Conference on, IEEE, pp. 2850–2853.

[125] S. Ahmed, L. M. Merino, Z. Mao, J. Meng, K. Robbins, Y. Huang, A deep learning

method for classification of images rsvp events with eeg data, in: Global Conference on
Signal and Information Processing (GlobalSIP), 2013 IEEE, IEEE, pp. 33–36.

[126] P. W. Mirowski, Y. LeCun, D. Madhavan, R. Kuzniecky, Comparing svm and convolu-

tional networks for epileptic seizure prediction from intracranial eeg, in: Machine Learning
for Signal Processing, 2008. MLSP 2008. IEEE Workshop on, IEEE, pp. 244–249.

[127] H. Cecotti, A. Graeser, Convolutional neural network with embedded fourier transform
for eeg classification, in: Pattern Recognition, 2008. ICPR 2008. 19th International Con-
ference on, IEEE, pp. 1–4.

[128] U. R. Acharya, H. Fujita, O. S. Lih, Y. Hagiwara, J. H. Tan, M. Adam, Automated detec-

tion of arrhythmias using different intervals of tachycardia ecg segments with convolutional
neural network, Information sciences 405 (2017) 81–90.

[129] A. L. Goldberger, L. A. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov, R. G. Mark,

J. E. Mietus, G. B. Moody, C.-K. Peng, H. E. Stanley, Physiobank, physiotoolkit, and
physionet, Circulation 101 (2000) e215–e220.

27
[130] J. H. Tan, Y. Hagiwara, W. Pang, I. Lim, S. L. Oh, M. Adam, R. San Tan, M. Chen, U. R.
Acharya, Application of stacked convolutional and long short-term memory network for
accurate identification of cad ecg signals, Computers in Biology and Medicine (2018).

[131] U. R. Acharya, H. Fujita, S. L. Oh, U. Raghavendra, J. H. Tan, M. Adam, A. Gertych,

Y. Hagiwara, Automated identification of shockable and non-shockable life-threatening
ventricular arrhythmias using convolutional neural network, Future Generation Computer
Systems (2017).

[132] B. Pourbabaee, M. J. Roshtkhari, K. Khorasani, Deep convolutional neural networks and

learning ecg features for screening paroxysmal atrial fibrillation patients, IEEE Transac-
tions on Systems, Man, and Cybernetics: Systems (2017).

[133] G. Moody, A. Goldberger, S. McClennen, S. Swiryn, Predicting the onset of paroxys-

mal atrial fibrillation: The computers in cardiology challenge 2001, in: Computers in
Cardiology 2001, IEEE, pp. 113–116.

[134] G. Zheng, S. Ji, M. Dai, Y. Sun, Ecg based identification by deep learning, in: Chinese
Conference on Biometric Recognition, Springer, pp. 503–510.

[135] G. B. Moody, R. G. Mark, The impact of the mit-bih arrhythmia database, IEEE
Engineering in Medicine and Biology Magazine 20 (2001) 45–50.

[136] A. Majumdar, R. Ward, Robust greedy deep dictionary learning for ecg arrhythmia
classification, in: Neural Networks (IJCNN), 2017 International Joint Conference on,
IEEE, pp. 4400–4407.

[137] S. P. Shashikumar, A. J. Shah, Q. Li, G. D. Clifford, S. Nemati, A deep learning approach

to monitoring and detecting atrial fibrillation using wearable technology, in: Biomedical
& Health Informatics (BHI), 2017 IEEE EMBS International Conference on, IEEE, pp.
141–144.

[138] K. Luo, J. Li, Z. Wang, A. Cuschieri, Patient-specific deep architectural model for ecg
classification, Journal of Healthcare Engineering 2017 (2017).

[139] U. R. Acharya, H. Fujita, M. Adam, O. S. Lih, V. K. Sudarshan, T. J. Hong, J. E. Koh,

Y. Hagiwara, C. K. Chua, C. K. Poo, et al., Automated characterization and classification
of coronary artery disease and myocardial infarction by decomposition of ecg signals: a
comparative study, Information Sciences 377 (2017) 17–29.

[140] S. D. Greenwald, The development and analysis of a ventricular fibrillation detector, Ph.D.
thesis, Massachusetts Institute of Technology, 1986.

[141] F. M. Nolle, F. K. Badura, J. M. Catlett, R. W. Bowser, S. M. H, Crei-gard, a new

concept in computerized arrhythmia monitoring systems, Computers in Cardiology 63
(1986) 515–518.

[142] U. R. Acharya, S. L. Oh, Y. Hagiwara, J. H. Tan, M. Adam, A. Gertych, R. San Tan,

A deep convolutional neural network model to classify heartbeats, Computers in Biology
and Medicine 89 (2017) 389–396.

[143] M. Cheng, W. J. Sori, F. Jiang, A. Khan, S. Liu, Recurrent neural network based classi-
fication of ecg signal features for obstruction of sleep apnea detection, in: Computational
Science and Engineering (CSE) and Embedded and Ubiquitous Computing (EUC), 2017
IEEE International Conference on, volume 2, IEEE, pp. 199–202.

28
[144] T. Penzel, G. B. Moody, R. G. Mark, A. L. Goldberger, J. H. Peter, The apnea-ecg
database, in: Computers in cardiology 2000, IEEE, pp. 255–258.

[145] B. Taji, A. D. Chan, S. Shirmohammadi, Classifying measured electrocardiogram signal

quality using deep belief networks, in: Instrumentation and Measurement Technology
Conference (I2MTC), 2017 IEEE International, IEEE, pp. 1–6.

[146] P. R. Muduli, R. R. Gunukula, A. Mukherjee, A deep learning approach to fetal-ecg signal

reconstruction, in: Communication (NCC), 2016 Twenty Second National Conference on,
IEEE, pp. 1–6.

[147] S. Kiranyaz, T. Ince, M. Gabbouj, Real-time patient-specific ecg classification by 1-d

convolutional neural networks, IEEE Transactions on Biomedical Engineering 63 (2016)
664–675.

[148] Y. Zheng, Q. Liu, E. Chen, Y. Ge, J. L. Zhao, Time series classification using multi-
channels deep convolutional neural networks, in: International Conference on Web-Age
Information Management, Springer, pp. 298–310.

[149] L.-H. Du, W. Liu, W.-L. Zheng, B.-L. Lu, Detecting driving fatigue with multimodal deep
learning, in: Neural Engineering (NER), 2017 8th International IEEE/EMBS Conference
on, IEEE, pp. 74–77.

[150] J. Zhang, S. Li, R. Wang, Pattern recognition of momentary mental workload based
on multi-channel electrophysiological data and ensemble convolutional neural networks,
Frontiers in Neuroscience 11 (2017).

[151] B. Xia, Q. Li, J. Jia, J. Wang, U. Chaudhary, A. Ramos-Murguialday, N. Birbaumer,

Electrooculogram based sleep stage classification using deep belief network, in: Neural
Networks (IJCNN), 2015 International Joint Conference on, IEEE, pp. 1–5.

[152] B. Kemp, A. H. Zwinderman, B. Tuk, H. A. Kamphuisen, J. J. Oberye, Analysis of a

sleep-dependent neuronal feedback loop: the slow-wave microcontinuity of the eeg, IEEE
Transactions on Biomedical Engineering 47 (2000) 1185–1194.

[153] X. Zhu, W.-L. Zheng, B.-L. Lu, X. Chen, S. Chen, C. Wang, Eog-based drowsiness
detection using convolutional neural networks., in: IJCNN, pp. 128–134.

[154] M. Längkvist, L. Karlsson, A. Loutfi, Sleep stage classification using unsupervised feature
learning, Advances in Artificial Neural Systems 2012 (2012) 5.

[155] S. Min, B. Lee, S. Yoon, Deep learning in bioinformatics, Briefings in bioinformatics

(2016) bbw068.

[156] J. Dean, G. Corrado, R. Monga, K. Chen, M. Devin, M. Mao, A. Senior, P. Tucker,

K. Yang, Q. V. Le, et al., Large scale distributed deep networks, in: Advances in neural
information processing systems, pp. 1223–1231.

[157] S. Tokui, K. Oono, S. Hido, J. Clayton, Chainer: a next-generation open source framework
for deep learning, in: Proceedings of workshop on machine learning systems (LearningSys)
in the twenty-ninth annual conference on neural information processing systems (NIPS),
volume 5, pp. 1–6.

[158] K. Chatfield, K. Simonyan, A. Vedaldi, A. Zisserman, Return of the devil in the details:
Delving deep into convolutional nets, arXiv preprint arXiv:1405.3531 (2014).

29
[159] D. Erhan, Y. Bengio, A. Courville, P.-A. Manzagol, P. Vincent, S. Bengio, Why does
unsupervised pre-training help deep learning?, Journal of Machine Learning Research 11
(2010) 625–660.

[160] X.-W. Chen, X. Lin, Big data deep learning: challenges and perspectives, IEEE access 2
(2014) 514–525.

[161] M. M. Najafabadi, F. Villanustre, T. M. Khoshgoftaar, N. Seliya, R. Wald,

E. Muharemagic, Deep learning applications and challenges in big data analytics, Journal
of Big Data 2 (2015) 1.

[162] J. Bergstra, F. Bastien, O. Breuleux, P. Lamblin, R. Pascanu, O. Delalleau, G. Desjardins,

D. Warde-Farley, I. Goodfellow, A. Bergeron, et al., Theano: Deep learning on gpus with
python, in: NIPS 2011, BigLearning Workshop, Granada, Spain, volume 3, Microtome
Publishing., pp. 1–48.

[163] C. Schaffer, Selecting a classification method by cross-validation, Machine Learning 13

(1993) 135–143.

[164] B. Zadrozny, Learning and evaluating classifiers under sample selection bias, in: Proceed-
ings of the twenty-first international conference on Machine learning, ACM, p. 114.

[165] J. Huang, A. Gretton, K. M. Borgwardt, B. Schölkopf, A. J. Smola, Correcting sample

selection bias by unlabeled data, in: Advances in neural information processing systems,
pp. 601–608.

[166] J.-G. Lee, S. Jun, Y.-W. Cho, H. Lee, G. B. Kim, J. B. Seo, N. Kim, Deep learning in
medical imaging: general overview, Korean journal of radiology 18 (2017) 570–584.

Appendix A. Cluster description

.
Table A.8: Author supplied keywords mapped to topics

Topic Author supplied keywords

arrhythmia arrhythmia, arrhythmic signal identification
cvd cardiovascular diseases, myocardial infarction, non-shockable, shockable,
ventricular arrhythmias
db mit-bih arrhythmia database, physiobank mit-bih arrhythmia database
dnn deep neural network, deep neural networks
ecg apnea- electrocardiogram signals, ecg, ecg measurement system, ecg
signal, ecg signals, electrocardiogram, electrocardiogram signal quality
classification, electrocardiogram signals, electrocardiography
heartbeat heartbeat, heartbeat interval features, rr interval
measurement computerised instrumentation, pollution measurement, power line inter-
ference
noise denoising autoencoder, noise, noise measurement, snr
testing testing
training training, training data
wearable wearable computers, wearable electrocardiogram measurement system,
wearable fnirs

30
algorithm ada-boost, algorithm, algorithms, classification algorithms, component
analysis, compression, hidden markov models, independent component
analysis
dbn belief networks, boltzmann machines, convolutional deep belief net-
works, dbn, deep belief network, rbm, restricted boltzman machine
(rbm), restricted boltzmann machine, deep neural network, deep neural
networks
eeg brain signals, eeg, eeg data, eeg-based hand squeeze task, electroen-
cephalography, electroencephalography (eeg), encephalogram signals
feature feature clustering, feature extraction, feature learning, feature-
extraction, features, unsupervised feature learning
learning deep feature learning, deep learning, ensemble learning
machine machine, machines
model mathematical model, model, models
motor imagery motor imagery
sleep apnea detection, automatic sleep stage classification, obstructive sleep
apnea detection, sleep apnea, sleep stage classification, slow eye-
movements
classification classification, emotion classification, patient-specific ecg classification,
pattern classification
control multifunction myoelectric control, myoelectric control
emg electromyogram, electromyography, emg-based speech recognition, non-
stationary emg, surface emg
ml machine learning, machine learning algorithms
prostheses prostheses, prosthetics, upper-limb, upper-limb prostheses
recognition brain activity recognition, pattern recognition, pattern-recognition,
recognition, signal recognition
signal medical signal detection, medical signal processing, signal classification,
signal reconstruction, signal representation, signals, time preserving sig-
nal representation strategy
bci bci brain computer interface, brain computer interface (bci), brain
computer interfaces, brain-computer interface, brain-computer interface
(bci), brain-computer interfaces
cnn cnn, convolution, convolution neural network, convolutional neural net-
work, convolutional neural networks, convolutional neural networks
(cnns)
epilepsy epilepsy, seizure
nn batch normalization layers, biological neural networks, feedforward neu-
ral nets, neural-networks, neurons
real-time real-time, real-time heart monitoring, real-time systems, truenorth-
enabled real-time classification
system ibm truenorth neurosynaptic system, low-power platform, system

31