Research ArticleResearch on Improved Depth Belief Network-BasedPrediction of Cardiovascular Diseases

Peng Lu ,1,2 Saidi Guo ,2,3 Hongpo Zhang,2,4 Qihang Li ,1,2 Yuchen Wang ,1

Yingying Wang ,2,3 and Lianxin Qi 2,3

1School of Electrical Engineering, Zhengzhou University, Zhengzhou 450001, China2Cooperative Innovation Center of Internet Healthcare, Zhengzhou University, Zhengzhou 450001, China3Industrial Technology Research Institute, Zhengzhou University, Zhengzhou 450001, China4State Key Laboratory of Mathematical Engineering and Advanced Computing, Zhengzhou Science and Technology Institute,Zhengzhou 450003, China

Correspondence should be addressed to Peng Lu; lupeng@zzu.edu.cn

Received 15 December 2017; Revised 27 March 2018; Accepted 5 April 2018; Published 9 May 2018

Academic Editor: Vincenzo Positano

Copyright © 2018 Peng Lu et al.)is is an open access article distributed under the Creative Commons Attribution License, whichpermits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Quantitative analysis and prediction can help to reduce the risk of cardiovascular disease. Quantitative prediction based ontraditional model has low accuracy. )e variance of model prediction based on shallow neural network is larger. In this paper,cardiovascular disease prediction model based on improved deep belief network (DBN) is proposed. Using the reconstructionerror, the network depth is determined independently, and unsupervised training and supervised optimization are combined. Itensures the accuracy of model prediction while guaranteeing stability. )irty experiments were performed independently on theStatlog (Heart) andHeart Disease Database data sets in the UCI database. Experimental results showed that themean of predictionaccuracy was 91.26% and 89.78%, respectively. )e variance of prediction accuracy was 5.78 and 4.46, respectively.

1. Introduction

Cardiovascular disease has become themost pathogenic diseasein our country [1]. )e establishment of a prediction model ofcardiovascular disease and the quantitative analysis of the riskof disease can effectively reduce the incidence of the disease [2].

In the past few decades, researchers have conducted a lotof research on the computer classification of ECG, such assupport vector machines (SVMs), artificial neural networks(ANNs), decision trees, Bayesian networks, support featuremachines (SFMs), and regression analysis. Cardiovasculardisease prediction model is divided into two categories; oneis the traditional prediction model based on probability. Forexample, in Framingham Heart Study (FHS) [3], the modelis characterized by the adoption of a mathematical formula,which has good stability, but its effect is poor and the ac-curacy is low in the multiclassification and nonlinearcomplex factors. And the other is based on shallow neuralnetwork prediction model of cardiovascular disease. In

Munster Heart Study (PROCAM) [4], two neural networkmodels are used: a multilayer perceptron network with onehidden layer (MLP) [5] model and a probabilistic neuralnetwork (PNN) [6]. )e characteristics of this kind of modelare that it can effectively expand the forecasting factor,quickly process the fuzzy data and nonlinear data, andprovide high rate of accuracy[7]. However, due to therandomness of initialization of shallow neural networkparameters, the prediction results will be much lower thanthe average accuracy, and the variances of multiple pre-diction results will be larger.

Recently, deep learning has been widely used in differentfields and has made great progress [8–11]. )e deep learningmodel uses multiple samples to extract high-level featuresand learns hierarchical representations by combining low-level inputs more effectively. )e learned features charac-terize more intrinsic features of the data, avoiding theprocess of artificial feature design and selection, and have thecharacteristics of many varieties and high accuracy [12].

HindawiJournal of Healthcare EngineeringVolume 2018, Article ID 8954878, 9 pageshttps://doi.org/10.1155/2018/8954878

mailto:lupeng@zzu.edu.cnhttp://orcid.org/0000-0002-1125-9507http://orcid.org/0000-0003-1943-635Xhttp://orcid.org/0000-0002-7629-1866http://orcid.org/0000-0003-1922-2634http://orcid.org/0000-0003-2793-8958http://orcid.org/0000-0002-2429-693Xhttps://doi.org/10.1155/2018/8954878

)is paper takes deep learning as the point of penetrationand uses multilayer network architecture to abstract thecharacteristics of layers and establish a cardiovascular dis-ease prediction model based on deep belief network. At thesame time, the prediction model based on deep trust net-work is improved by using reconstruction error to achievebetter prediction.

Overall, the major contributions of this work can besummarized in three aspects. First, we use the deep beliefnetwork to build predictive models of cardiovascular dis-ease, skip the morphological feature extraction step, andclassify the original ECG data directly, thus solving theproblem that the cardiovascular disease prediction model isnot robust due to the large difference in waveform char-acteristics between patients with the same disease. Second,we adopt the best network parameters trained to initializethe neural network, so as to solve the instability problemcaused by stochastic initialization. Finally, we use re-construction error to improve the prediction model which isbased on the deep trust network, so that it can independentlydetermine the network depth and achieve better predictedresults.

2. Related Work

)e literature related to this classification application wasstudied, and it can be seen that a great variety of methodswere used, which reached high classification accuracies.

Algorithms for R-peak extraction tend to use wavelettransforms to compute features from the original ECGfollowed by a fine-tuned threshold-based classifier. Since theaccurate estimation of heart rate and heart rate variabilitycan be extracted from the R-peak feature, the speciallydesigned algorithm is usually used for the classification ofcoarse-grained heart rhythm. Sundar et al. [13] proposeda prototype using data mining techniques, namely, NäıveBayes and WAC (weighted associative classifier). )e rec-ognition rate of 84% and 78% was obtained from weightedassociative classifier and Näıve Bayes. Iftikhar et al. [14]present a hybrid approach using a supervised learningmodelbased on a well-known classifier SVM and evolutionaryoptimization techniques (genetic algorithm (GA) and par-ticle swarm optimization (PSO)). )e results have shownconsiderably improved accuracy of more than 88%.

However, because of the differences in the ECG wave-forms of different people and the great differences in ECGwaveform characteristics of different diseases, the featureextraction of the waveform is inaccurate. )erefore, thesecharacteristics are not sufficient to distinguish most car-diovascular diseases.

With the rapid development of artificial intelligence,inspired by automatic speech recognition, hidden Markovmodel with Gauss observation probability distribution hasbeen applied to the beat detection task [15], and the hottestartificial neural network is also used for the task of beatdetection [16]. Elsayad proposed an approach which usedthe learning vector quantization (LVQ) neural network toestablish the ECG positive anomaly model and obtained anaccuracy of 74.12% [17]. Olaniyi et al. [18] designed a neural

network for diagnosis of heart diseases with the heart diseasesample obtained fromUCI machine learning repository.)esystem is a multilayer neural network model based onbackpropagation training and is simulated on a feed-forwardneural network. )e recognition of 85% was obtained fromtesting of the network.

Although the self-learning ability of backpropagation(BP) neural network is strong, the convergence speed is slow,and the result is easily affected by the random initializationof network parameters. In particular, there has been nounified and complete theoretical guidance for the selectionof BP neural network structure. Generally, it can only beselected by experience.

)e DBN model not only has the self-adaptive ability ofthe self-adjustment of the general neural network but alsoavoids the defects of the BP neural network, which is easy tofall into the local minimum. DBN uses a network structurecomposed of multiple RBM networks, which is more ef-fective for modeling one-dimensional data [19].

3. Description of the Proposed Approach

3.1. Deep Belief Network. Deep belief network (DBN) is oneof the main tools for deep learning, which is based on therestricted Boltzmann machine (RBM) [20], to propose. )estructure of RBM includes only the visible layer and thehidden layer; the neurons between two layers are fullyconnected, and the neurons in the same layer are notconnected [21].

In Figure 1, v(v1, v2, . . . , vn) represents the visible layer, viis the visible unit; h(h1, h2, . . . , hm) denotes the hidden layer,hj is the hidden unit; and W is the connection weight matrixbetween two layers. )e data are input from the visible layer,(v1, v2, . . . , vn) represents the feature set of the data, and thehidden layer data are generated by the random initializationof the weight value w and the state of each neuron. Due to thedisconnection between neurons at the same level, when de-termining the neuron state, it has the following properties:when the visible cell state is determined, the hidden unitcondition is activated independently; otherwise, if the state ofthe hidden cell is determined, the conditions of the visibleunits are activated independently.

v1

v2

v3

vnVisiblelayer

Hiddenlayer

hm

h2

h1

W

Figure 1: Restricted Boltzmann machine undirected configurationdiagram.

2 Journal of Healthcare Engineering

Given a set of states (v, h), the energy function of theRBM model can be defined by the following equation:

E(v, h) � − n

i�1aivi −

m

j�1bjhj −

n

i�1m

j�1viwijhj,

� −aTv− bTh− hTwv,(1)

where a � (a1, a2, . . . , an) denotes the offset vector of thevisible unit, b � (b1, b2, . . . , bm) denotes the bias vector ofthe hidden unit, and v � (v1, v2, . . . , vn) denotes the statevector of the visible layer, h � (h1, h2, . . . , hm) denotes thestate vector of the hidden layer, w � (wi,j) denotes the con-nection weight matrix, and wi,j denotes the weight of the ithvisible unit and the jth hidden element.

For the state (v, h), according to (1), the joint probabilitydistribution can be given as follows:

P(v, h; θ) �1z

e−E(v,h)

,

z � v

h

e−E(v,h)

,

(2)

where θ � a, b, w{ } is the RMB network parameters and Z iscalled the normalization factor or the partition function.

In practical applications, the probability distributionp(v) of training data v is generally used, that is, the edgeprobability distribution of P � (v, h, θ):

P(v) �1Z

h

e−E(v,h)

. (3)

Similarly, the edge probability distribution P � (h) of thehidden layer state can be obtained:

P(h) �1Z

v

e−E(v,h)

. (4)

RBM training data are obtained by solving the modeloptimal parameters in (3), so that the model can better fit thedistribution of training data even if the sample reaches themaximum probability in the distribution. Constructing log-likelihood functions:

lnP(v) � ln h

e−E(v,h)⎛⎝ ⎞⎠− ln

v

h

e−E(v,h)⎛⎝ ⎞⎠. (5)

)e model parameters are respectively solved by themaximum likelihood function method:

z lnP(v)za

� EPd[v]−EPm[v],

z lnP(v)zb

� EPd[h]−EPm[h],

z lnP(v)zw

� EPd vhT

−EPm vhT

,

(6)

where EPd denotes the expectation of the input conditionalprobability distribution of training data, and EPm denotesthe expectation of the joint probability distribution of the

model. )e expected computation is done by the Gibbssampling method, while the computation cost is too large inthe computation process of each iteration. Hinton proposedthe contrastive divergence (CD) algorithm [21] for theapproximate calculation after sampling.

According to the above formula, when the neuron stateof the given layer is given, it can be inferred that the acti-vation probability of hidden units is

P hj � 1 ∣ v � σ i

viwij + bj⎛⎝ ⎞⎠. (7)

After obtaining the hidden element state matrix, thereconfigurable visible element state probability can be cal-culated according to the CD algorithm:

P vi � 1 | h( � σ j

hjwij + ai⎛⎝ ⎞⎠, (8)

where σ is a sigmoid function σ(x) � 11/(1 + exp(−x)).)e maximum value of the likelihood function is

gradually approximated by gradient ascent. )e formula ofthe RBM parameter is updated as follows:

θi+1 � θi + ηz lnP(v)

zθ, (9)

where η is the parameter learning rate for the model, and i isthe current iteration. )e parameters θ are iterativelyupdated according to the rules of (9), and the maximumvalue of the gradient of the likelihood function is reachedquickly, which is the optimal parameter.

DBN is composed of a plurality of RBM units connectedto the bottom layer of the RBM visible layer as the inputlayer, the underlying RBM hidden layer of the upper RBMvisible layer. )e tuning of global training parameters iscarried out by the BP neural network.

RBM is a probabilistic neural network that determines theprobability generation of DBNs, this is establishing a jointprobability distribution between the feature and the lables:

P v, h1, h2, . . . , hl(

� P v ∣ h1( P h1 ∣ h2( , . . . , P hl−2 ∣ hl−1( P hl−1 ∣ hl( ,(10)

where P(hk ∣ hk+1) is the conditional probability distributionof hk for the given hk+1 state; P(hl−1, hl) is the joint prob-ability distribution of hl−1 and hl. P(v, h) is the jointprobability distribution of a single RBM.)e hidden layer oflow-level RBM in DBN is the visual layer of high-level RBM.So (10) is the probability distribution for the whole model.

)e use of DBN to establish a deep learning-basedcardiovascular disease prediction model is an importantentry point to solve the problem of accuracy and stability ofprediction models.

3.2. Phase 1: ForecastingModel Based onDeep Belief Network.)e use of deep trust network to establish a cardiovasculardisease prediction model is divided into two stages, as shownin Figure 2, respectively, upward training and downwardadjustment.

Journal of Healthcare Engineering 3

(1) Training section: use the greedy layer-by-layertraining algorithm to learn the parameters of eachlayer of RBM θ � a, b, w{ } in turn by unsupervisedlearning. First, the training data are received by thevisible layer of the first layer RBM, and the state v1 isgenerated. )e hidden state h1 is generated upwardsby the initialized weight matrix w1, and the visiblelayer state v1′ is reconstructed by h1. Generating newhidden units, the new layer is generated by w1remapping to the hidden unit h1′. )e parameters areupdated using the CD algorithm until the re-construction error is least, that is, to complete thefirst layer RBM training. Stacked RBMs are trainedlayer by layer according to greedy learning rules,each layer maps different feature spaces )e topmostRBM bidirectional connections make up the asso-ciative memory layer, which can be associated withthe optimal parameters of memory layers. By un-supervised learning, the DBNs gains a priorknowledge, obtains more abstract features at the toplevel, and better reflects the real structure in-formation of the training data. Stacked RBM pre-training input is as follows: training data x, DBN;and output is as follows: unsupervised DBN.

(2) Tuning section: taking the pre-trained parametersof the network as initial values, the labeled samplesare used to supervise the DBN model and the top-down reverse propagation error of the network isused as the standard to further optimize the RBMparameters of various layers. )e initial value of BPnetwork is the high abstract feature set obtained bythe pretraining of DBN, which solves the problemof falling into local optimum and overfitting causedby random initialization of the traditional neural

network.)e parameters are finetuned based on theBP algorithm, and the input is the parameters ofeach layer of the DBN pretraining and the outputvector of the top RBM; the output is the DBN afterfinetuning the parameters.

)rough the above steps, a globally optimal DBN modelis constructed and fully trained. To sum up the abovelearning phase, a complete DBN model is established, andthe input is as follows: number of DBN structure layers,training samples; output is as follows: fully trained DBN.

Cardiovascular disease training samples without labelvalues were entered into the visible layer of the bottom RBMwithout any characteristics of supervised learning data. )etop RBM will learn the optimal characteristic parameters asthe initial value of the neural network solves the defectscaused by random initialization and improves the stability ofthe model prediction.

3.3. Phase 2: Improved Deep Belief Network ForecastingModel. )e more complex the network structure of DBN,the stronger the ability to solve complex problems. Simul-taneously, the higher the number of network layers, theharder the training will be, the greater the training erroraccumulates, and the lower the correctness of the model[22]. In application, in order to establish suitable DBNstructure for specific tasks, due to lack of correspondingtheoretical support and effective training mode, the depth ofnetwork and the number of hidden units need to be set byexperience, which leads to the deviation in the modelingprocess and the high cost [23].

Aiming at the problem of determining the number oflayers of DBN, based on the reconstruction error of eachRBM training, this paper improves the prediction model ofdeep trust network and establishes a DBN which can au-tomatically select the network depth to improve the auto-matic analysis ability of the cardiovascular disease predictionmodel. Specific methods are as follows.

In each RBM, the input data of the visible layer arereconstructed andmapped to the hidden layer again, and thereconstruction error is calculated based on the differencebetween the reconstructed output data and the initialtraining data.

Rerror �

ni�1

mj�1 pij − xij

nmpx, (11)

where Rerror denotes the reconstruction error, n denotes thenumber of training samples, m denotes the number offeatures in each group of samples, Pij denotes the recon-structed value of RBM training sample per layer, xij denotesthe true value of the training sample, and Px denotes thecalculation of the number of values.

In order to prevent the training data from overfitting orreconstructing large deviation of the data and at the sametime to balance the training cost of the network model,when the difference between the two reconstruction errorsis less than the present value, the depth accumulation isstopped.

Output theresult

W3

W2

W1

Entertraining data

BP

RBM3

RBM2

RBM1

Labeled data

Back propagation

Fine-tuning

Fine-tuning

h2

h1

v3

v2

v1

Figure 2: DBNs training flow chart.

4 Journal of Healthcare Engineering

L � NRBM + 1,

Rerror(k− 1)−Rerror(k)

> ε,

L � NRBM,

Rerror(k− 1)−Rerror(k)

< ε,

(12)

where L denotes the hidden layer number of DBN, Rerror (k)denotes Rerror of current layer, and ε denotes the defaultvalue. )e selection of the preset value is one of the keys todetermining the accuracy of the model. )e value of thedefault value ε is too large, which can cause inaccuratelyfinding the optimal number of network layers. If the value istoo small, the number of layers in the deep neural networkmay be too large and the calculation amount is too large. Forthe number of cardiovascular disease prediction modelparameters and the performance of laboratory equipment,we determined that ε ∈ [0.01, 0.05]. Compared with manyexperimental results, when ε � 0.03, the prediction modelcan determine the network depth independently.

In the pretraining phase of the unsupervised, when itreaches the number of layers of target value, the top-leveltrained output is used as input of the BP algorithm and thereverse fine-tuning parameters are started. )e process ofbuilding a network relies on Rerror, as shown in Figure 3.

Rerror is positively related to the network energy E(v, h),and this coupling characteristic also proves feasibility ofDBN depth with the reconfiguration error as the standard. Itis proved as follows.

Let P be the calculated value, and X be the actual labelvalue, then P � P(v) and X � P(v1); according to theconditional probability formula, there is

P � P(v) � P v1( P h ∣ v1( P(v ∣ h). (13)

According to the total probability formula, there is

P(v ∣ h) �P(v, h)

P(h). (14)

According to (14), to rewrite (13), there is

P � P v1( P v1, h(

P v1(

P(v, h)

P(h),

� P v1, h( P(v, h)

P(h).

(15)

According to (14) again, there is

P � P v1 ∣ h( P(h)P(v, h)

P(h),

� P v1 ∣ h( P(v, h).

(16)

Substituting the above formula in (11) to reconstruct theerror:

Rerror �

ni�1

mj�1 pij −xij nmpx

,

� P−X,

� P v1 ∣ h( P(v, h)−P v1( ,

� P v1( [P(v, h)− 1].

(17)

As the energy of the neural network is proportional to theprobability distribution, that is, P(v, h)∞E(v, h), there is

Rerror∞P(v, h)∞E(v, h). (18)

Equation (18) shows that there is a coupling relationshipbetween Rerror and network mechanism, and it is reasonableto rely on reconstruction error to determine the networkdepth of DBN autonomously. )e number of neurons ineach layer also has an impact on the network. At present,there is a lack of a clear theory to prove that the appropriatenumber of cells is set and the improvement is achieved. )eDBN structure focuses on the ability to determine the depth ofa network, and the number of neurons in each layer is fixed.

4. Experiment Analysis

4.1. Database Description. Experimental data select theStatlog (Heart) data set and the Heart Disease Database dataset for the UCI Machine Learning Library. )e Statlog(Heart) data set contains 270 sets of instances and the HeartDisease Database data set contains 820 sets of instances. )eproperties of both data sets contain continuous, two-category, ordered multiclass, and unordered multiclassvariables. As shown in Table 1, select the same 13 attributesand 1 classified label values in two data for experiments.

)e physical meaning, data unit, and order of magnitudeof each attribute in the selected data set are different and needto be normalized before the experiment. Text-based data are

Begin

Enter training data

Use CD algorithmfor training

Rerror satisfies(13)

BP algorithm finetuningnetwork parameters

End

Parameterinitialization

Figure 3: Calculation flowchart of DBNs depth.

Journal of Healthcare Engineering 5

directly converted to numeric data.)e reference standard formedicine is the data attribute of the hierarchical classificationstructure. )e normalized assignment is the correspondingdiscrete arithmetic progression or geometric progression. Forthe data attributes of the range type, we proposed improvedmin-max normalization due to the existence of data imbal-ances: take the average of the first k large values of the featureterm as the maximum value, and take the average of the first ksmall values as the minimum value. )e feature item isnormalized to the interval (0, 1) as min-max.

In the two data sets, 70% of the instances are selected astraining samples, and the remaining instances are testsamples. )e data set is divided into two mutually exclusivecollections, and the consistency of data distribution ismaintained as much as possible.

4.2. Improved DBN Model Network DepthAnalysis Experiment

4.2.1. Improved DBNs Model Experiment. Using trainingdata, improved DBN models are built and tested with test

data. Inputs are as follows: training sample risk factor data x{ },training sample tag value y , and testing samples x′, y′ ; theoutput is the forecast results. )e steps are as follows:

(1) Set the initial value of the network, the learning rateis set to 1, the initial error is 0, the setting error of thereconstruction error ε is set to 0.03, the maximumtraining period of each RBM is set to 10 times. )eweight (w), the visible layer bias (a), and the hiddenlayer offset (b) are all randomly generated values thatare smaller, and the training batch is set to 100.

(2) )e training data x{ } with the label value removed isinput as the first layer network and the unsupervisedpretraining phase is started. )e number of neuronsin the input layer automatically takes the value of thesample feature dimension, that is, 13 risk factors inthe data set. Perform the following steps using Gibbssampling and CD algorithms, as shown in Table 2.

Update the parameters and calculate the error and repeatthe above steps until the end conditions are met. In this case,the first layer of RBM is trained, and the principle of

Table 1: Data set attributes.

Features Description Data types Normalization Value

Age Age Continuousdata Min-max scaling 16–80; 0∼1

Sex Gender Text-based data Direct mapping 0: female1: male

Cp Chest pain type Text-based data Direct mapping

0: typical angina1: typical type angina2: nonangina pain3: asymptomatic

Trestbps Trest blood pressure Range data Improved min-maxscalingMmHg on admission to the

hospital

Chol Serum cholesterol Range data Improved min-maxscaling (mg/dl)

Fbs Fasting blood sugar Hierarchicaldata Hierarchical mapping0: 120mg/dl

Restecg Resting electrographic results Text-based data Direct mapping

0: normal1: having ST-T wave abnormality2: showing probable or definite left

ventricular hypertrophy

)alach Maximum heart rate achieved Range data Improved min-maxscaling —

Exang Exercise-induced angina Text-based data Direct mapping 0� no1� yes

Oldpeak ST depression induced byexercise relative to rest Range dataImproved min-max

scaling

Slope Slope of the peak exercise ST segment Text-based data Direct mapping0: unsloping

1: flat2: downsloping

Ca Number of major vessels colored byfluoroscopy Text-based data Direct mapping 0–3

)al Text-based data Direct mapping0: normal

1: fixed defect2: reversible defect

Num Predicted attribute — — 0, 1, 2, 3, 4

6 Journal of Healthcare Engineering

reconfiguring the error method to determine the depth ofthe network is used to calculate whether the condition is met;if it is satisfied, it stops; if it is not, h1′ is used as the input forthe next layer of training.

(3) Use step (2) to determine the final depth of thenetwork, and remember the optimal parameters ofeach layer. )e trained DBN structures and theparameters are passed to the BP network to build thesame depth of backpropagation network.

(4) )e top RBM output for the BP network input, whileinputting the training data label value, began tomonitor the tuning phase and further adjust theparameters of the DBN layers.

(5) Put the unlabeled test data into the constructedimproved DBN, and compare the value of the labelvalue of the network to the true label value to cal-culate the prediction accuracy.

(6) )e algorithm ends.

4.2.2. Standard DBNs Model Experiment. In order to im-prove the correctness of the network depth determined byDBN autonomously, a standard DBN is established and theoptimal network layer number is determined by experiment.)e optimal number of cells in each layer is experimentallyselected according to

N � ⌈���mn

√⌉ + k, (19)

where m denotes the dimension of the input data, that is,the number of CVD risk factors; n denotes the number ofoutput layer units and CVD predicts the probability as theoutput, that is, n � 1; N is the number of hidden units; ⌈ ⌉ isthe uplift symbol; and k is an integer between [1, 5], which isused to increase the interval of units selection and avoidblind selection.

4.2.3. Experimental Results and Analysis. )e improvedDBN prediction model was tested in two data sets andstopped increasing when Statlog (Heart) was added to thethird layer, with a depth of 4; the Heart Disease Databasestopped increasing when it increased to the fourth level witha model depth of 5. )e Rerror curve in the RBM computingprocess of each layer is shown in Figures 4 and 5.

In order to improve the performance of DBN, a standardDBN model with the same structure was established, that is,a 4-layer neural network was established for Statlog (Heart)and a 5-layer neural network was established for the HeartDisease Database. )e number of network units per layer wasbased on (19), and the best number of units is selected by the

experimental method. )e number of input layer units isequal to 13 feature latitudes of the data set, that is, m � 13; thenetwork output is a label probability obtained by regressioncalculation, that is, n � 1; and the number of second layerunits ranges from 5 to 9 experiments to select the smallestreconstruction error as the optimal unit number, the numberof units under the reconstruction error shown in Figure 6.

Table 2: Algorithm implementation steps.

h1 � sigm(vT1 w1 + b)Use the input data to construct

a hidden unit state

v1′ � sigm(hT1 w1 + a)Reconstruct the input usingthe hidden layer structure

h1′ � sigm(v′T1 w1 + b)

Construct the hidden layer withthe reconstructed input again

2 4 6 8 10 12 140

0.51

1.52

2.5

R error

33.54

Numepochs

RBM1RBM2RBM3

Figure 4: )e Rerror curve of each layer RBM on Statlog (Heart).

R error

5

4

3

2

1

02 4 6 8 10 12 14

Numepochs

RBM1RBM2

RBM3RBM4

Figure 5: )e Rerror curve of each layer RBM on Heart DiseaseDatabase.

Statlog (heart)Heart Disease Database

5

6

7

8

9

13

2.092.34

1.742.11

1.61.94

1.661.82

1.721.73

1.871.95

Figure 6: Reconstruction error of RBM1 with different number ofhidden units.

Journal of Healthcare Engineering 7

As shown in Figure 6, the Rerror of RBM1 in the Statlog(Heart) data set is the smallest at the 7th implicit unit, andthe number of units is determined to be 7.)e Heart DiseaseDatabase has the smallest Rerror at the 9th implicit unit, andthe number of units is determined to be 9. Similarly, theDBN structure finally determined according to the abovemethod is Statlog (Heart): 4-layer network, the number ofunits of per layer is 13-7-6-4; Heart Disease Database: 5-layernetwork, the number of units of per layer is 13-9-8-5-4.

To further improve the correctness of the network depthdetermined by DBN, we increase the hidden layer number ofthe standard DBN model in Figure 6.

Reconstruction error of RBM1 with different numbers ofhidden units turns and judges the correctness of the testdata. To ensure that the number of layers is the only in-dependent variable, the number of units in each layer is thesame as that of the improved DBN model. )e results areshown in Table 3.

Analysis of Table 3 shows that increasing the networkhierarchy reduces Rerror and training time will increase. )eaccuracy of the test data was maximized for Statlog (Heart)at depth 4, maximum for the Heart Disease Database atdepth 5, and in line with the improved network depth thatDBN automatically determines; it further proves that theprediction model of cardiovascular diseases based on im-proved DBN has better performance.

Table 4 presents the overall results of the proposedStatlog (Heart) data set evaluation using the UCI Machine

Learning Library for the proposed improved DBN pre-diction model and other different hybridization and non-hybrid techniques for cardiac classification andidentification of relevant risk factors.

From the comparison of the tables, we can see that thetraditional feature extraction algorithm is more specific toa specific data set. Based on the experimental accuracy rate,a special manually set feature combination is used. )ismethod is to dig out the characteristics of the data set itself,not the essential characteristics of ECG data; the general-ization ability of the method is weak, the portability is poor,and the accuracy is relatively poor.

)e traditional classification model based on proba-bility uses a combination of multiple feature extractionmethods. However, the deep learning method can learna kind of deep-level nonlinear network structure and caneffectively obtain the deep-level essential feature repre-sentation of ECG from the sample. )e effectiveness of themodel based on deep learning is better than that of thetraditional classification models based on probability andshallow neural networks.

)is paper constructs a deep confidence network whichcan independently determine the network structure. )eperformance of the model is evaluated on two data sets, andthe highest accuracy is achieved. )e algorithm has stronggeneralization ability, and it can fully tap the deep-levelcharacteristics of ECG and achieve an accurate and stableautomatic classification of cardiovascular diseases incomplex individuals and complex environments. )eperformance of heart disease classification is superior toother technologies.

5. Conclusion

For these issues, the probabilistic-based predictive modelcannot integrate multiclass and nonlinear factors, and thestability of shallow neural network is poor. A predictionmodel based on deep learning is proposed and improved toenable it to independently determine the network param-eters. )e proposed prediction model was validated with theStatlog (Heart) data set and the Heart Disease data set, whichproves that the predictionmodel has high accuracy and goodstability.

Our further research is to apply the prediction modelbased on improved depth learning to actual cardiovasculardisease predictions. By analyzing the prediction results indetail, we can quantify the proportion of each risk factor tothe risk of cardiovascular disease and provide personalizedadvice to reduce the risk of cardiovascular disease.

Conflicts of Interest

)e authors declare that they have no conflicts of interest.

Acknowledgments

)e work was supported by the National Natural ScienceFoundation of China, under Contract 60841004, 60971110,61172152, and 61473265; the Program of Scientific and

Table 3: DBNs test results at different network depths.

Data set Networkdepth RerrorAccuracy

(%)Runtime

(s)

Statlog (Heart)

2 1.8700 81.43 17.303 0.5189 85.71 19.904 0.4912 91.26 22.705 0.4306 87.14 25.206 0.3877 82.58 29.50

Heart DiseaseDatabase

2 1.9450 80.40 21.403 1.1931 83.20 22.904 0.6594 85.60 25.205 0.6372 89.78 27.806 0.5765 87.20 31.50

Table 4: Comparison of classification results of different tech-niques on the Statlog (Heart) data set.

Classification algorithms Accuracy (%)SVM linear [14] 86.62SFM [24] 83.31OCSFM [24] 86.73Naı̈ve Bayes [13] 78.93WAC [13] 84.00ANN [14] 86.04PSO-SVM-SMO-RBF [14] 88.24Naı̈ve Bayes [25] 82.31Decision tree [25] 84.35LVQNN [17] 74.12BPNN [18] 85.00Improved DBN 91.26

8 Journal of Healthcare Engineering

Technological Research of Henan Province, China, underContract 172102310393; the Support Program of Science andTechnology Innovation of Henan Province, China, underContract 17IRTSTHN013; the Key Support Project Fund ofHenan Province, China, under Contract 18A520011; theFund for “Integration of Cloud Computing and Big Data,Innovation of Science and Education,” China, under Con-tract 2017A11017; and the CERNET Innovation Project,China, under Contract NGII20161202.

References

[1] W.-W. Chen, R.-L. Gao, and L.-S. Liu, “China cardiovasculardiseases report 2015: a summary,” Journal of Geriatric Car-diology, vol. 1, pp. 1–10, 2017.

[2] N. Cohnj, “Prevention of cardiovascular disease,” Trends inCardiovascular Medicine, vol. 25, no. 5, pp. 436–442, 2015.

[3] M. J. Pencina, R. B. Agostino, M. G. Larson et al., “Predictingthe 30-year risk of cardiovascular disease: the Framinghamheart study,” Circulation, vol. 119, no. 24, pp. 3078–3084,2009.

[4] R. Voss, P. Cullen, H. Schulte et al., “Prediction of risk ofcoronary events in middle-aged men in the prospectivecardiovascular munster study (PROCAM) using neural net-works,” International Journal of Epidemiology, vol. 31, no. 6,pp. 1253–1262, 2002.

[5] A. V. Bahena, M. Mejialavalle, and J. R. Ascencio, “Towardoptimal parameter selection for the multi-layer perceptronartificial neural network,” in Proceedings of the InternationalConference on Mechatronics, Electronics and Automotive En-gineering, pp. 103–108, Cuernavaca, Mexico, November 2013.

[6] M. E. Pastell and M. Kujala, “A probabilistic neural networkmodel for lameness detection,” Journal of Dairy Science,vol. 90, no. 5, p. 2283, 2007.

[7] R. Deo, “Machine learning in medicine,” Circulation, vol. 132,no. 20, pp. 1920–1930, 2015.

[8] A. Khumoyun, Y. Cui, and L. Hanku, “Spark based distributeddeep learning framework for big data applications,” in Pro-ceedings of the International Conference on Information Sci-ence and Communications Technologies, pp. 1–5, Tashkent,Uzbekistan, November 2016.

[9] M. A. Alsheikh, D. Niyato, S. Lin et al., “Mobile big dataanalytics using deep learning and apache spark,” IEEE Net-work, vol. 30, no. 3, pp. 22–29, 2016.

[10] B. M. Wilamowski, B. Wu, and J. Korniak, “Big data and deeplearning,” in Proceedings of the Jubilee International Confer-ence on Intelligent Engineering Systems, pp. 11–16, Budapest,Hungary, June 2016.

[11] Q. Zhang, M. Lin, L. T. Yang et al., “Energy-efficientscheduling for real-time systems based on deep Q-learningmodel,” IEEE Transactions on Sustainable Computing, vol. 99,p. 1, 2017.

[12] Q. Zhang, L. T. Yang, Z. Chen et al., “A survey on deeplearning for big data,” Information Fusion, vol. 42, pp. 146–157, 2018.

[13] N. A. Sundar, P. P. Latha, and M. R. Chandra, “Performanceanalysis of classification data mining techniques over heartdisease database,” International Journal of Engineering Scienceand Advanced Technology, vol. 2, no. 3, 2012.

[14] S. Iftikhar, K. Fatima, A. Rehman, A. S. Almazyad, andT. Saba, “An evolution based hybrid approach for heartdisease classification and associated risk factor identification,”Biomedical Research, vol. 28, no. 8, pp. 3451–3455, 2017.

[15] D. A. Coast, R. M. Stern, G. G. Cano et al., “An approach tocardiac arrhythmia analysis using hidden Markov models,”IEEE Transactions on Biomedical Engineering, vol. 37, no. 9,pp. 826–836, 1990.

[16] S. L. Melo, L. P. Caloba, and J. Nadal, “Arrhythmia analysisusing artificial neural network and decimated electrocardiographic data,” Computers in Cardiology, vol. 27, pp. 73–76,2000.

[17] A. M. Elsayad, “Classification of ECG arrhythmia usinglearning vector quantization neural networks,” in Proceedingsof the International Conference on Computer Engineering &Systems, pp. 139–144, Cairo, Egypt, 2009.

[18] E. O. Olaniyi, O. K. Oyedotun, A. Helwan et al., “Neuralnetwork diagnosis of heart disease,” in Proceedings of theInternational Conference on Advances in Biomedical Engi-neering, pp. 21–24, Beirut, Lebanon, September 2015.

[19] A. Mohamed, G. E. Dahl, G. Hinton et al., “Acoustic modelingusing deep belief networks,” IEEE Transactions on AudioSpeech and Language Processing, vol. 20, no. 1, pp. 14–22,2011.

[20] G. Hinton, S. Osindero, and Y. Teh, “A fast learning algorithmfor deep belief nets,” Neural Computation, vol. 18, no. 7,pp. 1527–1554, 2006.

[21] F. Asja and I. Christian, “Training restricted Boltzmannmachines: an introduction,” Pattern Recognition, vol. 47, no. 1,pp. 25–39, 2014.

[22] J. Schmidhuber, “Deep learning in neural networks: anoverview,” Neural Networks, vol. 61, pp. 85–117, 2015.

[23] P. Guangyuan, C. Wei, and Q. Junfei, “Depth determinationmethod of DBN network,” Journal of Control and Decision,vol. 30, no. 2, pp. 256–260, 2015.

[24] A. Adeli and M. Neshat, “A fuzzy expert system for heartdisease diagnosis,” in Lecture Notes in Engineering andComputer Science, Springer, Berlin, Germany, 2010.

[25] V. Chaurasia and S. Pal, “Data mining approach to detectheart disease,” International Journal of Advanced ComputerScience and Information, vol. 2, no. 4, pp. 56–66, 2013.

Journal of Healthcare Engineering 9

International Journal of

AerospaceEngineeringHindawiwww.hindawi.com Volume 2018

RoboticsJournal of

Hindawiwww.hindawi.com Volume 2018

Hindawiwww.hindawi.com Volume 2018

Active and Passive Electronic Components

VLSI Design

Hindawiwww.hindawi.com Volume 2018

Hindawiwww.hindawi.com Volume 2018

Shock and Vibration

Hindawiwww.hindawi.com Volume 2018

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawiwww.hindawi.com Volume 2018

Hindawiwww.hindawi.com Volume 2018

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawiwww.hindawi.com

Volume 2018

Hindawi Publishing Corporation http://www.hindawi.com Volume 2013Hindawiwww.hindawi.com

The Scientific World Journal

Volume 2018

Control Scienceand Engineering

Journal of

Hindawiwww.hindawi.com Volume 2018

Hindawiwww.hindawi.com

Journal ofEngineeringVolume 2018

SensorsJournal of

Hindawiwww.hindawi.com Volume 2018

International Journal of

RotatingMachinery

Hindawiwww.hindawi.com Volume 2018

Modelling &Simulationin EngineeringHindawiwww.hindawi.com Volume 2018

Hindawiwww.hindawi.com Volume 2018

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawiwww.hindawi.com Volume 2018

Hindawiwww.hindawi.com Volume 2018

Navigation and Observation

International Journal of

Hindawi

www.hindawi.com Volume 2018

Advances in

Multimedia

Submit your manuscripts atwww.hindawi.com

https://www.hindawi.com/journals/ijae/https://www.hindawi.com/journals/jr/https://www.hindawi.com/journals/apec/https://www.hindawi.com/journals/vlsi/https://www.hindawi.com/journals/sv/https://www.hindawi.com/journals/ace/https://www.hindawi.com/journals/aav/https://www.hindawi.com/journals/jece/https://www.hindawi.com/journals/aoe/https://www.hindawi.com/journals/tswj/https://www.hindawi.com/journals/jcse/https://www.hindawi.com/journals/je/https://www.hindawi.com/journals/js/https://www.hindawi.com/journals/ijrm/https://www.hindawi.com/journals/mse/https://www.hindawi.com/journals/ijce/https://www.hindawi.com/journals/ijap/https://www.hindawi.com/journals/ijno/https://www.hindawi.com/journals/am/https://www.hindawi.com/https://www.hindawi.com/