LITERATURE REVIEW

1. “ Providing Healthcare-as-a-Service Using Fuzzy Rule Based Big Data

Analytics in Cloud Computing ”, Anish Jindal, Amit Dua, Neeraj Kumar, Ashok
Kumar Das, Athanasios V. Vasilakos, and Joel Rodrigues -2018

Proposed System
A Healthcare-as-a-Service system is developed using vast collection of
heterogeneous data which is one of the biggest challenges which requires a
specialized approach. To address this challenge, a new fuzzy rule based classifier is
presented in this paper with an aim to provide Healthcare-as-a-Service. The pro-
posed scheme is based upon the initial cluster formation, retrieval, and processing of
the big data in cloud environ- ment. Then, a fuzzy rule based classifier is designed for
efficient decision making for data classification in the proposed scheme. To perform
inferencing from the collected data, membership functions are designed for
fuzzification and defuzzification processes.

Block Diagram

Fig 1.1: Network Model Framework (healthcare data acquisition and processing
module).
Methodology
The cloud is divided into a set of sub-clouds based on the parameters which are used
to find out the list of patients who are suffering from a particular disease. For
example, sub-cloud SA may be an age cloud, sub-cloud SB can be Blood Pressure
(BP) cloud and similarly other sub-clouds can also be formed for other. These sub-
clouds are the collection of various parameters that cover most of the diseases and
these sub-clouds are further divided into different clusters as shown in Fig. 2. The
clustering is done based on the modified EM[1] algorithm by calculating the
membership of each value with specified clusters, and updating the clusters
accordingly. For example, sub-cloud SA representing age cloud can be classified into
three clusters of different age groups such as - infants, adults and old aged per- sons
(Different clusters would be formed after consulting with an expert).

Fig. 1.2 : Flow of Information

Moreover, a fuzzy rule-based classifier has also been presented to classify fuzzy data
and to retrieve fuzzy queries as specified by the expert. Once the clusters are formed,
whenever a new record is received, cloudlets send different parameters to their
respective sub-clouds, where data is stored in different clusters according to their
membership value. The detailed de- scription about membership value of data is
given in the next section. The major benefit of using this scheme of different sub-
clouds for storing various parameters is that it classifies the records intelligently and
retrieve the results quickly. For exam- ple, the doctor or the expert can specify the
symptoms (in terms of parameters) and cloudlets will generate the results based on
the parameters of the patients stored in various sub-clouds. The novelty of our
approach lies in the fact that whenever a new disease breaks out, the doctors can
simply pass the symptoms for that disease in the system and the cloud will generate
the probable patients which manifest those symptoms. A doctor can easily quantify
the parameters according the symptoms of the disease. The parameters that cannot be
exactly quantified are fuzzified and processed accordingly.

The flow of information in the proposed scheme is as shown in Fig. 1.2. The data
gathered from the patient using various sensors is stored in the cloud as explained
above. The doctor then searches for the queries in the cloud of a particular disease.
The cloud provides the search results of those patients to the doctor which satisfies
the query.

Modules Description

1. Data Acquisition Layer: The data acquisition layer is responsible for gathering the
data of the patients from different geographical domains. This data can be generated
from the body sensor network, vehicular ad-hoc network, and network of devices
present in homes or at various hospitals. The data from vehicular network is
generated either by using body sensors or the sensors that are placed in the vehi- cles.
These vehicles are generally hospital vans and ambulances; however the patients can
send the data from their personal vehi- cles if these are equipped with appropriate
sensors. The devices in these networks send the data individually or can choose a
network head which oversees the transmission of data for each of these networks.
The intra-network communication amongst the devices is done using short range
communication techniques such as Bluetooth, ZigBee, Passive Radio Frequency
IDentifier (RFID), Ultra WideBand (UWB), and 60 GHz Millimeter Wave .

2. Transmission Layer:Second layer is the transmission layer which is responsi- ble

for all the transmission and communication from or to the network connected to the
cloud. For this purpose, differ- ent gateways are assumed to be located at various
places such as hospitals, building, and other parts of a geographical region. For
network to gateway communication, short range commu- nication techniques are
used such as Dedicated Short-Range Communication/ Wireless Access in Vehicular
Environment (DSRC/WAVE), and Dynamic Spectrum Access[2] (DSA). For
gateway to cloud interaction, long range communication tech- niques like Wi-Fi,
Worldwide Interoperability for Microwave Access (WiMax), and Long term
Evolution (LTE), and its vari- ant LTE Advanced are used [3].

3.Computational Layer: All the data generated from the networks or devices go to the
cloud server through transmission layer, where the programmed cloudlets receive the
data and store it on different sub-clouds based on its context. The results are then
computed in order to serve the requests arising from different clients. Cloud
computing is used for man- agement of health care services which is the basis of
Healthcare as-a-service (HaaS). The cloud services are used for real-time storage of
data sensed from patients’ body. This data is collected through various body sensors
which are deployed in patient’s body, or in the form of wearable sensors. Extending
to traditional storage utility of HaaS, our work not only gives fast and reliable
information to doctors, but can also predict the future diseases that a patient may
suffer. This requires providing symptoms of a disease that are given by an expert or
specialized doctor.

Open Issues

The data is widely heterogenous and there is imminent requirement of efficient

mining techniques for context retrieval and processing. Security issues for the storage
of sensitive data and accuracy can be further enhanced.

2. “Modeling Healthcare Quality via Compact Representations of

Electronic Health Records”,Jelena Stojanovic,Djordje
Gligorijevic,Vladan Radosavljevic, Nemanja Djuric, Mihajlo Grbovic,
and Zoran Obradovic-2017
Proposed System

A novel approach for learning low- dimensional, distributed representations of patient

EHRs is presented. As a first step, it is described how to apply state-of-the-art,
unsupervised neural language models for learning embed- dings of diseases and
applied clinical procedures from the EHR data of individual patients. Then, the
obtained embeddings are employed to find useful inpatient feature vectors, used to
train predictive models of the healthcare quality indicators in a supervised manner.

Block Diagram

Fig. 2.1 Pipeline of the proposed approach

Methodology

Assume there is a given a set R of N hospital inpatient discharge records

(representing a single hospital visit) and sets D of possible diseases and P procedures.
Then, a discharge record ri of the ith patient is defined as a sequence of diseases d
belonging to D and procedure p belonging to P at the end of hospital stay.
Here, Di is the number of diagnosed diseases and Pi is the number of applied
procedures in the sequence, so that Di + Pi = Hi and that record is represented as
ri=[(hi1,..,hiHi)], where hil can be a disease or a procedure in the sequence. Then,
using the set R, the objective is to find M-dimensional real-valued rep- resentations
vd belonging to RM for every disease d and vp belonging to RM for every procedure
p, such that similar diseases and procedures lie nearby in the joint M-dimensional
vector space and to use them to build a patient vector representation xi belonging to
RM for training predictive models of the healthcare quality indicators.

Before discussing applications to specific healthcare related prediction problems, it is

intuitive to introduce neural language models as applied to NLP. These methods take
advantage of word order, and assume that closer words in the word sequence are
statistically more dependent. Typically, a neural language model learns the
probability distribution of the next word given a fixed number of preceding words
that act as the context. A typical approach to approximate the probability distri-
bution is to use a neural network model architecture . The neural network is trained
by projecting the vectors for context words into a latent representation with multiple
non-linear hidden layers and the output softmax layer comprising W nodes, where W
is the vocabulary size (in our task equal to the number of diseases and procedures
(|D|+|P|) while attempting to predict word wt with high probability.

When working with large-scale data, the vocabulary size W can easily reach millions.
In those cases, training of the neural network becomes a challenging task, as updates
of word vectors become computationally expensive. For that reason, recent
approaches propose log-linear models which aim to reduce the computational
complexity. The use of hierarchi- cal softmax or negative sampling is shown to be
effec- tive in substantially speeding up the training process.

Fig. 2.2 Graphical

representation of the disease+procedure model
Modules

1. disease+procedure2vec Method : In this section , a disease+procedure2vec (dp2v)

approach for learning diseases and procedures representations is proposed (step 1 in
Fig. 2.1) that extend models of the recently proposed word2vec algorithm. The key
insight is that the patients’ lists of diseases and procedures from EHRs as sequences
of tokens can be represented, and each sequence as a sample from some unknown
language can be viewed. disease+procedur- e2vec uses central disease/procedure hi to
predict b dis- eases/procedures that come before and b diseases/ procedures that come
after it in the discharge record, an architecture known as the SkipGram. As a result,
diseases and procedures that often co-occur and have similar con- texts (i.e., with
similar neighboring diseases and proce- dures) will have similar representations as
learned by our model. Additionally, a continuous bag of words architecture (CBOW),
that uses context diseases and procedures to predict a central disease or procedure is
considered .The disease+prodedure2vec model was optimized using stochastic
gradient ascent, suitable for large-scale problems. However, computation of gradients
is proportional to the number of unique disease and procedures in the datasets, which
may be computationally expensive in practical tasks.

2. Vector Representation: The Skip-Gram[4] model was consistently more accurate

than the CBOW model, thus we opted to use this model in disease +procedures2vec
approach. Varying parameter b did not introduce much variation in the results for
SkipGram, thus we chose to set context neighborhood size to b =40, such that model
captures larger context and most of the diseases and procedures in that record. It was
observed that increasing parameter M improves the accuracy, however
dimensionality is increased, leading to a more complex model that is more difficult to
train. Dimensionality of the embedding space was set to M= 200. Finally, we used 25
negative samples in each vector update for negative sampling. Similarly to the
approach presented in [20], the most frequent diseases and proce- dures were sub-
sampled during the training phase.

3. Predictive Model: Several penalized linear models for regression and classifi-
cation tasks are used in our experiments. In particular, for regression problems we
apply linear regression. On the other hand, for the classification problem we use the
logistic regression model. Vector w is an unknown set of weights for both prediction
models, and IðÞ is an indicator function equal to 1 if the argument is true and 0
otherwise. In addition, for both models we explored a number of regularization
approaches, ranging from l1 Lasso to over- lapping group Lasso penalizations. We
summarized the training objectives of five penalized line models in Table 3, where l1
indicates Lasso norm and lq is norm of the non- overlapping groups, wi and wGi
indicate a single dimension of the weight vector and a group of dimensions defined
by the index set Gi, respectively. The index sets Gi for group Lasso models were
defined in groups of ten consecutive features, indexed from 1 to 10, 11 to 20, and so
on until M 9 to M (smaller groups showed better performance). For the overlapping
group Lasso the index sets were defined as 1 to 20, 11 to 30, and so on. All
parameters were set to be equal and chosen from range [0.01,0.1], determined
through cross-validation. In the conducted experiments, an implementation from the
efficient SLEP package is done.

Open Issues
3. “ A Hybrid Feature Selection with Ensemble Classification for
Imbalanced Healthcare Data: A Case Study for Brain Tumor
Diagnosis”, Shamsul Huda, John Yearwood, Herbert Jelinek,
Mohammad Mehedi Hassan, Giancarlo Fortino, and Michael
Buckland-2017

Proposed System

The proposed approach develops a globally optimized Arti- ficial Neural Network Input
Gain Measurment Approxima- tion (GANNIGMA) based hybrid feature selection which
is combined with an ensemble classification (GANNIGMA- ensemble) technique to
generate the diagnostic decision rule. The GANNIGMA[5] hybrid feature selection in
the proposed approach finds the significant features which help to generate a simplified
rule. Ensemble classifier improves the classification accuracy.
Block Diagram

Fig. 3.1 Feature Selection with Ensemble Classification Network

Methodology
The data collection procedure from different patients was approved by the University
of Sydney Human Ethics Committee. Neurology patients at the Royal Prince Alfred
Hospital, Sydney Australia with and without the 1p19q co-deletion variant of
ologodendroglioma .The system develops a globally optimized Arti- ficial Neural
Network Input Gain Measurment Approximation (GANNIGMA) based hybrid
feature selection which is combined with an ensemble classification (GANNIGMA-
ensemble) technique to generate the diagnostic decision rule. The GANNIGMA
hybrid feature selection in the proposed approach finds the significant features which
help to generate a simplified rule. Ensemble classifier improves the classifi-cation
accuracy.

Modules
1.Feature Selection: Filter approach can find the intrinsic relationships between the
individual diagnosis feature and tumor class. However, filter approach did not use
any performance evaluation criteria based on accuracies. In contrast, wrapper
approach uses accuracy based performance evaluation. Since the wrapper approach
uses a classification accuracy based performance evaluation criteria during training, it
can be ensured from wrapper approach that selected subset by the wrapper can
achieve a better peformance; however, it may take more computational cost. The
proposed hybrid approach integrates the knowledge about the intrinsic relationship
between a particular feature with corresponding class estimated by the filter in the
wrapper search process and takes advantages of the complementary properties of both
approaches. In our proposed tumor feature selection approach, a mutual information
(MI) based Maximum Relevance Minimum Redundancy (MRMR) [47], [48] filter
ranking heuristic is combined with the wrapper heuristic. The wrapper is taken as an
Artificial Neural Network (ANN) .

2. Computation of Global ANNIGMA score: Computation of Global ANNIGMA

score: Artificial Neural Network Input Gain Measurement Approx- imation
(ANNIGMA) [47] is computed from the wrapper training of ANN. In general, a
three-layer ANN has a struc- ture as presented in Fig 3. Let us assume that the input,
hidden and output layer are denoted as i, j, k. The logistic activation function is
denoted as ‘‘Q’’ as below

Q(z) = (1/(1 + exp(−z)))

The standard back propogation training algorithm of ANN provides locally optimized
parameters, which could be worse for imbalanced dataset. Therefore, a global
optimization approach has been adopted with the standard backpropagation training.
An Algorithm for Global Optimization Prob- lem (AGOP) proposed in [45] and [46]
is applied for optimal estimation of ANN parameters in the training of ANN. An
average optimal wrapper heuristic for a Global ANNIGMA (GANNIGMA) scores is
computed using an n-fold cross validation during the training of ANN.
 Fig 3.2 Multilayer Perceptron in Artificial Neaural Network

3.Ensemble Classification and Rule Generation : Following the feature selection,

classification of the test examples is performed using decision tree in combination
with Bootstrap aggregating or bagging machine learning algorithms. Bagging is a
simple algorithm which uses bootstrap sampling. Given a training dataset T
containing n examples, a sample of training examples, Tm, where m is 1 to M is
created by selecting n examples uniformly at random with replacement from T(some
examples can be selected repeatedly while some may not be selected at all). A
particular classifier Hm : m = 1.., M is learned based on the actual training set Tm .
Then a compound classifier (H ) is created by aggregating the particular classifiers. A
new instance ti is then classified to class cj according to the number of votes obtained
from particular classifiers Hm. A decision tree is constructed and based on that further
results are obtained.

Open Issues
Complications in combining filter approach and wrapper approach although accuracy
increases manifolds. GANNINGMA feature selection based approach ensemble is
computationally intensive. The imbalanced dataset in healthcare data is an inherent
limitation.
4. “Healthcare Big Data Voice Pathology Assessment Framework”, M
Shamim Hossain, and Ghulam Muhammad-2016

Proposed System
A cloud server receives the medical data from many different hospitals and clinics. In
this paper, the publisher mainly focuson on speech signals from patients with or
without vocal fold pathology. The speech signals are processed on the cloud server.
Feature extraction is done. These features are fed into a classification unit in the
cloud. In the classification unit, three machine learning algorithms, SVM, GMM[6],
and ELM, are used. A ranking sys- tem based on the scores from these algorithms is
introduced to provide a classification.
Block Diagram

Fig. 4.1 Block diagram of the proposed big data for voice pathology assessment.

Methodology
Big healthcare data is contributed to by various data sources. From these data
sources, relevant information is sent to the healthcare cloud datacenters for analysis
by a big data application. The data is delivered through intermediate communication
and processing, where it has been pre-processed for noise reduction, unreliability,
inconsistency, and analog-to-digital conversion. Sometimes this pre-processing is
done based on the opinion of the health- care professionals.

With the increasing volume of collected data extracted from heterogeneous sources, it
is necessary that this large dataset should be stored in a database or a distributed file
system for aggregation or processing by different healthcare stakeholders. In the
aggregation phase, at first, the feature is extracted from the original data (e.g., signal)
and then further processed by classification, normalization, fusion, or mod- elling.
The features can be extracted based on the advice of healthcare professionals or
domain experts. Then processed data (as training) and test data are fed through the
machine learning algorithm. Training is sometimes based on previous historical data.
Classification algorithms have the potential to detect abnormalities, such as with
voice pathol- ogy detection, or to assist with diagnosis as a reference for physicians.
A popular platform for parallel programming is Google’s MapReduce, which is
characterized by its sophisticated techniques of load balancing and fault tolerance.
Apache Hadoop’s big data platform can implement the MapReduce algorithm for
healthcare big data analysis. MapReduce is used in processing healthcare data where
response time is not so critical. For large healthcare data, the MapR (MapReduce
version 2) can be used for implementing machine learning algorithms, to provide
useful and accurate data for improved patient care. However, for the fast pro- cessing
as well as for continuous streaming of data, Apache Spark may be used. It has a set of
Application program interfaces (APIs) for machine learning. After data processing,
the meaningful healthcare data recommendations are sent to the relevant stakeholder.

Modules
1.Hospital Data :The hospital data for the VPA may vary in many ways. As we are
concerned with only speech signals, we focus on the diversity of these signals
captured in many hospitals. The speech signals may contain only sustained vowels
/a/, /i/, or /o/, or a combination of them. Alternatively, they may contain more
complex speech such as whole words, phrases, or sentences from a conversation,
spontaneous speech, or reading aloud. Therefore, speech signals differ in both content
and length. Speech signals may be recorded by different media and can be captured in
different environments, such as in a sound-treated room or in a normal office room.
The sampling frequencies may also differ. For each patient, there is information
about gender, age, smoking or non-smoking, weight, diagnosis, severity, and so on;
however, some fields in this information may be missing. This results in a very large
volume of data that are then transferred to the cloud for processing. At the processing
stage, we process only voice or speech signals; EHR data are kept in a separate index.
Due to the unstruc- tured nature of the data, we also have to filter out the poor quality
or incorrectly labeled samples; this should be done at least for the training data.

2.Feature Extraction: There are many feature extraction techniques involved in the
VPA. These include Mel-frequency cepstral coefficients (MFCC) , multi-
dimensional voice program (MDVP) , MPEG-7 low-level audio descriptors , IDP ,
and glottal noise parameters. Each of them has its own advantages and disadvan-
tages; however, after a careful consideration, we find that both the MPEG-7 audio
features and the IDP features provided good results in the literature, and they are not
too much affected by the diversity of the recorded signals. Therefore, we adopt these
two features in the proposed assessment.

3. Classification:We use three machine learning algorithms, each of which is of

different characteristics, in the proposed system. The classifiers that we use are the
SVM, the ELM, and the GMM. A simple description of these algorithms are
provided below:
a: SVM : The SVM has been intensely used in the literature for many two-class
problems because of its flexibility for determining the threshold for separating the
classes, good generalization of out-of-samples, producing a unique solution contrary
to other forms of the neural networks. In the SVM, a kernel function is used to map
the features from a lower to a higher- dimensional space to find an optimal
hyperplane to separate the samples of the classes.
b: ELM : The ELM is a powerful machine learning algorithm [7], and it has several
important properties including (i) it is fast, (ii) the solution is dense, and (iii) the
feature mapping can be in an infinite or a finite space. If there are M number of
hidden neurons in the ELM, bias bj of the j-th hidden neuron, input weight is aj and
the output weight is θj, then the output function o(x) of x training samples can be
calculated.
c: GMM : The GMM is a stochastic modeling technique, which is fre- quently used
in speech processing applications such as speech recognition, speaker recognition,
and environment detection. The GMM-based approach is fast during testing, because
the models are represented by a few number of parameters. In this approach, each
class is modeled by a GMM having several Gaussian mixtures. The number of
mixtures varies depending on the hidden variables.

Open Issues
There are different combinations for feature extraction and feature classifications
which make it tedious to work through every combination and perform the
experiment in order to get one final output. The accuracy for different combinations
differ and sometimes it is difficult to tell which of the combinations is better than
which.
 APPLICATIONS

Helps to filter highly heterogenous healthcare data into coherent data to be used for
mining.

Helps to take general symptoms during data acquisition and using it to predict the
disease and required diagnosis.

Helps to filter all kinds of data be it text, speech or images.

Has huge implications in day to day life as everyone comes across medical issues
every now and then.
REFERENCES

[1] M.E.Boon,L.P.Kok,andS.Beck,‘‘Histologicalvalidationofneural- network assisted

cervical screening: Comparison with the conventional approach,’’ Cell Vis., vol. 2,
pp. 23–27, 1995.
[2] K. Chan, T.-W. Lee, P. A. Sample, M. H. Goldbaum, R. N. Weinreb, and T. J.
Sejnowski, ‘‘Comparison of machine learning and traditional classifiers in glaucoma
diagnosis,’’ IEEE Trans. Biomed. Eng., vol. 49, no. 9, pp. 963–974, Sep. 2002.

[3] M. Cruz-Monteagudo, M. N. D. S. Cordeiro, and F. Borges, ‘‘Com- putational

chemistry approach for the early detection of drug-induced idiosyncratic liver
toxicity,’’ J. Comput. Chem., vol. 29, no. 4, pp. 533–549, Mar. 2008.

[4] D.Delen,A.Oztekin,andZ.Kong,‘‘Amachinelearning-basedapproach to prognostic

analysis of thoracic transplantations,’’ Artif. Intell. Med., vol. 49, no. 1, pp. 33–42,
May 2010. L.J.Mango,‘‘Computer-assistedcervicalcancerscreeningusingneural
networks,’’ Cancer Lett., vol. 77, nos. 2–3, pp. 155–162, Mar. 1994.
[5] M. Gadji, D. Fortin, A.-M. Tsanaclis, and R. Drouin, ‘‘Is the 1p/19q deletion a
diagnostic marker of oligodendrogliomas?’’ Cancer Genet. Cytogenetics, vol. 194,
no. 1, pp. 12–22, Oct. 2009.

[6] C.Ramirezetal.,‘‘Lossof1p,19q,and10qheterozygosityprospectively predicts

prognosis of oligodendroglial tumors—Towards individualized tumor treatment?’’
Neurol. Oncol., vol. 12, pp. 490–499, Feb. 2010.

[7] J. F. Parkinson et al., ‘‘The impact of molecular and clinical factors on patient
outcome in oligodendroglioma from 20 years’ experience at a single centre,’’ J. Clin.
Neurosci., vol. 18, no. 3, pp. 329–333, Mar. 2011.