Future Generation Computer Systems 114 (2021) 23–43

Contents lists available at ScienceDirect

Future Generation Computer Systems

journal homepage: www.elsevier.com/locate/fgcs

An intelligent healthcare monitoring framework using wearable

sensors and social networking data
∗
Farman Ali a ,1 , Shaker El-Sappagh b,c ,1 , S.M. Riazul Islam d , Amjad Ali e , ,
∗
Muhammad Attique a , Muhammad Imran f , Kyung-Sup Kwak g ,
a
Department of Software, Sejong University, South Korea
b
Centro Singular de Investigación en Tecnoloxías Intelixentes (CiTIUS), Universidade de Santiago de Compostela, Santiago de Compostela, Spain
c
Information Systems Department, Faculty of Computers and Artificial Intelligence, Benha University, Banha, Egypt
d
Department of Computer Science and Engineering, Sejong University, Seoul, South Korea
e
Department of Computer Science, COMSATS University Islamabad, Lahore Campus, Lahore, Pakistan
f
College of Applied Computer Science, King Saud University, Riyadh, Saudi Arabia
g
Department of Information and Communication Engineering, Inha University, Incheon, South Korea

article info a b s t r a c t

Article history: Wearable sensors and social networking platforms play a key role in providing a new method
Received 17 June 2019 to collect patient data for efficient healthcare monitoring. However, continuous patient monitoring
Received in revised form 24 June 2020 using wearable sensors generates a large amount of healthcare data. In addition, the user-generated
Accepted 21 July 2020
healthcare data on social networking sites come in large volumes and are unstructured. The existing
Available online 24 July 2020
healthcare monitoring systems are not efficient at extracting valuable information from sensors and
Keywords: social networking data, and they have difficulty analyzing it effectively. On top of that, the traditional
Machine learning machine learning approaches are not enough to process healthcare big data for abnormality prediction.
Semantic knowledge Therefore, a novel healthcare monitoring framework based on the cloud environment and a big data
Big data analysis analytics engine is proposed to precisely store and analyze healthcare data, and to improve the
Healthcare monitoring system classification accuracy. The proposed big data analytics engine is based on data mining techniques,
Wearable sensors ontologies, and bidirectional long short-term memory (Bi-LSTM). Data mining techniques efficiently
Social network analysis
preprocess the healthcare data and reduce the dimensionality of the data. The proposed ontologies
provide semantic knowledge about entities and aspects, and their relations in the domains of diabetes
and blood pressure (BP). Bi-LSTM correctly classifies the healthcare data to predict drug side effects
and abnormal conditions in patients. Also, the proposed system classifies the patients’ health condition
using their healthcare data related to diabetes, BP, mental health, and drug reviews. This framework
is developed employing the Protégé Web Ontology Language tool with Java. The results show that the
proposed model precisely handles heterogeneous data and improves the accuracy of health condition
classification and drug side effect predictions.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction diabetes and BP can provide an opportunity to secure the patient

from cardiovascular diseases.
Diabetes and abnormal blood pressure (BP) are the two most Due to the advancements in information technology, all the
common and extremely dangerous diseases that affect the func- devices in the healthcare industry have been digitalized. These
tionality of the human body, and increase the risk of cardio- digital devices enable the living better and more comfortable.
vascular diseases. Early detection and classification of diabetes Therefore, people use various devices, such as smartphones and
and BP into their specific category is therefore essential to treat wearable sensors, in their daily lives. Smartphones can contain
the patient efficiently. In addition, accurate monitoring of the sensors that can be used to obtain information about the human
body [1]. Wearables can be utilized to collect a huge amount
∗ Corresponding authors. of patient vitals [2]. Both of these devices can be utilized to
real-time monitor the patient. To date, numerous healthcare sys-
E-mail addresses: [email protected] (F. Ali),
[email protected] (S. El-Sappagh), [email protected] (S.M.R. Islam),
tems have been proposed to monitor the diabetes and BP pa-
[email protected] (A. Ali), [email protected] (M. Attique), tients using smartphone and wearable sensors [3–10]. However,
[email protected] (M. Imran), [email protected] (K.S. Kwak). these systems are not well-equipped to collect the data in real
1 These authors contributed equally to this work and co-first authors. time. Furthermore, the digital devices generate a huge amount of

https://doi.org/10.1016/j.future.2020.07.047
0167-739X/© 2020 Elsevier B.V. All rights reserved.
24 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43

healthcare data and the existing systems are incapable of storing data and improves the accuracy of patient health condition clas-
and processing it efficiently for precise monitoring of patients. sification. The main contributions of our research work are as
Moreover, extracting valuable information from healthcare data follows.
and effectively analyzing them has become a new challenge for
the existing healthcare monitoring systems. The generated data • A novel framework is introduced that extracts large amounts
from these devices are unstructured and are therefore difficult to of the most useful healthcare data from various sources,
handle for chronic patient monitoring. In addition, the amount of such as smartphones, wearable sensors, medical records,
data is exponentially increasing, which requires a huge storage and social networks. In addition, a big data cloud repository
space. Therefore, a smart methodology and a cloud-based health- is utilized to store the extracted data, and MapReduce is
care architecture are required to accurately monitor diabetes and applied to intelligently handle and process structured and
BP patients, store their healthcare data, and perform predictive unstructured data.
analysis on the structured and unstructured data. • A big data analytics engine is proposed for the analysis of
Recently, the use of social networking in the healthcare indus- real-world big data. It is used to accurately handle health-
try has been rapidly increasing. The social network data can also care data containing inconsistencies and that have missing
be utilized to identify various factors such as emotional status values, noise, different formats, a large size, and high dimen-
and accrued stress, which might be translated into the status of a sionality. In addition, it is utilized to improve the quality of
patient health. People with diabetes and abnormal BP share their data processing and to save time. The proposed big data an-
emotions and experiences with others on social networking sites. alytics engine uses artificial intelligence (AI) approaches to
They share valuable information and motivate each other to fight extract useful features from big data that eventually reduce
against diabetes and high/low BP. In addition, diabetes patients the dimensionality of data.
publish their opinions about specific drugs. A new patient sees • A neural network-based word embedding model called
the opinions of others and responds to them about the same Word2vec is used to represent healthcare textual data with
drugs. Therefore, the healthcare monitoring systems for diabetes semantic meaning. In addition, specific domain ontologies
and abnormal BP need social networking data in order to identify are integrated with the Word2vec model. These ontologies
emotional disturbances in patients by using their posts, and to provide additional information for a neural network model
monitor drug side effects by using drug reviews. However, the that understands the semantic meaning of unusual words.
information on social networking sites about patient emotions A novel semantic knowledge with the Bi-LSTM model is
and drug experiences are unstructured and unexpected, and it used to precisely classify the unstructured and structured
would be a challenging task for healthcare monitoring systems healthcare data.
to extract the information and analyze it in order to monitor the • Several experiments are conducted using principal com-
patient’s mental health and to predict drug side effects. Therefore, ponent analysis (PCA) and information gain (IG) with the
there is a need of smart approaches that can automatically extract ontology and LSTM-based models, and the results are com-
the most relevant features, and reduce the dimensionality of the pared with the respective reference models. This compari-
datasets for better accuracy of healthcare monitoring system. son helps us to determine the advantages and limitations of
In recent years, machine learning (ML) techniques, such as the applied approaches and classification models.
the decision tree, the support vector machine (SVM), k-nearest
neighbors (KNN), fuzzy logic, and multi-layer perceptron (MLP), The rest of this paper is structured as follows. Section 2
are used to monitor diabetes and BP patients and provide compe- presents discussions of healthcare monitoring using wearable
tent treatments [11–14]. However, continuous patient monitoring sensors and ML approaches, healthcare monitoring systems based
generates a large amount of healthcare data, such as sensor on social networking data and ML approaches, and healthcare
readings, patient profiles, medical records, lab tests, and physician big data and ML approaches. Section 3 presents the proposed
notes. Both healthcare and social networking data have greatly in- framework and describes its various modules. Section 4 provides
creased in a few years, which is called big data (both unstructured the experimental results. Finally, Section 5 concludes our work.
and structured). The traditional approaches and ML techniques
may not handle these data very well for the extraction of mean- 2. Related work
ingful information and for abnormality prediction. In addition,
these data may not help the healthcare industry until they are Machine learning techniques and big data play a key role in
processed intelligently in real time. This necessitates a big data the healthcare monitoring systems of chronic patients. Due to
cloud platform and an advanced deep learning approach, such as the rapid increase in the usage of wearable sensors and social
long short-term memory (LSTM). networking data in the healthcare domain, this section looks
New advancements in the healthcare monitoring systems are at healthcare monitoring of diabetes and BP patients based on
still a big challenge because they need huge multidisciplinary wearable sensors and ML approaches, at healthcare monitoring
steps. With the new technologies in healthcare sectors and the systems based on social networking data and ML approaches, and
changes in society, traditional systems are not efficient enough at healthcare big data and ML approaches.
for these new conditions. They need a new framework based
on new methods to solve the problems. However, traditional 2.1. ML approaches and healthcare monitoring of diabetes and BP
techniques can be used together with a new system. patients using wearable sensors
In this paper, we propose an advanced healthcare monitor-
ing framework based on ontologies and on bidirectional long There have been various studies into wearable sensor-based
short-term memory (Bi-LSTM) to precisely analyze healthcare big physiological information extraction and the healthcare moni-
data in order to improve classification accuracy. The proposed toring of chronically ill patients. These studies introduced vari-
framework integrates different sources of information for effi- ous methods to analyze wearable sensor data [15–17]. However,
cient healthcare monitoring of chronic patients. It was applied to wearable sensors-based collected data are in large volume and
the classification of patients using their healthcare data related unstructured. These methods do not consider the advanced data
to diabetes, BP, mental health, and drug reviews. The results analysis approaches to discover useful information in the sensors
prove that the proposed model correctly handles heterogeneous data. Thus, they may not monitor the patient body effectively.
 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43 25

Wearable sensors and an ML-based personalized healthcare mon- 2.2. Healthcare monitoring systems based on social networking data
itoring system was presented for diabetes patient monitoring [5]. and ML approaches
This system utilizes Bluetooth Low Energy-based sensors to col-
lect data. The authors then applied multi-layer perceptron and Recently, researchers have been utilizing social networking
LSTM to classify the type of diabetes and to predict the glucose data and drug information in the field of healthcare monitoring
level of the input user. However, the diabetes dataset used in the systems, which are extensively discussed in this section [12,24,
existing system was limited to Pima Indian woman. Therefore, 25]. An emotional healthcare system was presented to detect
it may not be possible to use the dataset with robust prediction psychological disturbances in patients [26]. The authors utilized
model for different purposes. A framework based on ML methods patient messages published on social networking platforms and
and a Hadoop environment was proposed for diabetes predic- identified depression and stress levels. They applied the con-
tion [18]. In that work, the authors proposed an IG algorithm for volution neural network (CNN), the recurrent neural network
feature extraction. In addition, naive Bayes, the decision tree, and (RNN), and Bi-LSTM to detect stressful and depressive content.
random forest were utilized to predict diabetes. However, the ac- In addition, they proposed an ontology-based recommendation
system that sends text to the patient based on results from mon-
curacy of these classifiers largely decreased when the amount of
itoring. However, the continuous monitoring of patients based
data increased. This is because they did not use dimensionality re-
on their social networking data is a difficult task. This needs
duction. A 5G smart diabetes system based on wearable sensors,
effective statistical analysis, data normalization, feature selec-
big data, and ML was presented for personalized diabetes diag-
tion, and ML models to correctly predict the patient mood. Deep
nosis [12]. The system has five main goals (smartness, person-
learning-based sentiment analysis was presented for perinatal
alization, comfortability, effectiveness, and sustainability), which
depression [27]. These authors utilized WeChat friends’-circle
help patients detect diabetes early, and it provides personalized
data for LSTM training, and used emoticons as feature extrac-
treatment solutions. This model is costly in terms of implementa- tion to monitor perinatal depression in patients. This system
tion and lacks a personalized data analysis method. A wearable- abbreviated the screening time and decreased the costs of doctor–
technologies and Internet of Things (IoTs)-based recommendation patient communication. This method does not consider the sen-
system was proposed for proactive healthcare monitoring [19]. timent analysis of textual data which is potential to improve
The system solved two main problems in health monitoring: the performance of perinatal depression monitoring. An artificial
the identification of factors that must be monitored, and the intelligence approach was proposed to analyze patient posts to
identification of wearable technology that must be used for the healthcare social networking platforms and to identify critical
measurement of the factors. However, this system recommends problems in patients [28]. The system applied the preprocess-
tools or technologies using manual approaches, which is time- ing methods of text and ML classifiers to automatically analyze
consuming and may not suit with the ongoing trend of healthcare patient posts and inform their doctors when needed. The posts
technology. An LSTM-based model was presented to effectively on social networks are always unstructured. It is impossible to
mine sensor data and lab tests [20]. The authors trained LSTM precisely handle without using deep learning algorithms such as
and MLP techniques by utilizing different attributes of the inten- word embeddings and n-grams. A novel system was proposed
sive care unit (ICU) patient, including body temperature, systolic to identify diabetes risk based on social networking activities on
and diastolic blood pressure, blood glucose, and heart rate. The Twitter [29,30]. The authors presented a new approach to data
authors then verified the utility of these classifiers and compared generation, and introduced ML methods to classify the users’
their performance. However, this LSTM-based system failed to diabetes risk based on their Twitter activities. Both the systems
directly handle irregular sampling and missing values. It needs have utilized insufficient information for diabetes risk predic-
indicator variables to differentiate actual from missing values. tion, which may generate inaccurate results and thus perhaps
A framework for big data in Mauritius was proposed for dia- misguide the healthcare professionals. A Bi-LSTM and gated re-
betes [21]. That work presented the patient’s current diabetes current unit was utilized to intelligently process social media
status and solutions for diabetes management. In addition, the comments of a patient undergoing a medical cure [31]. These
authors also discussed the challenges of big data in the healthcare authors focused on patient remarks about drugs, and discov-
domain. A bi-directional LSTM was proposed to predict the future ered disease-related medical concepts from them. This system
level of blood glucose [22]. In that system, the authors compared considered different types of text to discover disease (e.g., a
depressive disorder was discovered in patients using text like
the results from simple LSTM with Bi-LSTM using 26 datasets
‘‘woke up too early’’). Feature-based sentiment analysis of drug
from 20 real patients. However, the system lacks features that
reviews was presented to detect adverse drugs reactions [32].
decrease the performance of diabetes prediction. In addition,
This system performs different tasks on drug reviews that identify
there is no alarm mechanism to predict the up-coming blood-
the effectiveness and side effects of the drugs. However, deep
glucose level. The authors presented a review of published work
learning approaches with more sophisticated features can im-
from 2001 to 2017 that summarized the research topics related to
prove the achieved results. A sentiment feature-based system
wearable sensor technology [23]. In the review paper, the authors was proposed to detect adverse drug reaction posts on social
discussed frameworks that covered various topics, including data networks [33]. The authors utilized a large amount of text and
gathering, data processing, and system responses. In addition, conducted experiments in order to develop a real approach to
the authors looked at the gap between wearable devices and identifying posts related to adverse drug reactions. In this system,
human factors. A knowledge discovery approach was proposed there is no data analysis method to efficiently process the data.
for healthcare assistance [14]. The system simplified the analysis Also, the system lacks semantic similarity approaches to precisely
of big data in the cloud environment, and the authors illustrated confirm the identification of drug-related posts.
a new method that determines the classifications to predict un- An ontology-based, feature-level, sentiment analysis system
usual conditions in blood pressure patients. However, the system was proposed in the domain of diabetes treatment [34]. The
utilized traditional approaches and may solve only specific cases. authors used the ontology to provide semantic relationships be-
In addition, the information gathered from sensors is so massive. tween features that successfully performed the classification task.
It is impossible to operate them for knowledge-discovery without The system has two main limitations. First, the general semantic
cloud environments and big data framework. lexicon cannot capture the meanings of diabetes-related text.
26 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43

Fig. 1. Different layers in the proposed healthcare monitoring framework. The data source layer provides different sources of data. The data collection layer gathers
data from different sources about patients. The data-storage layer is responsible to store the collected data. The analytics engine layer analyzes the healthcare data
and predicts the patient health conditions. The data representation layer presents the analysis results to physicians.

Second, it requires fuzzy ontology-based semantic knowledge to However, their rule-based system is limited to specific cases.
identify the aspects. An ontology mapping framework was pre- Therefore, instead of using the traditional fuzzy rule-based sys-
sented to discover the relationships between ontology features tem, there is a need of the deep learning model that can learn
and required features in text [35]. In this system, the authors from the existing data and provide the results for precise de-
used various ontologies from bioportals, and proposed a novel ap- cision making. A recommendation system based on numerical
proach to extracting features from ontologies for semantic word reputation and context investigation were proposed to guarantee
embedding. A deep learning method was proposed to present the veracity of big data [49]. In addition, ML methods and ana-
medical knowledge embedding [36]. In this work, biomedical lytic algorithms were used for handling the volume of data and
ontologies related to epilepsy were employed to automatically managing the velocity of data, respectively. A disease diagnosis
find medical concepts from clinical text. However, clinical text healthcare framework based on the cloud-centric IoT was pro-
is very noisy and the generated triples for finding medical con- posed in order to predict diseases in students [50]. In this system,
cepts may not be correct. Therefore, the integration of semantic student healthcare data were generated using medical sensors
attributes is required to address the above issues. An ontology and a University of California, Irvine (UCI) dataset, and ML algo-
was proposed to boost the deep learning performance for disease rithms were trained to predict diseases in students. In addition,
name extraction from Twitter text [37]. The authors presented the system utilized medical measurements for disease diagnoses.
the architecture of an ontology-boosted neural network to auto- However, same measurement algorithms for six different diseases
may generate the wrong results. Semantic knowledge with deep
matically extract disease names from Tweets. The system used
learning method is required to handle the issues of disease diag-
different types of features for sentiment classification. However,
noses. A big data analytics system was proposed for an effective
they failed to achieve high accuracy for standard disease names.
health recommendation system [51]. This system handles large
A System called BO-LSTM was presented to detect biomedical
amounts of structured and unstructured data extracted from pa-
information in text [38]. This approach overcomes the problem
tient social networking activities. In addition, ML algorithms were
of missing semantic entities in word embedding. The authors
applied to recommend centric treatments for patients. However,
utilized a domain ontology with word embedding in order to en-
this system is possibly incapable to filter the unstructured data
hance the extraction of biomedical relations using deep-learning properly.
LSTM. Dwivedi et al. [52], discussed the effects of the IoT on the
results of big data sentiment analysis. At first, they introduced
2.3. Healthcare big data and deep learning approaches the notion of big data sentiment analysis, features, and the value
of decisions, and then, they described the framework required to
The analysis and representation of healthcare big data with handle big data. However, this framework is complex and needs
low dimensions is another main issue in healthcare monitor- more management to reduce the task on the Hadoop cluster.
ing systems. Recently, various systems have been presented to A random forest and MapReduce were used to present a big
analyze big data [39–44], to decrease the dimensionality of big data-analytics-based IoT healthcare system [53]. This system con-
data [45], and to manage large amounts of heterogeneous data sidered patients with different diseases for analysis. In addition,
using ML [46,47]. Jindal et al. [48], collected patient data using an improved dragonfly algorithm was applied to select optimal
remote healthcare applications, and then applied a fuzzy-rule attributes for better classification. This random forest classifier is
classifier to the data for efficient decision making. They discussed computationally slow because of healthcare big data. This issue
various issues in information clustering, information retrieval, can be addressed using deep learning approaches. A new IoT-
and parallel processing of big data in the cloud environment. based framework was proposed to precisely collect the sensor
 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43 27

data for healthcare applications [54]. The system utilizes Apache 3.1.1. Wearable devices
HBase and Apache Pig for the collection and storage of sensor Smart sensors and wearable devices, such as a blood pressure
data, and it applies a prediction model of MapReduce to predict monitor, a glucometer sensor, the smart watch, pulse oxime-
heart disease. However, it is not enough to transfer sensor data ters, accelerometers, temperature sensors, and weigh scales, can
to the cloud for storage and processing. It needs a smart data an- be utilized to collect the patient’s real-time body signals for
alytics strategy with secure architecture for precise heart disease the monitoring of physiological signs in the patient, such as
prediction. blood pressure, blood sugar level, heart rate, stress rate, oxy-
Most of the aforementioned systems were based on traditional gen saturation rate, temperature, and weight. In our system,
methods, and they handle the problem of big data to some extent. wearable devices are placed on the patient’s body to monitor
However, there is no clear framework for using big data analytics the bodily functions. In addition, the use of mobile devices has
to process and analyze big data with high accuracy. In addition, largely increased, and the applications related to health have
these systems are still not good enough to deal with different been well developed in recent years. A mobile device contains
types of structured and unstructured data. Due to the limitations sensors that provide the opportunity to collect accurate infor-
of medical semantic knowledge for ML classifiers, these systems mation regarding different parts of the human body. Therefore,
can incorrectly classify social networking data. a smartphone is used to collect information on diet, exercise, and
other activity information from diabetes and BP patients, along
3. The proposed healthcare monitoring framework for dia- with personal information (age, gender, height, and other infor-
betes and abnormal BP patients mation). This information is valuable for healthcare and disease
prevention. However, smartphone data are highly unstable and
The proposed framework is presented in Fig. 1. This frame- can be damaged easily. Therefore, it is difficult to use smartphone
work contains different layers, namely, the data source layer, data for accurate healthcare and relevant information extraction.
the data collection layer, the data storage layer, the analytics We have utilized data mining techniques and ontology-based
engine layer, and the data presentation layer. The data source semantic knowledge to manage sensor data and extract relevant
layer deals with heterogeneous data. Sensor devices, medical information for efficient healthcare.
records, and social networking platforms are the main sources
of data. The data collection layer is responsible for gathering 3.1.2. Medical records
data from different domains about diabetes and BP patients. This Medical records are the data about treatments undergone
layer collects physiological information from wearable devices, by diabetes and BP patients. We collect the medical records of
drug and symptom information from smartphones, and patients’ patients, which contain the patient history (treatments, lab tests,
and doctors’ discussions on social networks through application and drug intake). These records can be analyzed to extract valu-
programming interfaces (APIs). The third layer is the data storage able information that can help provide insights into improving
layer. All the monitored data of patients using wearable devices medical guidelines for diabetes and BP diagnosis. However, the
and the collected data from social networks are offloaded to volume of medical records is usually large, and each record com-
the cloud server through a wireless communications network. prises data with distributed variables and of high dimensionality.
The fourth layer is the analytics engine layer, which is the most In addition, diabetes and BP patients may face other complica-
important layer of the proposed framework. It is divided into tions, such as kidney and cardiovascular disease, neuropathy, and
two sub-layers: the data computation layer and the data clas- skin and eye problems. Therefore, the medical records should be
sification layer. The data computation layer has sub-modules, analyzed to identify patients affected by the above complications,
namely, data preprocessing and analysis, feature extraction, word and to monitor their status with more specific examinations.
embedding, and dimensionality reduction. Here, ontology-based
semantic knowledge, along with soft computing approaches, are 3.1.3. Data from social networking platforms and webpages
employed to process and analyze the data for the extraction of The proposed system first extracts patient contents from hos-
required information. In the data classification layer, Bi-LSTM pital social networking platforms. However, this task needs extra
with ontologies is utilized for the classification of diabetes, BP, work, and completely depends on social network privacy settings.
mental health, and drug side effects. This layer intelligently an- The APIs of some specific social networks are not publicly avail-
alyzes the multidimensional big data about diabetes and BP to able. In this situation, special software, such as wrappers, can be
get insights for decision making from the data, and provides a used to extract information (e.g., patient posts) [28]. Generally,
personalized diabetes and BP healthcare system for patients. The diabetes and BP patients regularly contact their physicians, but
proposed system uses Hadoop MapReduce with ML to reduce they also need support, knowledge, and skills for personal mon-
large-sized data about patient treatments. The final layer is the itoring of their healthcare condition. In addition, if patients do
data representation layer, which presents the analysis results to not get efficient information from their physicians, social media
physicians. This layer combines the generated results with the can play an important role in fulfilling their needs. Therefore,
physician’s suggested treatment, and then, the personal diabetes social networking platforms like Facebook and Twitter provide
and BP healthcare treatment is recommended to the patient. The opportunities for patients to get enough knowledge regarding
proposed system helps warn diabetes and BP patients before diabetes and BP and to connect with people who have similar
their health risk reaches a high level. It supports physicians in health problems and experiences. Social networks provide plat-
offering actual treatments to their patients by monitoring health forms for both patients and physicians to share their knowledge
conditions smartly. The whole idea of the proposed system is regarding diabetes treatments. We collect social media data, such
displayed in Fig. 2. We briefly discuss the different modules of as drug reviews and emotional posts of patients, to predict their
the proposed system in the following subsections. stress and depression levels, to identify the side effects of diabetes
drugs in the context of diet and lifestyle, and to improve patient
3.1. Data collection care and knowledge. The procedure used for data collection from
Facebook and Twitter is discussed in the following sections.
This proposed system collects the data from four different Facebook: The popularity of social networks has increased in the
sources. In this section, we extensively discuss data collection healthcare domain. Therefore, patients connect with online com-
from all four sources, one by one. munities to meet their needs. Various communities on Facebook,
28 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43

Fig. 2. The workflow of the proposed system. The first step is to collect patient data using wearable sensors and social media platform. The data are then stored in
the cloud server. An ontologies-based semantic knowledge is applied to transform the data into a readable form. Next, deep learning models classify the transformed
data in order to predict the patient health condition for precise healthcare monitoring.

such as the American Diabetes Association, diabetes recipes and Webpages: Diabetes patients are interested in knowing about
food hints, Diabetes Daily, and Diabetes Health, provide platforms any adverse side effects, diseases, and symptoms from taking
for patients to share information with each other. Patients can a specific drug. Usually, patients share their experiences about
respond to posts and share them with others. We employed adverse drug reactions in social media websites such as https://
the Graph API along with a Java client, RestFB, to extract data www.askapatient.com/, https://www.webmd.com/, http://www.
from Facebook pages [55,56]. The Graph API permits us to au- druglib.com/, and https://www.drugs.com/. These websites con-
tomatically extract information from Facebook. We first select tain a huge amount of information related to adverse drug re-
those communities’ pages that contain information about dia- actions from anti-diabetes drugs. Therefore, these websites were
betes patients and drugs. Then, we extracted all posts from the chosen as data sources for the proposed system. We used the
communities’ pages that were published between January 2017 names of intake drugs for diabetes and BP as queries, retrieved
and January 2019. The responses (reactions and emotions) made 1600 posts from https://www.askapatient.com/ and https://www.
in these posts were collected and stored for further processing. webmd.com/, and used a keyword-based search engine to extract
Twitter: The Streaming API and REST API of Twitter were em- posts that contain only drug-related information. In addition,
ployed to retrieve Tweets holding diabetes data. We retrieved the the proposed system employed an automatic web clawer that
most recent Tweets using the REST API with various queries [57]. gathered 25,000 user comments (reviews) about drugs from http:
In these queries, the most specific terms related to diabetes //www.druglib.com/ and https://www.drugs.com/.
should be used to retrieve the most-related Tweets. Thus, we
built queries that are based on keywords with the Boolean op- 3.2. Data store in big data cloud server
erators AND and OR, e.g., patient AND (blood pressure OR blood
sugar) for diabetes, patient AND (heart rate OR stress rate OR In the proposed system, four different sources of data are con-
blood sugar) for hypertension, and patient AND (diabetes OR sidered to monitor patient health and provide useful information.
drugs) for diabetes drugs. More than 30 keywords were used to Here, wearable devices data are used to extract physiological in-
construct queries related to diabetes [58]. We fetched 300,000 formation about the individual patient. Second, patients’ medical
Tweets about diabetes patients, treatments, drugs, and symptoms records are collected to determine their treatments and med-
by using the aforementioned keywords-based queries. ical histories. Third, the patients’ contents are extracted from
 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43 29

social networking sites to understand their feelings, emotions, patient’s body. However, for simplicity, only these parameters
and stresses. Fourth, patient reviews about drugs are collected are mentioned in the proposed work. In pre-analysis, various
from medical webpages to identify the side effects of the current steps are performed. First, the dataset is converted into comma
drug intake. These large volumes of data from four different separated value (CSV) files for easy parsing. Names are assigned
sources are difficult to store and handle. In addition, the number to each attribute (parameter), and they are presented in the
of patients with various diseases is largely increasing, which form of columns along with numerical values. Identities (IDs) of
generate a large volume of health-related data. Therefore, a smart sensor data are then plotted to the actual sensor name (e.g., S1 is
approach is needed that handles and processes the data compe- the ID for blood sugar). We use various attributes for different
tently. Hence, a big data cloud repository is utilized to store the purposes. For example, we utilize blood sugar, blood pressure,
extracted data so it can easily be accessed from any place and heart rate, body mass index (BMI), age, gender, and activities
at any time. The data must be stored in an intelligent way so attributes to classify the health condition of diabetes and BP
the information can be retrieved correctly and rapidly. The data
patients. However, instead of using an explicit year and time for
used in our proposed work are in large volumes with versatility
age and activities, respectively, we divide the attribute age into
and velocity. Therefore, the traditional methods of storage and
day, months, and years, and divide activities time into morning,
retrieval may not process the data from different sources. The
afternoon, and evening. These generated features provide valu-
proposed system is connected with a personal cloud server called
able insights into patient health, such as any pattern of normal
Amazon S3 [14], which is highly scalable and secure to store
and abnormal conditions based on whether it is the start or the
patient information. Amazon S3 stores the data in buckets, and
processes them for different purposes. A unique name and URL end of the month, or if there is any serious condition during
is assigned to each bucket, and the S3cmd method is utilized the starting and ending months of the year. Afterward, the final
to upload the data into Amazon S3. The S3cmd method allows datasets are uploaded into the Hadoop cloud environment for
users to upload, retrieve, and manage the data in the cloud further processing.
database [54]. Amazon S3 provides a facility to upload the data
into an HBase cluster. As shown in Fig. 3, wearable sensor and 3.3.2. Preprocessing and filtering of sensor data
social networking data are transferring from the cloud server to Data collected using wearable sensors have various limita-
a Hadoop Distributed File System (HDFS). HDFS is a distributed tions. They contain a lot of incorrect and useless information.
file storage system that provides high bandwidth in order to In addition, sensor data are corrupted by signal artifacts, such
transfer the data to HBase cluster [52,59]. HBase runs on top of as noise and missing values, which highly decrease classification
HDFS. It is distributed database management, which stores the performance. Therefore, the data are preprocessed and filtered
data in the form of rows and columns. Flume is used to transfer
before analysis. We filter the data to remove inconsistencies and
the data from the cloud server to the Hadoop ecosystem. Flume
noise. The data are cleaned by removing ASCII characters. A well-
contains three units (source, decorator, and sink), which present
known filtering approach called Kalman filtering is utilized to
the sources of the data, decorations of the data (e.g., compression
remove such noise from the data [60]. Furthermore, the unsu-
and decompression), and targets of the data for specific purposes,
pervised filter called ReplaceMissingValues is utilized to replace
respectively.
all missing numeric values in the dataset with means and modes
We use Apache Pig to transfer the data extracted by wear-
able devices into MapReduce. Apache Pig handles structured and from the available data. Useless attributes are removed with a
unstructured data and represents them for data analysis. MapRe- maximum variance of 90% using an unsupervised filter called the
duce is a parallel processing system in Hadoop for large datasets, RemoveUseless filter. The numerical values are then normalized
which works in a distributed environment. It has two main func- using a Normalize filter to limit them to between 0 and 1 for
tions, namely, Map and Reduce. The map task collects values any classification. After these steps, the EmEditor is applied to
from the big data in key–value pairs. The reduce function stores divide the dataset into n data files, and they are uploaded into
the value set for a specific key. MapReduce intelligently handles the Hadoop cloud environment for further processing.
structured and unstructured data.
3.3.3. Preprocessing of medical records
3.3. Data analysis using the big data analytics engine The Medical Records (MR) contain complete patient history in
the digital format. It consists of different medical data describing
The data of the proposed system consists of sensing data, med- the patient’s health, such as laboratory tests, self-examination
ical records, and social networking data. However, it is extremely
answers, and medications taken. Laboratory tests are data from
difficult to handle real-world big data due to its inconsistencies,
medical devices that can be utilized to judge the health sta-
missing values, noise, different formats, large size, and high di-
tus of the patient in terms of reference values. In addition, the
mensionality. Low-quality and noisy data produce low-quality
evaluation of the patient’s health status can change depending
results. The data preprocessing step is applied before actual pro-
on the patient’s disease history and family disease history. Self-
cessing, which improves the quality of the data processing and
saves time. Our system includes pre-analysis of sensor data, pre- examination data contain information, which is collected from
processing and filtering of sensor data, preprocessing of medical patients through questions. The questions are regarding periods
records, and preprocessing of social network content, as shown of indigestion, drinking and smoking habits. The medication data
in Fig. 4. include information about prescribed drugs for diagnoses. Some
of the attributes of the MR can be used for patient classification.
3.3.1. Pre-analysis of sensor data Therefore, the ID and reference value (0, 1, or 2) are assigned to
The physiological information of the patient is extracted using each attribute of the MR for data analysis. The reference values
wearable biomedical and behavioral sensors. Different param- 0, 1, and 2 indicate the patient’s health condition (normal, pre-
eters are sensed from diabetes and BP patients using sensors diabetes, and diabetes). In addition, we also use the MR data
and a smartphone, as shown in Table 1. The sensed parameters to overcome the limitations of sensors-based generated data;
cover most of the symptoms of diabetes, abnormal BP, and other for example, we replace a missing value with the current MR
diseases. In addition, other parameters are extracted from the attribute value in the dataset.
30 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43

Fig. 3. Big data cloud server and the Hadoop Distributed File System. This system is based on HDFS and MapReduce. The HDFS provides high bandwidth in order to
transfer the data to HBase cluster, which stores the data in the form of rows and columns. The MapReduce utilizes Apache Pig to handle structured and unstructured
data and represents them for data analysis.

Fig. 4. The flow chart of the big data analytics engine. This analytics engine handles three different types of data: wearable sensors data, medical records, and
social network contents. First, the preprocessing and filtering techniques are applied to preprocess the data. Second, the information gain approach is used to extract
features from sensor’s data. In addition, the method of feature polarity identification is applied to label the textual data. After this, the data dimensionality is reduced
using PCA. Finally, word embedding model is trained to represent textual data, and machine learning classifiers are applied to predict patient health conditions and
drug side effects.

3.3.4. Preprocessing of social network content corpus data, and transforms the data into a set of words that
Preprocessing of social network content is an important task are useful representations for ML classifiers. The proposed system
before text classification. It reduces the noise that occurs in the collects social network content and drug reviews, and stores
 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43 31

Table 1
Different health parameters collected from sensors, smartphones, and medical records.
Resource ID Parameter Explanation
S1 Blood sugar Blood glucose (mg/dL)
S2 Body temperature Patient’s current body temperature
S3 Blood pressure Systolic blood pressure (mmHg)
Sensors Diastolic blood pressure (mmHg)
S4 Oxygen saturation Patient’s SpO2 consumption (mmHg)
S5 Heart rate Patient’s heart rate (bpm)
S6 ECG Patient’s electrocardiography using ECG sensor
S7 EEG Patient’s electroencephalography using EEG sensor
S8 Stress Patient’s stress calculation using ECG+EEG pattern
SP1 Age Age of the patient in days, months, and years
SP2 Height Stadiometer to find the patient’s height
Smartphone SP3 BMI Body mass index (kg/m2 )
SP4 Gender Patient’s gender (0/1)
SP6 Activities Lifestyle: sedentary, slightly active, moderately active, active, or very active
MR1 Lipoprotein level Low-density lipoprotein level (LDL cholesterol)
High-density lipoprotein level (HDL cholesterol)
MR 2 Hemoglobin Glycated hemoglobin (A1c) of patient (%)
MR 3 Blood sugar Patient’s blood test
MR 4 Serum creatinine Patient’s blood test
MR5 Triglycerides Patient’s blood test
Hospital medical records MR6 Cholesterol Patient’s blood test
MR7 AST (SGOT) Blood test to check for liver damage
MR8 ALT (SGPT) Blood test to check for liver damage
MR9 Drugs intake Extracted from prescription list
MR10 Smoking Yes/No (self-examination)
MR11 Drinking Yes/No (self-examination)
MR12 Indigestion Yes/No (self-examination)
MR13 Family’s disease history Yes/No (self-examination)
MR14 Patient’s disease history Yes/No (self-examination)

them in the HDFS. The collection procedures for this content 3.4. Feature polarity identification and document labeling
are explained in Section 3.1. The following steps are applied to
transform the data into structured form. In addition, these steps The proposed system detects the patient’s stress and depres-
remove useless data, which helps to easily extract features and sion using their published contents on social networks. In ad-
opinion words. dition, the system performs multiple tasks on drug reviews to
Stop word removal: Words such as prepositions (to, in, and of), all identify their opinions on the efficiency and side effects of di-
articles (a, an, and the), symbols (#, @, etc.), and URLs in the cor- abetes drugs. We use a sentiment analysis approach for the
pus data do not disturb the meaning of the document. However, aforementioned two tasks. Therefore, it is important to find the
they reduce the accuracy of text classification. Therefore, it is feature polarity and document labeling for sentiment classifica-
essential to remove them to decrease noise in the text. We utilize tion. After preprocessing of social networking data and webpage
a well-known method called Rainbow to delete this content [58]. contents, we use SentiWordNet (SWN) to identify the polarity
Tokenization: Tokenization separates a complex text in the corpus of the opinion words of the features [62,63]. The results of the
into small terms or tokens by removing white space and delim- feature polarity are then accumulated to find the polarity of the
whole document. SWN is a resource of lexicons, which links each
iters. Generally, white space and delimiters occur in a complex
synset of a WordNet with three numeric values: positive, nega-
text. Therefore, the n-gram tokenizer is applied to delete white
tive, and objective. However, SWN contains senses for each word
space and delimiters. The output is then saved for further anal-
to indicate if it is a noun, verb, adjective, or adverb. Therefore,
ysis, such as extracted words’ part of speech (PoS) tagging, and
Word Sense Disambiguation (WSD) is used to handle this issue by
lemmas [61].
extracting the needed category sense for each word. In addition,
PoS tagging: PoS tagging defines words in the text. We split the
it assigns a zero value to the input word if SWN does not hold
corpus text into sentences, and then use Stanford Core Natural any sense for it. After WSD, we extract the SWN scores for the
Language Processing (CoreNLP) for POS tagging. After tagging, it identical senses of the opinion words, and then calculate the
is confirmed that every sentence has a complete clause with a polarity of each feature using the following equations [64]:
noun and a verb [55].
n
Stemming and lemmatization: Stemming converts the words in ∑
Posscore (Fi ) = Posscore SWNw (1)
the corpus text into their own basic forms. The system applies
w∈w fi
a suffix-dropping algorithm for stemming. Lemmatization ex-
n
presses the lemma of words used in the text. After lemmatization, ∑
Negscore (Fi ) = Negscore SWNw (2)
the system easily obtains the lexical information of each word.
w∈w fi
For example, blood sugar is related to blood glucose. Therefore, the
n
stem and lemma words are utilized for further processing. ∑
Character conversion: Patients employ unusual words (e.g., de- Nuescore (Fi ) = Neuscore SWNw (3)
w∈w fi
pressssssed) on social networks that affect the results of classi-
fiers. Therefore, we convert a series of characters that appear where Posscore SWNw , Negscore SWNw , and Neuscore SWNw indicate
more than twice into a general word (e.g., depressssed becomes positive, negative, and neutral scores of the feature, respectively.
depressed). This score is computed by arithmetic means of SWN for individual
32 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43

word w . A detailed example is provided in [64]. If Posscore (Fi ) > If the TF (Term, Onto) value is greater than zero, this step will
Negscore (Fi ) and Neuscore (Fi ), then the system considers the feature be repeated to extract the specific concept. TF-IDF selects the
polarity as positive. In contrast, the feature polarity is negative if illustrative terms for information extraction. The IDF reduces the
Negscore (Fi ) > Posscore (Fi ) and Neuscore (Fi ). Lastly, it is considered significance of the word that appears mostly in all ontologies
neutral if Neuscore (Fi ) > Posscore (Fi ) and Negscore (Fi ) .W1,1 (e.g., patient, disease, hospital, etc.). If the term occurs in more
ontologies or in more concepts in a single ontology, it means it is
3.5. Word embedding and ontology-based features extraction from a regular term and may not be the required term for information
textual data extraction. Therefore, the result of the log function for the input
word will decrease to zero. This shows that the value of TF-IDF is
Word embedding is a word representation approach where small for this term. The statistical description of the IDF is shown
words in the document are encoded with numeric values for in the following equation:
advanced analysis. Two well-known approaches can be applied
Nf
for word representation: one-hot word encoding and word em- IDF (Term, Ontos) = log (5)
bedding [55,65]. In this work, the word embedding approach 1 + |{Onto ∈ Ontos: Term ∈ Onto}|
is applied to represent words as numeric values. It sets the where Nf indicates the total number of ontologies in the database
dimensionality, d, and then uses the d-dimensional values to or the total number of concepts in the ontology, e.g., Nf = |O|
express the word. For example, when d is set to 3, ‘‘metformin’’ is and |{Onto ∈ Ontos: Term ∈ Onto}| is the number of ontologies or
expressed as {0.3, 0.6, and 0.8}. This method drastically decreases the number of concepts in the ontology where Term appears. We
the dimensions. However, the embedding approach avoids the remove the common terms using the following TF-IDF equation:
relationships between words in a matrix. Therefore, we use a
neural network-based word embedding model called Word2vec TF − IDF = TF (Term, Onto) .IDF (Term, Ontos) (6)
for word representation. Word2vec contains two architectures:
The TF-IDF result shows how essential the ontology feature is
continuous bag-of-words (CBoW) and the skip-gram model. The
to the ontology in the ontologies corpus, or shows how important
CBoW model uses the surrounding context of the word for word
the ontology feature is to the concept in the ontology. Based
prediction. The skip-gram model predicts the surrounding con-
on the values of TF-IDF terms (X=2), the top concept can be
text using the current word, as shown in Fig. 5. We trained the
extracted. Let us suppose that the system gets the text related to
skip-gram model of Word2vec with 200-dimensional vectors on
diabetes, as shown in Fig. 5. In this figure, the ambiguous words
the dataset. It identifies the relationships between words, and
for information extraction are in bold type (e.g., hyperglycemia).
helps predict the neighboring words of the input word. This
Once the semantic information about these words is identified,
Word2vec model is then utilized to train ML classifiers to detect
the system then extracts that specific concept from the ontology
the association between features. Generally, ML classifiers miss
to provide additional information to word embedding and ML
the basic semantics of the features according to their particular
classifiers (e.g., hyperglycemia is a disease complication).
domain. An ontology can represent the semantics of a given
domain in some cases.
3.6. Features extraction from wearable sensors data
An ontology aims to provide semantic knowledge of concepts
and their relations in a particular domain. Our online available
The huge size of the data is another main issue associated with
biomedical ontologies cover various topics related to diabetes,
depression, hypertension, drugs, and foods. In this paper, we use sensor-based healthcare monitoring systems. Raw data from the
particular domain ontologies, where each class of the ontology patient’s body extracted in large quantities is a burden for data
is a concept or feature of the domain, and their properties are processing. It is important to decrease the size of the data without
relations between features. This is the basic representation of losing useful information. The dataset contains a large number
the online available biomedical ontologies in the National Cen- of attributes about diabetes and BP patients. However, all the
ter for Biomedical Ontology BioPortal, which is currently the attributes will not be needed for patient classification. Unneces-
normal method to formalize knowledge about features. We use sary attributes are time consuming and decrease the accuracy of
the latest releases of the Ontology for Nutritional Studies (ONS), classification. Different methods are utilized for feature selection,
the BioMedBridges Diabetes Ontology (DIAB), the Diabetes Mel- such as the dragonfly algorithm and recursive feature elimina-
litus Treatment Ontology (DMTO), the Human Disease Ontology tion [71,72]. We use the Information Gain (IG) method, which
(DOID), the Drug Target Ontology (DTO), and the Fast Health- affects the classifier by reducing noise and irrelevant features.
care Interoperability Resources (FHIR) and semantic sensor net- Information gain (IG): IG chooses features based on the informa-
work (SSN)-based Type 1 Diabetes Ontology (FASTO), which con- tion’s contribution related to the variables of the class without
tain lots of concepts and relationships, as shown in Table 2. seeing attribute interactions [73]. Every attribute or feature in
These ontologies provide additional information for a neural net- the dataset has importance, and based on that importance, the
work model that understands the semantic meanings of unusual system can learn about specific problems. Waikato Environment
words. In addition, they affect the word-level semantics in word for Knowledge Analysis (WEKA) is utilized to compute the IG [74].
embedding and text classification. It contains various methods for data processing. We apply the IG
In this section, the system retrieves all the ontologies and filter ‘‘infoGainAtrributeVal’’ as an evaluator on the diabetes and
attempts to integrate the ontology information with word em- BP dataset to obtain the results. However, IG cannot be used for a
bedding for classifiers. Generally, the bag-of-words model is used numerical dataset. Therefore, it is important to convert numeric
for features. Therefore, we consider each ontology as a bag-of- data into nominal data before using IG. When the value of the
words. To extract information from the ontology, we employ attribute is identified, the IG measure is linked with a decrease
well-known statistical methods called term frequency (TF), and in entropy in the training dataset. This approach finds the value
term frequency and inverse document frequency (TF-IDF) [35,55, of an attribute by calculating the IG according to classification.
70]. Here, we consider each concept or feature of the ontology as The proposed IG method utilizes entropy to measure system
a term, and the ontology as a document. Therefore, TF is the terms uncertainty and finds the difference between prior entropy and
(ambiguous words) found in the ontology. TF is mathematically post entropy [75]. It specifies the amount of extra information
defined in the following equation: about A provided by B, as shown in Eq. (7):
TF (Term, Onto) = 1 + log (Featureterm , Onto) (4) IG (A|B) = H (A) − H (A|B) (7)
 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43 33

Fig. 5. Ontologies and Word2vec-based word representation. Ontologies provide additional information for Word2vec model to understand the semantic meanings
of unusual words.

Table 2
BioPortal ontologies of different domains as datasets.
Ontology and acronym Definition Number of Number of Number of
classes properties individuals
Ontology for Nutritional Studies It provides a description of composite 3442 66 104
(ONS) [66] nutritional studies.
BioMedBridges Diabetes Ontology It represents the relations between diabetes 375 4 0
(DIAB) [67] phenotypes for text mining.
Diabetes Mellitus Treatment It provides interoperable facts for diabetes 10 700 315 63
Ontology (DMTO) [68] treatment.
Human Disease Ontology (DOID) It represents the concept of rare diseases. 12 694 15 0
[67]
Drug Target Ontology (DTO) [69] It provides information for the classification of 10 075 0 0
drug target data.
FHIR And SSN-based Type 1 It provides management details on insulin for 9577 822 460
Diabetes Ontology (FASTO) [6] diabetes patients.

where A and B are discrete variables. A is a feature, and its prior The IG can be computed by putting Eqs. (8) and (10) into
entropy can be measured using Eq. (8): Eq. (7), as shown in Eq. (11):
∑
∑ IG (A|B) = − P (Ai ) log2 P (Ai )
H (A) = − P (Ai ) log2 P (Ai ) (8) i
i ⎛ ⎞
∑ ( )∑( ( ) ( ))
where P (Ai ) represents the prior probability for the discrete value − ⎝− P Bj P Ai |Bj log2 P Ai |Bj ⎠ (11)
j i
of Ai . The conditional entropy of A, after given the post entropy
B, can be defined as shown in Eqs. (9) and (10):
3.7. Principal component analysis
∑ ( )
H (A|B) = −
( )
P Bj H A|Bj (9)
j
The dimensionality of the dataset is increased after prepro-
∑ ( )∑( ( ) ( )) cessing and extra-attribute generation. This creates problems,
=− P Bj P Ai |Bj log2 P Ai |Bj (10) such as a decrease in classification accuracy, over-fitting, and
j i time complexity. Therefore, we use principal component analysis
34 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43

(PCA), which is a statistical approach to dimensionality reduc- cell (C t ), and output gate (Ot ). These components control the up-
tion [76]. It converts p-dimensional features into q-dimensional date and use of previous data. The LSTM output can be computed
features (q < p) with the least loss. The q-dimensional features using the following equations [58]:
are entirely new orthogonal features, called principal comments. t
C = tanh wxc X t + whc ht −1 + bc
( )
(14)
The new orthogonal feature has its own exclusive meaning. The
it = σ wxi X t + whi ht −1 + wCi C t −1 + bi
( )
information of features can be imitated in variance. Feature with (15)
high variance rate shows that the feature contains the main
f = σ wxf X t + whf ht −1 + wCf C t −1 + bf
t
( )
(16)
information. This can be identified by cumulative variance rate.
t t t −1 t t
The main aim of PCA is to find the projection vectors with the C =f ⊙C +i ⊙C (17)
highest variance. Projection vector X on vector Y on the same O = σ wxo X + who ht −1 + wCo C t −1 + bo
t
( t
)
(18)
plane can be illustrated using Eq. (12):
where X , wx and wh , and b are input, weight matrices, and bias
t
⟨X , Y ⟩
f = .Y (12) vectors of LSTM, respectively; tanh (.) and σ (.) are the hyperbolic
|X | tangent function and sigmoid function, respectively. The final
Let X is the sample of n-dimensional drug side effect dataset, output of LSTM can be calculated using the following equation:
where X = {x1 , x2 , x3 , . . . .., xm }. The sample of the dataset
is arranged and therefore, it has a zero mean. In addition, we
ht = Ot ⊙ tanh C t
( )
(19)
assume that after projection transformation the new coordinate
system is W = {w1 , w2 , w3 , . . . .., wn }, where wi represents the where ⊙ is features-wise multiplication between the output gate
standard vector of orthogonal basis, which is ∥wi ∥2 = 1. The and input cell state. The LSTM output is linked with a softmax
projection of xi on the hyper-plane in the new space is W T X , function to identify the probabilistic outputs (0, 1, and 2), where
where W T X = w1 x1 . . . . . . wn xm . However, the main objective is 0, 1, and 2 show that the patient is normal, pre-diabetes, or
to find the projection of all points of X on the hyper-plane with diabetes, respectively. The softmax function can be calculated
maximum variance. The projected points with maximum variance using the following equation:
can be calculated using Eq. (13). exp(xst )
n
softmaxst = ∑c ( ) (20)
∑ ḱ
exp xśt
T 2
u = argmax (W X ) (13)
∥w∥2 =1
i=1
where c and xst are feature category and the input of time step
k, respectively. After preprocessing, feature extraction, dimen-
where u is an eigenvector related to the maximum eigenvalue sionality reduction, and word embedding, a sequence of inputs
(λ) of X T X . It is important to achieve eigenvalue decomposition (features and word vectors) representing the patient’s physiolog-
on the covariance matrix X T X , and arrange the eigenvalues in ical information and drug reviews is then fed into the Bi-LSTM
descending order. Usually, the number of principal components layer. We developed four Bi-LSTM-based classifier models: LSTMa ,
is selected based on the cumulative variance rate. In this paper, LSTMb , LSTMc , and LSTMd .
the PCA with cumulative variance rate of 90% is applied to re- In this study, wearable devices and smartphones are utilized
duce the dimensionality of datasets. The original drug side effect to collect the patient’s physiological and personal information.
dataset contains 200 feature dimensions. By using PCA and IG, The collected data are preprocessed and converted into structured
the dimensionality of the input data is reduced to 67 features form for further processing. At each time step, t, feature Xt is
(principal components). In addition, it shortens the processing fed into the LSTMa units to predict diabetes class {normal, pre-
time by reducing the dimension of the dataset before feeding it diabetes, and diabetes}. The first LSTM unit of the LSTMa model
to Bi-LSTM models for classification. may predict the patient category based on a specific feature.
However, all the features should be used to get important in-
3.8. Bi-LSTM-based diabetes and BP classification and drug side formation for classification. Therefore, the sequence of diabetes
effect prediction features, along with the results of the previous unit, is assigned
to the next LSTM unit. This procedure is repeated with each input
feature. In this way, the units of LSTMa save the valuable features
Machine learning approaches, such as the CNN, MLP, the SVM,
and generate the output. The output is linked with a softmax
logistic regression, decision trees, and KNN, are useful algorithms
activation function, which predicts the diabetes patient category.
for classification, feature selection, data normalization, and statis-
The medical rules to classify diabetes and blood pressure patients
tical analysis [21,77–79]. These approaches conduct tasks using
are presented in Table 3. Wearable sensors for blood pressure and
supervised and unsupervised methods. However, they are time
an ECG are used to identify the patient’s BP and heart rate (HR),
consuming, and their performance decreases when the data size
respectively. The patient can be evaluated from systolic BP and
is large [58]. Therefore, in the proposed healthcare monitoring diastolic BP values in millimeters of mercury, and HR values in
system, Bi-LSTM is used to classify the patient’s diabetes, BP, beats per minute, as shown in Table 3. However, these parameters
mental health, and side effects from diabetes drugs. The recom- are sometimes affected by physical activity, sleep, temperature,
mendation system is activated in cases of diabetes, high BP, stress, stress, eating habits, etc. [5]. In addition, patients may have BP
or depression, and from severe side effects of diabetes drugs. and HR problems due to other factors, such as family profile and
Fig. 6 shows the architecture of Bi-LSTM for the classification of disease history. For example, HR is always high during exercise.
diabetes, BP, mental health, and drug side effects. The grid search This means that an abnormal health condition in the patient is
optimization algorithm is applied to identify optimal values for sometimes not dangerous. Therefore, it is essential to use the
the hyper-parameters of the LSTM model. The selected optimal patient’s medical records and daily activities to detect abnor-
values for the hyper-parameter dropout rate, epochs, batch size, mal conditions. We select main features from medical records,
and learning rate are 0.3, 30, 32, and 0.001, respectively. including age, gender, and family history, from the diagnosis
LSTM is a type of RNN that consists of five main components: and medical history. All the diagnosis and medical records are
t
memory cell (C ), input gate (it ), forget gate (f t ), current memory searched to extract information related to diabetes and BP.
 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43 35

Fig. 6. LSTM-based healthcare data classification. The system consists of four LSTM models that predict the patient health condition.

Table 3
Medical rules to classify diabetes and BP patients.
Category Age Family Gender Activity BMI Blood sugar Heart rate Systolic BP Diastolic
(years) (Y/N) (M/F) (in steps) (MG/dL) (bpm) (mmHg) BP (mmHg)
Normal 20–50 N M/F 5000+ 18–25 ≤ 100 60–80 – –
Diabetes
(fasting)
classification
Pre-diabetes 41–65 Y M/F 3001–5000 25–30 or 100 to 125 140–160 – –
more (fasting)
Diabetes 60+ Y M/F ≤ 3000 30+ ≥126 140–160 – –
(fasting) or ≥160
Normal 20–50 N M/F ≤ 3000 – – 60–100 90–119 60–79
Blood pressure
classification Hypotension 41–65 Y M/F 3001–5000 – – – ≤ 90 ≤ 60
(Low BP)
Hypertension 60+ Y M/F 5000+ – – – ≥140 ≥ 90
(High BP)

The extracted information is then used for feature construc- label (such as positive, neutral, negative, or strong negative) for
tion. For example, if the BP data are found in the patient family each sentence of the published content. The rules to classify the
profile, then a 1 appears as the Family feature. If data about BP are patient’s mental health based on the predicted polarity are shown
not found in diagnosis and medical records, then a null value is in Table 4.
used in that feature. Later, all null values are replaced with a zero. In drug side effect prediction, the current intake diabetes and
After generating and preprocessing the features, we developed BP drugs are used as input queries, and we retrieved reviews
a model based on Bi-LSTM, named LSTMb , which is capable of about them from different websites. After filtering and prepro-
keeping temporal patterns present in historical sequenced data. cessing, the reviews are automatically labeled as having no side
For the classification of BP, all the parameters and generated
effects, moderate side effects, and severe side effects. For this la-
features at each time step t are converted into category vectors
beling, the sentiment polarity of each sentence is identified. Then,
{0, 1, or 2}.
positive, neutral, and negative sentiment polarity are considered
In mental health monitoring, patient messages and posts are
as having no side effects, moderate side effects, and severe side
extracted from social networks and are filtered using super-
vised and unsupervised approaches. However, text related to effects, respectively, as shown in Table 4. The 200-dimensional
depression and stress are usually short, unstructured, contain Word2vec model is trained to represent the drug reviews as a
negative emotions, and low sentiment polarity. Therefore, the sequence of words in the LSTMd model. The output of LSTMd is
text mining and ML approaches are used to efficiently handle then fed to the softmax function, which predicts drugs side effects
the social network content and identify the sentiment polarity accurately.
of a text with the terms happy, normal, depression, or stress.
Here, first the emotional texts are identified, and the sentiment
4. Experiment and results
analysis method is then applied to classify the text [55,58,80].
The Word2vec model with ontologies is used in this work to
represent the text for deep learning classifiers. We trained a In this section, the evaluation procedure of the proposed
200-dimensional Word2vec model, and then fed the sequence healthcare monitoring system is presented, and the results are
of words into the LSTMc model. The output of LSTMc is then discussed. Experiments were conducted on four different types
assigned to the softmax function, which predicts the sentiment of dataset, which were extensively discussed in Section 3.1.
36 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43

Table 4
Rules to classify a patient’s mental health and drug results.
Category Sentiment polarity about patient posts and Sentiment polarity about
comments (Positive/Neutral/Negative/Strong anti-diabetic drugs
Negative) (Positive/Neutral/Negative)
Happy Positive –
Patient mental Normal Neutral –
health classification Depressed Negative –
Stressed Strong Negative –
No side effects – Positive
Drug side effect
Moderate side effects – Neutral
prediction
Severe side effects – Negative

4.1. Performance evaluation This algorithm may not handle the textual data to generate the
accurate results. It consumes a lots of time for data processing and
To evaluate the efficiency of the proposed approach, various the classification results are worse than others [78]. CNN has been
experiments were conducted on the four different datasets. The compared with RNN in terms of sentiment classification [77]. The
datasets were extracted from the patient’s body and from so- authors described that CNN and RNN are useful for informative
cial networks using wearable sensors and APIs, respectively. The features extraction and sequences of unit modeling, respectively.
extracted datasets were then sent to the Hadoop cloud environ- CNN utilizes the fixed size of a window that moves over a textual
ment, which is a platform for big data analysis and processing. data to extract features from a sequence of terms. However, CNN
Hadoop MapReduce analyzed the extracted datasets for further misses the semantic meaning of words and unable to extract
processing. Furthermore, the generated datasets were then used the valuable words from lengthy sentences. In this evaluation,
to build the models for the classification of diabetes, blood pres- we selected the following algorithms based on the comparative
sure, mental health, and drug side effects. The word embedding study and discussions along with their limitations reported in our
approach was used to represent the social networking data and recent work [55,58].
drug reviews for the classification models. The number of features
utilized from diabetes and BP datasets, and the dimensionality of • Convolution neural network (CNN): CNN has been applied
for various classification tasks using sensors and textual
word embedding for drug reviews and social networks datasets
data [86–88]. We compared it with our proposed model in
are discussed as follows.
order to understand the efficiency of the proposed classifier.
Diabetes and BP datasets: For diabetes classification, the Pima
We used the CNN of the WEKA [74] library along with a
Indians Diabetes dataset was acquired from the UCI machine
sigmoid activation function.
learning repository [81]. The Pima Indians dataset contains the
• Multilayer perceptron (MLP): MLP is a neural network model
records of 768 patients, out of which 268 tested positive for
that learns the data pattern using several layers with con-
diabetes, while 500 tested normal. There are eight input at-
nected perceptrons [89]. We trained the MLP of WEKA
tributes in the dataset. However, only six attributes are used for
library with 3 hidden layers and a squared error function.
training the diabetes classification model. The age, family, gender,
The sigmoid function was used as an activation function.
activities, BMI, blood pressure, and blood sugar features are used
• Support vector machine (SVM): SVM with linear kernel
for diabetes classification. The dataset from the PhysioNet multi-
function is used to classify the datasets, and compare it with
parameter intelligent monitoring in intensive care II (MIMIC-II)
the proposed model [83]. SVM is suitable for light training
database was used to train the BP classification model [82]. This
data, which comprises non-zero values. However, it is time-
dataset comprises a large number of features, including BP and
consuming during sparse data training. SVM with a training
HR. However, only nine attributes are utilized for training the BP
parameter kernel-type radial basis function is considered in
classification model. The age, gender, BMI, heart rate, systolic BP,
the proposed work.
diastolic BP, activities, blood sugar, and family history features
• Fuzzy classifier: This classifier is a rules-based classifier that
are used for BP classification. Furthermore, we also combined our
categorizes the data using fuzzy rules in the form of if–
new datasets regarding diabetes and BP with the abovementioned
then. We utilized fuzzy classifier with fuzzy unordered rule
datasets. Thus, the total number of instances for diabetes and BP
induction algorithm [90].
classification are 868 and 550, respectively.
• Logistic regression: This algorithm is applied to define the
Drug reviews and social networking datasets: The dataset for drug
association between a binary response variables [84]. It pre-
side effect prediction was acquired from the UCI repository [32].
dicts the categorical outcomes based on certain predictors. It
This dataset consists of six attributes. However, only two at-
utilizes the relationship function that transforms the prob-
tributes (the name of the drug and patient reviews) are consid-
ability range [0,1] into (−∞, + ∞). We applied logistic
ered in the proposed work. We extracted patient messages and
regression classifier with the training parameter ridge es-
posts from social networks for patient mental health monitoring.
timator.
The social networking and drug reviews dataset are extensively • Random forest: This algorithm constructs a forest of random
discussed in Section 3.1. We trained a 200-dimensional Word2vec trees for classification [53]. It combines multiple decision
model to represent both the social networking and drugs reviews trees to achieve the accurate final results. We used it with
datasets for classification. number of iterations 100 and seed 1.
Machine learning algorithms such as logistic regression and • K-nearest neighbors (KNN): The KNN is non-parametric ap-
SVM have been utilized to classify the structured and unstruc- proach that is utilized for the purpose of both regression and
tured data [83–85]. These algorithms are shallow and trained on classification. KNN first stores all cases, and then classifies
sparse and high-dimensional features, which may not achieve the new cases based on the similarity measures. We utilized
the required results. Deep learning algorithms including MLP and KNN at k = 3 with the function of Euclidean distance.
CNN have been utilized for the purpose of textual data classifi-
cation. However, MLP performed well for small amount of input The above-mentioned models were trained by randomly dis-
data. MLP input is constant and link only with the current instant. tributing the datasets into training and testing sets at 70% and
 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43 37

30%, respectively. The proposed models were used both with an the best results because its memory function arranges the text
ontology and without an ontology, and performance was eval- in two directions that affect the text classification results. It is
uated and compared, each model with the others. In addition, important to note that the LSTMc outputs of positive, neutral,
PCA and IG were utilized with the proposed models and the negative, and strong negative indicate the patient’s mental health
results are compared. The WEKA API was applied to train the as happy, normal, depressed, and stressed, respectively, as listed
word embedding model and to evaluate the effectiveness of the in Table 4.
classifiers. This system was developed employing WEKA and the To evaluate the classification results of the proposed model
Protégé OWL tool with Java. from using drug reviews, the LSTMd model was compared with
other classifiers, as shown in Table 8. LSTMd and logistic regres-
4.2. Results sion obtained high accuracy at 90% and 83%, respectively. Other
classifiers achieved lower accuracy in comparison with LSTMd
To assess the usefulness of the aforementioned models, dif- and logistic regression. RMSE and MAE of LSTMd are 32 and 13,
ferent evaluation metrics were used, including precision, recall, respectively, which is lower than the other classifiers. The accu-
accuracy, function measures, root mean square error (RMSE), racy of the CNN is very low, and its RMSE is very high, compared
and mean absolute error (MAE). In this section, the results of to the other classifiers. This indicates that LSTMd can handle
the above-mentioned experiments are presented. The obtained longer sequences of word vectors as the dimensions of word
results from diabetes classification are shown in Table 5. The vectors increase. In contrast, the CNN results in over-fitting when
different classifiers (CNN, MLP, SVM, fuzzy logic, logistic regres- the sequence of words and the dimension of word embedding
sion, random forest, and KNN) were compared with the proposed increase. It is important to note that the LSTMd model outputs
LSTMa model using the Pima Indians dataset. The results show of positive, neutral, and negative show the drug side effects as
that the proposed LSTMa obtained the highest accuracy (75%) and having no side effects, moderate side effects, and severe side
the lowest MAE (26%) in comparison with the other classifiers. effects, respectively.
This high accuracy indicates that LSTMa stores more important
information in the memory cells for diabetes classification. In 4.3. Accuracy and MAE of the proposed models with ontology, PCA,
addition, the lowest MAE shows the better performance of the and IG
LSTMa model. The fuzzy classifier obtained the lowest RMSE (46)
We utilized ontology-based semantic knowledge with Bi-LSTM
compared to the other classifiers. Based on this experiment, we
in order to classify the social networking data and drug reviews.
observe that the other classifiers are time consuming, and their
The proposed ontology presented semantic knowledge of the
performance decreased for even a small number of features. How-
features related to depression and diabetes drugs. Fig. 7 presents
ever, LSTM outperformed the other classifiers in the prediction of
the accuracy and MAE of the proposed LSTMc and LSTMd models
diabetes.
using an ontology and without using an ontology. As can be seen,
Table 6 shows the results obtained by the proposed model and
the proposed ontology-LSTM shows significant improvement over
the other classifiers in terms of BP classification. These classifiers
simple LSTM. In terms of emotional text classification, the ac-
were trained using the PhysioNet MIMIC-II dataset. We observe
curacy of LSTMc was 84%, which increased to 87% when using
that LSTMb achieved the highest accuracy (88%) compared to
the ontology. However, MAE decreased by just 1. In addition,
the CNN (70%), MLP (80%), the SVM (73%), the fuzzy classifier
the accuracy and MAE of LSTMd was 90% and 14, respectively,
(83%), logistic regression (72%), random forest (70%), and KNN
in terms of drug side effect classification. However, this accu-
(58%). However, the MAE and RMSE of the proposed model were
racy increased to 93%, and MAE decreased by 1 when using
12 and 34, respectively, which are lower than the other classi- the ontology with LSTMd . The obtained results indicate that the
fiers. As shown in Table 6, the accuracy of the fuzzy classifier is proposed system can detect the most efficient features in the text
surprisingly high. This is because the training data are handled and provide additional information about them. In addition, the
by generated fuzzy rules. In addition, the results with KNN are proposed ontologies explore the domain information of diabetes
surprisingly poor. The value of K affected the results from KNN. If drugs and depression. Therefore, they identify depression and
the value of K is too big, then KNN misclassifies the test sample. drug features in the dataset, and help word embedding and
This is because of the large distance between data points and the LSTM to understand the exact meanings of the features in the
test sample of the neighborhood. task of text classification. The generated results indicate that the
We extracted patient emotional comments and posts, and proposed system outperforms in comparison with simple LSTM
drug reviews from social networking and drug-related webpages, and the other classifiers.
as discussed in Section 3.1. The emotional posts were labeled We applied PCA and IG to further improve the accuracy of
automatically using our proposed method, as explained in Sec- the proposed models. PCA combined the original features of each
tion 3.4. The patient’s mental health was classified using social dataset to generate new independent features. However, only a
network data, and the side effects of diabetes drugs were pre- few features are selected from the newly generated features. The
dicted using drug reviews. To accomplish these tasks, Word2vec selected features are less in number than the original features and
models were trained to represent these data in the form of word that is how the dimensionality is reduced.
vectors with high dimensionality. PCA first normalized features using mean and standard de-
Table 7 shows the performance of the classifiers in terms of viation function, which is the square root of the variance. PCA
mental health classification. According to Table 7, LSTMc using then used covariance function in order to know the correlation
softmax shows the highest accuracy (89%) in comparison with between those features. Finally, the new features are selected
the other classifiers. In addition, the MAE and RMSE of LSTMc are based on their eigenvalues to reduce the dimension. For example,
15 and 35, respectively, which is lower than the other classifiers. PCA reduced the dimensionality of the diabetes dataset by trans-
However, the fuzzy classifier and random forest obtained lower forming six features into five features, as shown in Table 9. Vector
accuracy in comparison with the other classifiers. Random forest Y1 , Y2 , Y3 , Y4 , and Y5 are called principal components. These new
faces over-fitting problems due to its complex representation and vectors are selected because of their high eigenvalues compared
noise in the dataset. Based on this experiment, we observe that to other vectors. In this research work, the PCA with cumulative
Word2vec and LSTM with softmax are better than the other clas- variance rate of 90% is applied to reduce the dimensionality of
sifiers in terms of emotional text classification. LSTMc presented datasets.
38 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43

Table 5
Comparison of the LSTMa model with other classifiers in terms of diabetes classification based on six features.
Proposed models and other classifiers Precision (P) (%) Recall (R) (%) Function Measure (FM) (%) Accuracy (Ac) (%) RMSE MAE
CNN 62 66 63 66 58 34
MLP 67 67 67 68 51 34
SVM 67 70 68 70 54 30
Fuzzy classifier 73 72 65 72 52 26
Logistic regression 70 68 69 69 55 32
Random forest 67 70 66 70 46 42
KNN 57 65 59 65 52 50
LSTMa 74 75 75 75 50 26

Table 6
Comparison of the LSTMb model with other classifiers in terms of BP classification based on nine features.
Proposed models and other classifiers Precision (P) (%) Recall (R) (%) Function Measure (FM) (%) Accuracy (Ac) (%) RMSE MAE
CNN 71 70 69 70 54 30
MLP 80 80 80 80 37 22
SVM 83 75 73 74 50 25
Fuzzy classifier 84 83 83 83 40 16
Logistic regression 72 72 71 72 51 29
Random forest 78 70 67 70 45 42
KNN 60 58 56 58 64 42
LSTMb 89 87 87 88 34 12

Table 7
Comparison of the LSTMc model with other classifiers in terms of mental health classification based on 200-dimensional Word2vec model.
Proposed classifier model and other classifiers Precision (P) (%) Recall (R) (%) Function Measure (FM) (%) Accuracy (Ac) (%) RMSE MAE
CNN 66 66 66 66 56 34
MLP 72 71 71 72 47 29
SVM 68 68 68 68 56 32
Fuzzy classifier 60 58 56 59 67 45
Logistic regression 70 75 72 71 44 34
Random forest 60 60 59 60 58 40
KNN 64 63 64 64 60 36
LSTMc 84 83 85 84 35 15

Table 8
Comparison of the LSTMd model with other classifiers in terms of drug side effect classification based on 200-dimensional Word2vec model.
Proposed classifier models and other classifiers Precision (P) (%) Recall (R) (%) Function Measure (FM) (%) Accuracy (Ac) (%) RMSE MAE
CNN 76 68 65 68 62 39
MLP 82 81 81 81 39 20
SVM 82 82 82 82 36 18
Fuzzy classifier 79 78 78 78 41 28
Logistic regression 84 82 83 83 36 17
Random forest 76 66 63 66 58 33
KNN 77 75 75 76 49 24
LSTMd 88 90 89 90 32 14

Table 9
PCA-based selected features from diabetes dataset.
Vectors New features Eigenvalues
Y1 0.205 Glucose = 172+0.205, Diabetes Pedigree Function = 0.702+0.205, Insulin = 2.799
579-0.194, Pregnancies = 17-0.194, Diabetes Pedigree Function = 0.817...

Y2 −0.304 Diabetes Pedigree Function = 0.817-0.304, Pregnancies = 17-0.299, BMI = 2.107

40.9-0.27, Insulin = 114-0.224, Glucose = 163...

Y3 0.311 Insulin = 240+0.269, BMI = 45.4+0.269, Diabetes Pedigree Function = 1.500

0.721+0.242, Blood Pressure = 110+0.217, Glucose = 171...

Y4 0.216 Glucose = 172+0.216, Diabetes Pedigree Function = 0.702+0.216, Insulin = 1.092

579+0.212, Insulin = 240+0.201, Diabetes Pedigree Function = 0.721...

Y5 0.185 Skin Thickness = 7+0.176, Pregnancies = 12+0.162, Diabetes Pedigree 0.677

Function = 0.926+0.162, Insulin = 258-0.158, BMI = 43.1...

After applying PCA, IG evaluated the value of each feature by features. The dataset of diabetes, BP, mental health, and drug
computing the IG with respect to the class. IG did not reduce the side effect contain 6, 9, 200, and 200 features, respectively. The
PCA-based selected features in case of diabetes classification, and PCA and IG reduced the input data features to 5, 7, 99, and 67,
considered that all the five features provide maximal informa- respectively. The obtained results using the classification models
tion. Similar procedure of PCA and IG is applied for other three along with PCA and IG are shown in Fig. 8. PCA and IG along with
datasets. However, IG selected the best features in case of other ontology-LSTM reduced the dimensionality of a large dataset that
datasets that give useful information, and removed the unrelated still contains most of the useful information. This shows that PCA
 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43 39

processing, data preprocessing using data mining techniques, fea-

ture extraction, and dimensionality reduction, and the role of
domain-specific ontologies in deep learning-based classification.
The proposed big data analytics engine automatically detects
valuable information, extracts useful features from healthcare
data, reduces the dimensionality of data, classifies patient health
conditions, and predicts drug side effects. This system helps warn
diabetes and BP patients before risks to their health reach a high
level. In addition, this system supports physicians to offer actual
treatments to their patients by smartly monitoring their patients’
health conditions. This method can store and analyze healthcare
big data, extract valuable features, and provide the semantic
meanings of features in order to improve the performance of
health condition classification. In this context, this framework can
be useful in healthcare sectors for monitoring chronic patients
using their structured and unstructured data.
Despite its potentials, this work has a couple of limitations.
Fig. 7. Accuracy and MAE comparisons of the LSTMc and LSTMd models with
ontology-LSTMc and ontology-LSTMd .
First, the SentiWordNet lexicons are not enough to precisely
understand the textual data for the prediction of patient mental
health and drug side effects. Therefore, it would be worth to use
different sentiment lexicons for the improvement of the proposed
system. Second, the proposed framework does not consider the
multimodal data such as videos, emoticons, images, etc., which
might be integrated with the said model to further improve the
system performance. Also, the presented model can be extended
taking a fuzzy ontology with fuzzy LSTM into account to enhance
the classification performances in healthcare context.

CRediT authorship contribution statement

Farman Ali: Conceptualization, Methodology, Software, Visu-

alization, Writing - original draft. Shaker El-Sappagh: Resources,
Validation, Investigation. S.M. Riazul Islam: Writing - review
& editing. Amjad Ali: Formal analysis, Software. Muhammad
Attique: Data curation, Resources. Muhammad Imran: Writing -
review & editing. Kyung-Sup Kwak: Funding acquisition, Project
Fig. 8. Proposed classification model accuracy with and without PCA and IG. administration, Supervision.

Declaration of competing interest

leads to an approximately 2x reduction in training time of LSTM
models. Based on the obtained results in Fig. 8, we note that the The authors declare that they have no known competing finan-
accuracy of LSTMa , LSTMb , ontology-LSTMc , and ontology-LSTMd
cial interests or personal relationships that could have appeared
with PCA and IG increased by 4%, 1%, 3%, and 1%, respectively.
to influence the work reported in this paper.
This indicates that using PCA and IG with ontology-LSTM is more
efficient than the other classifiers in terms of numeric and textual
data classification. In addition, it also shows that Word2vec and Acknowledgments
LSTM without an ontology miss the semantic meanings of fea-
tures related to the drug and disease domains. The main reason This work was supported in part by National Research Founda-
is that many hidden features of diabetes and depression represent tion of Korea-Grant funded by the Korean Government (Ministry
different notions. of Science and ICT-NRF-2020R1A2B5B02002478), and in part by
Sejong university through its faculty research program. Attique’s
5. Conclusion work is supported by the Basic Science Research Program through
the National Research Foundation of Korea funded by the Ministry
In this work, a novel healthcare monitoring framework for of Education under grant 2020R1G1A1013221. Imran’s work is
chronic patients was presented, which integrates advanced tech- supported by the Deanship of Scientific Research at King Saud
nologies, including data mining, cloud servers, big data, ontolo- University through research group project number RG-1435-
gies, and deep learning. The proposed framework enhances the 051. Shaker’s work is supported by the Spanish Ministry of
performance of heterogeneous data handling and processing, and Science, Innovation and Universities (grants RTI2018-099646-B-
improves the accuracy of healthcare data classification. The pro- I00, TIN2017-84796-C2-1-R, TIN2017-90773-REDT, and RED2018-
posed method correctly examines diabetes and blood pressure 102641-T) and the Galician Ministry of Education, University and
(BP) patients using various sources for their data, such as smart- Professional Training (grants ED431F 2018/02, ED431C 2018/29,
phones, wearable sensors, medical records, and social networks ED431G/08, and ED431G2019/04), with all grants co-funded by
(called big data). Various reasonable issues were discussed, in- the European Regional Development Fund (ERDF/FEDER pro-
cluding big data storage in a cloud server, MapReduce-based data gram).
40 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43

References [23] L. Tsao, L. Li, L. Ma, Human work and status evaluation based on
wearable sensors in human factors and ergonomics: A review, IEEE Trans.
[1] C.W. Song, H. Jung, K. Chung, Development of a medical big-data mining Hum.-Mach. Syst. 49 (2019) 72–84, http://dx.doi.org/10.1109/THMS.2018.
process using topic modeling, Clust. Comput. (2017) 1–10, http://dx.doi. 2878824.
org/10.1007/s10586-017-0942-0. [24] M.S. Hossain, G. Muhammad, Emotion-aware connected healthcare big
data towards 5g, IEEE Internet Things J. 5 (2018) 2399–2406, http://dx.
[2] J. Peral, A. Ferrandez, D. Gil, R. Munoz-Terol, H. Mora, An ontology-oriented
doi.org/10.1109/JIOT.2017.2772959.
architecture for dealing with heterogeneous data applied to telemedicine
systems, IEEE Access 6 (2018) 41118–41138, http://dx.doi.org/10.1109/ [25] F. Ahamed, F. Farid, Applying internet of things and machine-learning for
ACCESS.2018.2857499. personalized healthcare: Issues and challenges, in: 2018 Int. Conf. Mach.
Learn. Data Eng., 2019, pp. 19–21, http://dx.doi.org/10.1109/icmlde.2018.
[3] T. Nguyen Gia, I. Ben Dhaou, M. Ali, A.M. Rahmani, T. Westerlund, P. Lilje-
00014.
berg, H. Tenhunen, Energy efficient fog-assisted iot system for monitoring
diabetic patients with cardiovascular disease, Future Gener. Comput. Syst. [26] R. Lopes Rosa, G.M. Schwartz, W. Vicente Ruggiero, D. Zegarra Rodriguez,
93 (2019) 198–211, http://dx.doi.org/10.1016/j.future.2018.10.029. A knowledge-based recommendation system that includes sentiment anal-
ysis and deep learning, IEEE Trans. Ind. Inform. 3203 (2018) 1–12, http:
[4] M. Saravanan, R. Shubha, A.M. Marks, V. Iyer, SMEAD: A secured mobile
//dx.doi.org/10.1109/TII.2018.2867174.
enabled assisting device for diabetics monitoring, in: 2017 IEEE Int. Conf.
Adv. Networks Telecommun. Syst, 2017, pp. 1–6. [27] Y. Chen, B. Zhou, W. Zhang, W. Gong, G. Sun, Sentiment analysis based on
deep learning and its application in screening for perinatal depression, in:
[5] G. Alfian, M. Syafrudin, M.F. Ijaz, M.A. Syaekhoni, N.L. Fitriyani, J. Rhee, A
Proc. - 2018 IEEE 3rd Int. Conf. Data Sci. Cyberspace, DSC 2018, 2018, pp.
personalized healthcare monitoring system for diabetic patients by utiliz-
451–456, http://dx.doi.org/10.1109/DSC.2018.00073.
ing BLE-based sensors and real-time data processing, Sensors (Switzerland)
[28] M. Villari, L. Carnevale, A. Celesti, A. Galletta, G. Fiumara, Applying
18 (2018) http://dx.doi.org/10.3390/s18072183.
artificial intelligence in healthcare social networks to identity critical
[6] S. El-Sappagh, F. Ali, S. El-Masri, K. Kim, A. Ali, K.S. Kwak, Mobile health
issues in patients’ posts, 2018, pp. 680–687, http://dx.doi.org/10.5220/
technologies for diabetes mellitus: Current state and future challenges,
0006750606800687.
IEEE Access 7 (2019) 21917–21947, http://dx.doi.org/10.1109/ACCESS.2018.
[29] D. Bell, E. Laparra, A. Kousik, T. Ishihara, M. Surdeanu, S. Kobourov,
2881001.
Detecting diabetes risk from social media activity, 2018, pp. 1–11.
[7] F. Ali, S.M.R. Islam, D. Kwak, P. Khan, N. Ullah, S. jo Yoo, K.S. Kwak,
[30] J.R. Reichert, K.L. Kristensen, R.R. Mukkamala, R. Vatrapu, A supervised
Type-2 fuzzy ontology–aided recommendation systems for iot–based
machine learning study of online discussion forums about type-2 diabetes,
healthcare, Comput. Commun. 119 (2018) 138–155, http://dx.doi.org/10.
in: 2017 IEEE 19th Int. Conf. E-Health Networking, Appl. Serv. Heal.
1016/j.comcom.2017.10.005.
2017. 2017-Decem, 2017, pp. 1–7, http://dx.doi.org/10.1109/HealthCom.
[8] S.A. Siddiqui, Y. Zhang, J. Lloret, H. Song, Z. Obradovic, Pain-free blood
2017.8210815.
glucose monitoring using wearable sensors: Recent advancements and
[31] E. Tutubalina, Z. Miftahutdinov, S. Nikolenko, V. Malykh, Medical concept
future prospects, IEEE Rev. Biomed. Eng. 11 (2018) 21–35, http://dx.doi.
normalization in social media posts with recurrent neural networks, J.
org/10.1109/RBME.2018.2822301.
Biomed. Inform. 84 (2018) 93–102, http://dx.doi.org/10.1016/j.jbi.2018.06.
[9] B.Y. Su, M. Enayati, K.C. Ho, M. Skubic, L. Despins, J. Keller, M. Popescu,
006.
G. Guidoboni, M. Rantz, Monitoring the relative blood pressure using a
[32] F. Gräßer, S. Kallumadi, H. Malberg, S. Zaunseder, Aspect-based sentiment
hydraulic bed sensor system, IEEE Trans. Biomed. Eng. 66 (2019) 740–748,
analysis of drug reviews applying cross-domain and cross-data learning,
http://dx.doi.org/10.1109/TBME.2018.2855639.
2018, pp. 121–125, http://dx.doi.org/10.1145/3194658.3194677.
[10] T. Arakawa, Recent research and developing trends of wearable sensors
[33] J. Liu, X. Jiang, Q. Chen, M. Song, J. Li, Adverse drug reaction re-
for detecting blood pressure, Sensors (Switzerland) 18 (2018) http://dx.
lated post detecting using sentiment feature, Iran. J. Public Health 47
doi.org/10.3390/s18092772.
(2018) 861–867, https://www.scopus.com/inward/record.uri?eid=2-s2.0-
[11] F. Ali, D. Kwak, P. Khan, S.H.A. Ei-Sappagh, S.M.R. Islam, D. Park, K.S. 85048895897&partnerID=40&md5=e603602caba00dab4ffdbfa1c85b8e32.
Kwak, Merged ontology and SVM-based information extraction and rec-
[34] M.D.P. Salas-Zárate, J. Medina-Moreira, K. Lagos-Ortiz, H. Luna-Aveiga, M.Á.
ommendation system for social robots, IEEE Access 5 (2017) 12364–12379,
Rodríguez-García, R. Valencia-García, Sentiment analysis on tweets about
http://dx.doi.org/10.1109/ACCESS.2017.2718038.
diabetes: An aspect-level approach, Comput. Math. Methods Med. 2017
[12] M. Chen, J. Yang, J. Zhou, Y. Hao, J. Zhang, C.H. Youn, 5g-smart diabetes: (2017) http://dx.doi.org/10.1155/2017/5140631.
Toward personalized diabetes diagnosis with healthcare big data clouds,
[35] M. Chandrashekar, R. Nagulapati, Y. Lee, Ontology mapping framework
IEEE Commun. Mag. 56 (2018) 16–23, http://dx.doi.org/10.1109/MCOM.
with feature extraction and semantic embeddings, in: Proc. - 2018 IEEE
2018.1700788.
Int. Conf. Healthc. Informatics Work. ICHI-W 2018, 2018, pp. 34–42, http:
[13] A. Adeli, M. Neshat, A fuzzy expert system for heart disease diagnosis, in: //dx.doi.org/10.1109/ICHI-W.2018.00012.
Proc. Int. MultiConference Engineeers Comput. Sci. Vol. I, 2010, pp. 1–6. [36] R. Maldonado, T.R. Goodwin, M.A. Skinner, S.M. Harabagiu, Deep
[14] A.R.M. Forkan, I. Khalil, A. Ibaida, Z. Tari, BDCaM: Big data for context- learning meets biomedical ontologies: Knowledge embeddings for
aware monitoring—A personalized knowledge discovery framework for epilepsy, in: AMIA... Annu. Symp. Proceedings. AMIA Symp. 2017, 2017,
assisted healthcare, IEEE Trans. Cloud Comput. 5 (2015) 628–641, http: pp. 1233–1242, http://www.ncbi.nlm.nih.gov/pubmed/29854192%0Ahttp://
//dx.doi.org/10.1109/tcc.2015.2440269. www.pubmedcentral.nih.gov/articlerender.fcgi?artid=PMC5977726.
[15] H.F. Nweke, Y.W. Teh, U.R. Alo, G. Mujtaba, Analysis of multi-sensor fusion [37] B. Pang, L. Lee, S.C.-1118704 D.O.-10. 3115/1118693. 111870.
for mobile and wearable sensor based human activity recognition, 2018, Vaithyanathan, Thumbs up?: sentiment classification using machine
pp. 22–26, http://dx.doi.org/10.1145/3224207.3224212. learning techniques, in: Proc. ACL-02 Conf. Empir. Methods Nat. Lang.
[16] G. Manogaran, R. Varatharajan, M.K. Priyan, Hybrid recommendation sys- Process. - 10. 2002, pp. 79–86.
tem for heart disease diagnosis based on multiple kernel learning with [38] A. Lamurias, L.A. Clarke, F.M. Couto, BO-LSTM: Classifying relations via
adaptive neuro-fuzzy inference system, Multimedia Tools Appl. 77 (2018) long short-term memory networks along biomedical ontologies, BMC
4379–4399, http://dx.doi.org/10.1007/s11042-017-5515-y. Bioinformatics 20 (2019) 10.
[17] M. Chen, W. Li, Y. Hao, Y. Qian, I. Humar, Edge cognitive computing based [39] V. Jagadeeswari, V. Subramaniyaswamy, R. Logesh, V. Vijayakumar, A study
smart healthcare system, Future Gener. Comput. Syst. 86 (2018) 403–411, on medical internet of things and big data in personalized healthcare
http://dx.doi.org/10.1016/j.future.2018.03.054. system, Health Inf. Sci. Syst. 6 (2018) 1–20, http://dx.doi.org/10.1007/
[18] N. Yuvaraj, K.R. SriPreethaa, Diabetes prediction in healthcare systems s13755-018-0049-x.
using machine learning algorithms on hadoop cluster, Clust. Comput. [40] A. Onal, O.B. Sezer, M. Ozbayoglu, E. Dogdu, A. Can Onal, Weather data
(2017) 1–9, http://dx.doi.org/10.1007/s10586-017-1532-x. analysis and sensor fault detection using an extended iot framework with
[19] S. Asthana, A. Megahed, R. Strong, A recommendation system for proactive semantics, big data, and machine learning spend project view project
health monitoring using iot and wearable technologies, in: Proc. - 2017 a multi-agent simulation and backtesting workbench software for deep
IEEE 6th Int. Conf. AI Mob. Serv. AIMS 2017, 2017, pp. 14–21, http: learning and evolutionary algor, 2017, http://dx.doi.org/10.1109/BigData.
//dx.doi.org/10.1109/AIMS.2017.11. 2017.8258150.
[20] Z.C. Lipton, D.C. Kale, C. Elkan, R. Wetzel, Learning to diagnose with LSTM [41] J.C. Kim, K. Chung, Mining health-risk factors using PHR similarity in
recurrent neural networks, 2015, pp. 1–18, http://dx.doi.org/10.14722/ndss. a hybrid p2p network, Peer-to-Peer Netw. Appl. 11 (2018) 1278–1287,
2015.23268. http://dx.doi.org/10.1007/s12083-018-0631-7.
[21] K. Vimalkumar, N. Radhika, A big data framework for intrusion detection, [42] S. Chakrabarti, A. Swetapadma, P.K. Pattnaik, Smart Techniques for a
2017, pp. 198–204. Smarter Planet, Springer International Publishing, 2019, http://dx.doi.org/
[22] S. Qingnan, M.V. Jankovic, B. Joao, B.L. Moore, P. Diem, C. Stettler, S. 10.1007/978-3-030-03131-2.
Mougiakakou, Predicting blood glucose with an LSTM and bi-LSTM based [43] M.I. Razzak, M. Imran, G. Xu, Big Data Analytics for Preventive Medicine,
deep neural network, in: 14th IEEE Symp. Neural Networks Appl., 2019. Springer London, 2020, http://dx.doi.org/10.1007/s00521-019-04095-y.
 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43 41

[44] F. Amalina, I.A. Targio Hashem, Z.H. Azizul, A.T. Fong, A. Firdaus, M. Imran, [66] F. Vitali, R. Lombardo, D. Rivero, F. Mattivi, P. Franceschi, A. Bordoni,
N.B. Anuar, Blending big data analytics: Review on challenges and a recent A. Trimigno, F. Capozzi, G. Felici, F. Taglino, F. Miglietta, N. De Cock,
study, IEEE Access 8 (2020) 3629–3645, http://dx.doi.org/10.1109/ACCESS. C. Lachat, B. De Baets, G. De Tré, M. Pinart, K. Nimptsch, T. Pischon, J.
2019.2923270. Bouwman, D. Cavalieri, ONS: An ontology for a standardized description of
[45] P.K. Sahoo, S.K. Mohapatra, S.L. Wu, SLA based healthcare big data analysis interventions and observational studies in nutrition, Genes Nutr. 13 (2018)
and computing in cloud network, J. Parallel Distrib. Comput. 119 (2018) 1–9, http://dx.doi.org/10.1186/s12263-018-0601-y.
121–135, http://dx.doi.org/10.1016/j.jpdc.2018.04.006. [67] W.A. Kibbe, C. Arze, V. Felix, E. Mitraka, E. Bolton, G. Fu, C.J. Mungall, J.X.
[46] K. Muhammad, A.K. Sangaiah, A. Abdelaziz, M. Elhoseny, A.S. Salama, A.M. Binder, J. Malone, D. Vasant, H. Parkinson, L.M. Schriml, Disease ontology
Riad, A hybrid model of internet of things and cloud computing to manage 2015 update: An expanded and updated database of human diseases for
big data in health services applications, Future Gener. Comput. Syst. 86 linking biomedical knowledge through disease data, Nucleic Acids Res. 43
(2018) 1383–1394, http://dx.doi.org/10.1016/j.future.2018.03.005. (2015) D1071–D1078, http://dx.doi.org/10.1093/nar/gku1011.
[47] N. El aboudi, L. Benhlima, Big data management for healthcare systems: [68] S. El-Sappagh, D. Kwak, F. Ali, K.S. Kwak, DMTO: A realistic ontology for
Architecture, requirements, and implementation, Adv. Bioinform. 2018 standard diabetes mellitus treatment, J. Biomed. Semant. 9 (2018) 1–30,
(2018) 1–10, http://dx.doi.org/10.1155/2018/4059018. http://dx.doi.org/10.1186/s13326-018-0176-y.
[48] A. Jindal, A. Dua, N. Kumar, A.K. Das, A.V. Vasilakos, J.J.P.C. Rodrigues, [69] Y. Lin, S. Mehta, H. Küçük-McGinty, J.P. Turner, D. Vidovic, M. Forlin,
Providing healthcare-as-a-service using fuzzy rule based big data analytics A. Koleti, D.T. Nguyen, L.J. Jensen, R. Guha, S.L. Mathias, O. Ursu, V.
in cloud computing, IEEE J. Biomed. Health Inform. 22 (2018) 1605–1618, Stathias, J. Duan, N. Nabizadeh, C. Chung, C. Mader, U. Visser, J.J. Yang,
http://dx.doi.org/10.1109/JBHI.2018.2799198. C.G. Bologa, T.I. Oprea, S.C. Schürer, Drug target ontology to classify and
[49] H. Habibzadeh, A. Boggio-Dandry, Z. Qin, T. Soyata, B. Kantarci, H.T. integrate drug discovery data, J. Biomed. Semant. 8 (2017) 1–16, http:
Mouftah, Soft sensing in smart cities: Handling 3vs using recommender //dx.doi.org/10.1186/s13326-017-0161-x.
systems, machine intelligence, and data analytics, IEEE Commun. Mag. 56 [70] M.Á. Rodríguez-García, R. Valencia-García, F. García-Sánchez, J.J. Samper-
(2018) 78–86, http://dx.doi.org/10.1109/MCOM.2018.1700304. Zapater, Ontology-based annotation and retrieval of services in the cloud,
[50] P. Verma, S.K. Sood, Cloud-centric iot based disease diagnosis healthcare Knowl.-Based Syst. 56 (2014) 15–25, http://dx.doi.org/10.1016/j.knosys.
framework, J. Parallel Distrib. Comput. 116 (2018) 27–38, http://dx.doi.org/ 2013.10.006.
10.1016/j.jpdc.2017.11.018. [71] C. Verma, M. Hart, S. Bhatkar, A. Parker-Wood, S. Dey, Improving scala-
[51] T. Lenc, P.E. Keller, M. Varlet, S. Nozaradan, Frequency tagging is sensitive bility of personalized recommendation systems for enterprise knowledge
to the temporal structure of signals, 9264 (2017) 17–24 http://dx.doi.org/ workers, IEEE Access 4 (2016) 204–215, http://dx.doi.org/10.1109/ACCESS.
10.5281/zenodo. 2015.2513000.
[72] A. Derungs, S. Soller, A. Weishaupl, J. Bleuel, G. Berschin, O. Amft,
[52] A. Dwivedi, R.P. Pant, S. Pandey, K. Kumar, Nternet of things’ (iot’s) impact
Regression-based, mistake-driven movement skill estimation in nordic
on decision oriented applications of big data sentiment analysis, in: Proc.
walking using wearable inertial sensors, in: 2018 IEEE Int. Conf. Pervasive
- 2018 3rd Int. Conf. Internet Things Smart Innov. Usages, IoT-SIU 2018,
Comput. Commun. PerCom 2018, 2018, http://dx.doi.org/10.1109/PERCOM.
2018, pp. 1–10, http://dx.doi.org/10.1109/IoT-SIU.2018.8519922.
2018.8444576.
[53] V. Vijayakumar, K. Shankar, S.K. Lakshmanaprabu, N. Chilamkurti, M.
[73] H. Uğuz, A hybrid system based on information gain and principal
Ilayaraja, A.W. Nasir, Random forest for big data classification in the
component analysis for the classification of transcranial doppler signals,
internet of things using optimal features, Int. J. Mach. Learn. Cybern. (2019)
Comput. Methods Programs Biomed. 107 (2011) 598–609, http://dx.doi.
http://dx.doi.org/10.1007/s13042-018-00916-z.
org/10.1016/j.cmpb.2011.03.013.
[54] G. Manogaran, R. Varatharajan, D. Lopez, P.M. Kumar, R. Sundarasekar, C.
[74] M. Hall, E. Frank, G. Holmes, P. Bernhard, P. Reutemann, I. Witten, The
Thota, A new architecture of internet of things and big data ecosystem for
WEKA Data Mining Software: An Update, 11 (2009) 10–18.
secured smart healthcare monitoring and alerting system, Future Gener.
[75] G. Rehman, D. Khan, A. Hussain, W. Ahmad, M. Hamayun, A. Ahmad, S.
Comput. Syst. 82 (2018) 375–387, http://dx.doi.org/10.1016/j.future.2017.
Khan, A. Iqbal, U.U. Khan, L. Huang, Intelligent hepatitis diagnosis using
10.045.
adaptive neuro-fuzzy inference system and information gain method, Soft
[55] F. Ali, D. Kwak, P. Khan, S. El-Sappagh, A. Ali, S. Ullah, K.H. Kim, K.-S. Kwak,
Comput. (2018) http://dx.doi.org/10.1007/s00500-018-3643-6.
Transportation sentiment analysis using word embedding and ontology-
[76] S. Balli, E.A. Sağbaş, M. Peker, Human activity recognition from smart
based topic modeling, Knowl.-Based Syst. (2019) http://dx.doi.org/10.1016/
watch sensor data using a hybrid of principal component analysis and
j.knosys.2019.02.033.
random forest algorithm, Meas. Control (United Kingdom) 52 (2018)
[56] A. Teixeira, Data extraction and preparation to perform a The example of
37–45, http://dx.doi.org/10.1177/0020294018813692.
a Facebook fashion brand page, (n.d.).
[77] W. Yin, K. Kann, M. Yu, H. Schütze, Comparative study of CNN and RNN for
[57] F. Ali, D. Kwak, P. Khan, S.M.R. Islam, K.H. Kim, K.S. Kwak, Fuzzy ontology- natural language processing, 2017, http://dx.doi.org/10.14569/IJACSA.2017.
based sentiment analysis of transportation and city feature reviews for 080657.
safe traveling, Transp. Res. Part C Emerg. Technol. 77 (2017) 33–48, http: [78] B.T. Pham, D. Tien Bui, I. Prakash, M.B. Dholakia, Hybrid integration of
//dx.doi.org/10.1016/j.trc.2017.01.014. multilayer perceptron neural networks and machine learning ensembles
[58] F. Ali, S. El-Sappagh, D. Kwak, Fuzzy ontology and LSTM-based text mining: for landslide susceptibility assessment at himalayan area (India) using GIS,
A transportation network monitoring system for assisting travel, Sensors Catena 149 (2017) 52–63, http://dx.doi.org/10.1016/j.catena.2016.09.007.
(Switzerland). 19 (2019) http://dx.doi.org/10.3390/s19020234. [79] S. Kim, D. Jin, H. Lee, Predicting drug-target interactions using drug-drug
[59] H. Htet, S.S. Khaing, Y.Y. Myint, Big Data Analysis and Deep Learning interactions, PLoS One 8 (2013) 1–12, http://dx.doi.org/10.1371/journal.
Applications, Springer Singapore, 2019, http://dx.doi.org/10.1007/978-981- pone.0080129.
13-0869-7. [80] F. Ali, E.-S. Shaker, A. Ali, K.S. Kwak, D. Kwak, Sentiment analysis of
[60] S. Din, A. Paul, Smart health monitoring and management system: Toward transportation using word embedding and LDA approaches, n.d, 2018.
autonomous wearable sensing for internet of things using big data analyt- [81] J.W. Smith, J.E. Everhart, W.C. Dicksont, W.C. Knowler, R.S. Johannes, Using
ics, Future Gener. Comput. Syst. 91 (2019) 611–619, http://dx.doi.org/10. the ADAP learning algorithm to forecast the onset of diabetes mellitus, in:
1016/j.future.2017.12.059. Proc. Annu. Symp. Comput. Appl. Med. Care. Am. Med. Informatics Assoc.,
[61] F. Ali, E.K. Kim, Y.G. Kim, Type-2 fuzzy ontology-based opinion mining 1988, pp. 261–265.
and information extraction: A proposal to automate the hotel reservation [82] C.M. Cabello, W.B. Bair, S.D. Lamore, S. Ley, S. Alexandra, S. Azimian, G.T.
system, Appl. Intell. 42 (2015) 481–500, http://dx.doi.org/10.1007/s10489- Wondrak, NIH Public Access 46 (2010) 220–231, http://dx.doi.org/10.1016/
014-0609-y. j.freeradbiomed.2008.10.025.The.
[62] S. Baccianella, A. Esuli, F. Sebastiani, SentiWordNet 3 . 0: An enhanced [83] F. Ali, K.S. Kwak, Y.G. Kim, Opinion mining based on fuzzy domain
lexical resource for sentiment analysis and opinion mining sentiwordnet, ontology and support vector machine: A proposal to automate online
Analysis (2010) 1–12, doi: 10.1.1.61.7217. review classification, Appl. Soft Comput. J. 47 (2016) 235–250, http://dx.
[63] D.C. Cavalcanti, R.B.C. Prudêncio, S.S. Pradhan, J.Y. Shah, R.S. Pietrobon, doi.org/10.1016/j.asoc.2016.06.003.
Good to be bad? Distinguishing between positive and negative citations in [84] G. Stewart, A. Kamata, R. Miles, E. Grandoit, F. Mandelbaum, C. Quinn, L.
scientific impact, in: Proc. - Int. Conf. Tools with Artif. Intell. ICTAI, 2011, Rabin, Predicting mental health help seeking orientations among diverse
pp. 156–162, http://dx.doi.org/10.1109/ICTAI.2011.32. undergraduates: An ordinal logistic regression analysis, J. Affect. Disord.
[64] F. Ali, S. Ei-Sappagh, P. Khan, K.S. Kwak, Feature-based transportation 257 (2019) 271–280, http://dx.doi.org/10.1016/j.jad.2019.07.058.
sentiment analysis using fuzzy ontology and sentiwordnet, in: 9th Int. [85] T. Young, D. Hazarika, S. Poria, E. Cambria, Recent trends in deep learning
Conf. Inf. Commun. Technol. Converg. ICT Converg. Powered By Smart based natural language processing [review article], IEEE Comput. Intell.
Intell. ICTC 2018, 2018, pp. 1350–1355, http://dx.doi.org/10.1109/ICTC. Mag. 13 (2018) 55–75, http://dx.doi.org/10.1109/MCI.2018.2840738.
2018.8539607. [86] S. Poria, E. Cambria, A. Gelbukh, Aspect extraction for opinion mining
[65] F. Ali, S. Ei-sappagh, L. Feng, K.S. Kwak, Onemli! - word2vec and with a deep convolutional neural network, Knowl.-Based Syst. 108 (2016)
LSTM-based offensive content detection, 2019, pp. 1480–1481. 42–49, http://dx.doi.org/10.1016/j.knosys.2016.06.009.
42 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43

[87] M. Porumb, S. Stranges, A. Pescapè, L. Pecchia, Precision medicine and Islam was with the University of Dhaka, Bangladesh as an Assistant Professor
artificial intelligence: A pilot study on deep learning for hypoglycemic and Lecturer at the Dept. of Electrical and Electronic Engineering (formerly Dept.
events detection based on ECG, Sci. Rep. 10 (2020) 1–17, http://dx.doi. of Applied Physics, Electronics & Communication Engineering) for the period
org/10.1038/s41598-019-56927-5. September 2005 to March 2014. In 2014, In 2014, he worked at the Samsung
[88] G. Hussain, M.K. Maheshwari, M.L. Memon, M.S. Jabbar, K. Javed, A CNN R&D Institute Bangladesh (SRBD) as a Chief Engineer at the Dept. of Solution Lab
based automated activity and food recognition using wearable sensor for for six months. His research interests include wireless communications, signal
preventive healthcare, Electronics 8 (2019) 1425, http://dx.doi.org/10.3390/ processing for communications, and enabling technologies for 5G and beyond.
electronics8121425.
[89] T. Moon, S. Hong, H.Y. Choi, D.H. Jung, S.H. Chang, J.E. Son, Interpolation
of greenhouse environment data using multilayer perceptron, Comput. Amjad Ali received the B.S. and M.S. degrees in
Electron. Agric. 166 (2019) 105023, http://dx.doi.org/10.1016/j.compag. computer science from the COMSATS Institute of In-
2019.105023. formation Technology, Pakistan, in 2006 and 2008,
[90] L.J.B. Caluza, Fuzzy unordered rule induction algorithm application ba- respectively, and the Ph.D. degree from the Electronics
sic programming language competence: A rule-based model, Indian J. and Radio Engineering Department, Kyung Hee Univer-
Sci. Technol. 12 (2019) 1–10, http://dx.doi.org/10.17485/ijst/2019/v12i12/ sity, South Korea, in June 2015. Since July 2015, he
142575. has been an Assistant Professor with the Department
of Computer Science, COMSATS University Islamabad,
Lahore Campus, Pakistan. From 2018 to 2019, he was
a Post-Doctoral Research Scientist with the UWB Wire-
Farman Ali is an Assistant Professor in the Depart- less Communications Research Center (formerly Key
ment of Software at Sejong University, South Korea. National IT Research Center), Department of Information and Communication
He received his B.S. degree in computer science from Engineering, Inha University, South Korea, and also with the Mobile Network
the University of Peshawar, Pakistan, in 2011, M.S. and Communications Lab, School of Electrical Engineering, Korea University,
degree in computer science from Gyeongsang National Anam-dong, Seoul, South Korea. His main research interests include multimedia
University, South Korea, in 2015, and a Ph.D. degree in transmission, cognitive radio networks, machine learning, Internet of Things,
information and communication engineering from Inha smart grid, and vehicular networks. He was a recipient of the Excellent Research
University, South Korea, in 2018, where he worked as a Award from the UWB Wireless Communications Research Center and from the
Post-Doctoral Fellow at the UWB Wireless Communica- Mobile Network and Communications Lab during his Post-doctorate studies.
tions Research Center from September 2018 to August
2019. His current research interests include sentiment Muhammad Attique is currently working as an Asso-
analysis / opinion mining, information extraction, information retrieval, feature ciate Professor in the department of Software, Sejong
fusion, artificial intelligence in text mining, ontology-based recommendation University, South Korea. He received the bachelor’s
systems, healthcare monitoring systems, deep learning-based data mining, fuzzy degree in information and communication systems en-
ontology, fuzzy logic, and type-2 fuzzy logic. He has registered over 4 patents gineering from the National University of Science and
and published more than 50 research articles in peer-reviewed international Technology, Pakistan, in 2008, and the Ph.D. degree
journals and conferences. He has been awarded with Outstanding Research in Computer Science and Engineering from Ajou Uni-
Award (Excellence of Journal Publications-2017) and the President Choice of the versity, South Korea, in 2017. His research interests
Best Researcher Award during graduate program at Inha University. include spatial queries, big data analysis, social net-
work analysis, information retreival, and health care
monitoring systems.
Shaker El-Sappagh was born in El-Behara, Egypt, in
1977. He received the bachelor degree in computer
science from Information Systems Department, Faculty Muhammad Imran is an Associate Professor in the
of Computers and Information, Cairo University, Cairo, College of Applied Computer Science at King Saud Uni-
Egypt, in 1997, and the master degree from the same versity, Saudi Arabia. He received a Ph. D in Information
university in 2007. He received the Ph.D. degrees in Technology from the University Teknologi PETRONAS,
computer science from Information Systems Depart- Malaysia in 2011. His research interest includes Mobile
ment, Faculty of Computers and Information, Mansura and Wireless Networks, Internet of Things, Big Data
University, Mansura, Egypt in 2015. In 2003, he joined Analytics, Cloud computing, and Information Security.
the Department of Information Systems, Faculty of His research is financially supported by several national
Computers and Information, Benha University, Banha, and international grants. He has completed a number
Egypt as a teaching assistant. In 2009, he joined the Collage of Science, King Saud of international collaborative research projects with
University as a lecturer. Since June 2016, he has been with the Department of reputable universities. He has published more than 250
Information Systems, Faculty of computers and Information, Benha University as research articles in peer reviewed, well-recognized international conferences
a lecturer. He has publications in clinical decision support systems and semantic and journals. Many of his research articles are among the highly cited and
intelligence. His current research interests include medical informatics, ontology most downloaded. He served as an Editor in Chief for European Alliance
engineering, distributed and hybrid clinical decision support systems, semantic for Innovation (EAI) Transactions on Pervasive Health and Technology. He is
data modeling, distributed database systems, big data, semantic query languages, serving as an associate editor for top ranked international journals such as
medical data encoding, medical terminology, semantic interoperability, descrip- IEEE Communications Magazine, IEEE Network, Future Generation Computer
tion logic, fuzzy logic, fuzzy mathematics, fuzzy database, semantic database, Systems, and IEEE Access. He served/serving as a guest editor for about two
cloud computing, data integration, semantic web, and fuzzy expert systems. dozen special issues in journals such as IEEE Communications Magazine, IEEE
Dr. El-Sappagh is a reviewer in many journals, and he is very interested in Wireless Communications Magazine, Future Generation Computer Systems, IEEE
the diseases diagnosis and treatment researches. He has built some publicly Access, and Computer Networks. He has been involved in about one hundred
available ontologies including diabetes diagnosis ontology which is publicly peer-reviewed international conferences and workshops in various capacities
available in BioPortal site at https://bioportal.bioontology.org/ontologies/DDO, such as a chair, co-chair and technical program committee member. He has
SNOMED CT OWL 2 ontology at https://bioportal.bioontology.org/ontologies/ been consecutively awarded with Outstanding Associate Editor of IEEE Access
SCTTO, and DMTO diabetes treatment OWL 2 ontology available at https:// in 2018 and 2019 besides many others.
bioportal.bioontology.org/ontologies/DMTO.
Kyung-Sup Kwak (M’81) received the Ph.D. degree
from the University of California at San Diego in 1988.
S.M. Riazul Islam (M’10) received the B.S. and M.S. From 1988 to 1989, he was with Hughes Network
degrees in Applied Physics and Electronics from Uni- Systems, San Diego, CA, USA. From 1989 to 1990, he
versity of Dhaka, Bangladesh in 2003, and 2005, was with the IBM Network Analysis Center, Research
respectively and the Ph.D. degree in Information Triangle Park, NC, USA. Since then, he has been with
and Communication Engineering from Inha University, the School of Information and Communication Engi-
South Korea in 2012. He has been working at Sejong neering, Inha University, South Korea, as a Professor,
University, south Korea as an Assistant Professor at the where he had been the Dean of the Graduate School of
Department of Computer Science and Engineering since Information Technology and Telecommunications from
March 2017. From 2014 to 2017, he worked at Inha 2001 to 2002. He has been the Director of the UWB
University, South Korea as a Research Professor at the Wireless Communications Research Center (formerly Key National IT Research
UWB Wireless Communications Research Center. Dr. Center), South Korea, since 2003. In 2006, he served as the President of Korean
 F. Ali, S. El-Sappagh, S.M.R. Islam et al. / Future Generation Computer Systems 114 (2021) 23–43 43

Institute of Communication Sciences, and in 2009, the President of Korea interests include wireless communications, UWB systems, sensor networks,
Institute of Intelligent Transport Systems. In 2008, he had been selected for WBAN, and nano communications. He was a recipient of the number of awards,
Inha Fellow Professor and now for Inha Hanlim Fellow Professor. Dr. Kwak including the Engineering College Achievement Award from Inha University, the
published more than 200 peer-reviewed journal papers and served as TPC/Track LG Paper Award, the Motorola Paper Award, the Haedong Prize of research, and
chairs/organizing chairs for several IEEE related conferences. His research various government awards from the Ministry of ICT, the President, and the
Prime Minister of Korea, for his excellent research performances