Preventive Healthcare: A Neural Network Analysis Of .

Healthcare 2017, 5, 83 of 13and duration of smoking has been associated with an increased risk of chronic obstructive pulmonarydisease [18]. In the U.S. in 2015, approximately 15% of all adults (36.5 million) were cigarette smokers,and more than 13 million live with a smoking-related disease [3]. Additionally, people diagnosed withsmoking-related chronic diseases were found to be current smokers. There is a need for evidence-basedapproaches that prevent smoking initiation or increase smoking cessation in the U.S. The behavioralrisk factors of smoking, alcohol consumption, improper diet, and lack of physical activity contributeto about half of the burden of diseases in developed countries [7,19]. These factors are not equallydistributed through the population but tend to concentrate and affect the most vulnerable segments [20].The exposure to behavioral risk factors is temporal and varies with demographic characteristics suchas age and income, among others [21]. Also, behavioral habits occur on a long-term basis and canhave an impact on the health of individuals [21]. Epidemiological studies emphasize the importanceof measuring the impact of multiple lifestyle risk behaviors on people’s health [21].Promoting good behavioral habits can positively influence the prevention or delay of disability,dementia, frailty, and non-communicable/chronic conditions [21]. Modifying behavioral habitsconsists of disrupting the cue-response association, the fundamental principle for habit formation [22].Avoiding exposure to everyday cues can help facilitate behavior change. Our research on behavioralhabits and the association with chronic conditions is based on the underlying premise that behavioralhabits, if identified, can be addressed and modified.2.2. Neural Networks in HealthcareHealthcare is a domain that has deployed health analytics for various areas, including preventivehealth and wellness and disease management [23]. In disease management, by identifying the affectedpopulations in different disease categories, analytics helps target customized management techniquesand practices that will mitigate the disease as well as prevent the onset of associated medical conditions.Because of their ability to perform input-output mapping of data without a priori knowledge ofdistribution patterns, neural networks are appropriate for applications that deal with large volumes ofdata and with fuzzy or noisy data. These networks have the ability to learn from experience, generalizefrom previous examples, and abstract relevant features from irrelevant data inputs [24]. Neural networkapplications in the domain of chronic disease management include automatic prediction of exacerbationsin Chronic Obstructive Pulmonary Disorder [25]; diagnosing myocardial infarction [26–29], coronaryartery disease [30–32], chronic heart failure [33]; predicting heart diseases [34]; classifying other typesof heart disease [35]; diagnosing diabetes on small mobile devices [36]; and identifying behavioralhealth problems of patients who are at high risk for hospital admission [37]. In most chronic diseases,early detection is beneficial for effective management of the conditions.2.3. The Neural Network ModelA neural network consists of a series of processing elements called neurons that are interlinked toform a network. Each link has a weight associated with it. Each neuron receives stimuli (information)from the surrounding neurons that are linked to it, processes the information, and produces anoutput [38]. A neural net consists of an input layer, one or more hidden layers, and the output layer.The neurons in the input layer receive stimulus from outside the network; the neurons in the hiddenlayer receive stimulus from the interconnected neurons and pass on the output to other neurons withinthe network; and the neurons in the output layer receive the stimulus from the linked neurons andpass on the output externally. Different neural network structures arise based on combinations ofneurons and layers [39].In this research, a Multilayer Perceptron (MLP) feed forward neural network was used andtrained with the error back propagation algorithm. The MLP consists of an input layer, one or morehidden layers, and an output layer. Information moves in a forward direction through the network.The number of neurons at the input layer is guided by the number of independent variables, while thenumber of neurons at the output layer correlates with the number of values that need to be predicted.

Healthcare 2017, 5, 84 of 13Unlike the input and output layers, there are no widely accepted rules for determining the optimalnumber of hidden layers. A less than optimal number of hidden units will result in hampering thenetwork’s learning of the input-output mapping. A more than optimal number of hidden units willresult in the network generalizing poorly on new data. The optimal configuration is most often derivedby trial and error approach [24].The network is initially fed an array of input-output values. It is then trained using the backpropagation algorithm to assign appropriate weights for the connections and calculate the outputs.The accuracy of the predicted outputs is then estimated by comparing with known values. Error signalsare created out of such comparisons and are propagated backwards through the various layers.The network then adjusts and updates the weights appropriately. These training iterations are repeateduntil the network learns to adjust the weights and arrives at predictions that show a minimal differencewith the actual values.3. Research Methodology3.1. Data CollectionData for 475,687 records were collected from the CDC’s Behavioral Risk Factor SurveillanceSystem (BRFSS) database for the year 2012. The indicators for behavioral habits include alcoholconsumption, regular soda consumption (sugar), frequency of smoking, frequency of drinking alcohol,weekly working hours, fruit consumption, vegetable consumption, and exercise. The indicators forchronic diseases include heart attack, stroke, asthma, and diabetes. The demographic variables ofmarital status, income level, and age are included. The data for the variables was extracted at a statelevel for the state of New York. The variables and their description are shown in Table 1.For the neural network analysis, the independent variables were the behavioral habits of alcoholconsumption, regular soda consumption (sugar), frequency of smoking, frequency of drinking alcohol,weekly working hours, fruits consumption, vegetables consumption, and exercise. The dependentvariables were heart attack, stroke, asthma, and diabetes.We analyzed the data for the following proposition. We have included demographic variables inthe analysis. Even though the demographic variables are not modifiable, they play a major role in theonset of chronic conditions. Also, analyzing demographics in relation to chronic diseases can facilitatetargeting and planning of future intervention and wellness programs.Chronic diseases have a positive association with alcohol consumption, soda consumption,weekly working hours, marital status, income level, and age; and a negative association with fruit andvegetable consumption, and exercise.3.2. Analytics Tool SelectionSPSS Modeler was utilized with its functions of Neural Networks, Association, and Bayesiannetworks. The model building stage consisted of experimenting with one and two hidden layerswith various combinations of nodes to determine the best model. The training-testing percentages of50–50, 60–40, and 70–30 were used. Neural Network builds the model by learning from the potentialcorrelation between independent (behavioral habits) and dependent (chronic diseases) variables.It then validates the model results by comparing the predicted values with the actual values. In suchapplications, neural network systems are better than conventional computers that follow a set ofinstructions to solve a problem.

Healthcare 2017, 5, 85 of 13Table 1. Variables in the research.VariableDescription of VariablesBehavioral Habits:Smoking historySmoked at least 100 cigarettes in the entire life or notFrequency of drinking alcoholNumber of days of having at least one alcoholic drink per week or permonth during the past 30 daysFrequency of drinking regular soda during the last 30 days:Frequency of drinking soda (sugar)1 - Times per day (00–99)2 - Times per week (00–99)3 - Times per month (00–99)Times per day, week, or month eating fruit (not counting juice):Frequency of eating fruits1 - Times per day (00–99)2 - Times per week (00–99)3 - Times per month (00–99)Times per day, week, or month eating vegetables (include tomatoes,tomato juice or V-8 juice, corn, eggplant, peas, lettuce, cabbage, andwhite potatoes that are not fried such as baked or mashed potatoes):Frequency of eating vegetables1 - Times per day (00–99)2 - Times per week (00–99)3 - Times per month (00–99)ExerciseParticipated in any physical activity or exercise, other than a regular job,such as running, calisthenics, golf, gardening, or walkingChronic Diseases:Heart attackIf the person had a heart attackStrokeIf the person had a strokeAsthmaIf the person had an asthma attackDiabetesIf the person had diabetesWeekly working hoursHours working per week at all jobs and businesses combinedDemographics:Marital statusMarried, Divorced, Widowed, Separated, Never married, a member ofan unmarried couple (1–6)Income levelAnnual household income levelAgeAge of the person4. Analysis and ResultsSPSS Modeler and Auto Classifier Model are used to analyze the dataset. The analyses for modelbuilding, training, and testing phases are described below.4.1. Neural Network Training and TestingNeural Network with Auto Classifier model was selected as the one that works best with noisyand fuzzy data. Independent variables were selected in accordance with the weights assigned bythe model. We adopted different combinations of hidden layers (one and two) and nodes, andexperimented with different partition rates of the data set for training and testing: 50–50, 60–40, and70–30 (training-testing %). Since we have a comparatively large dataset, we had the option of adoptingthe most strict partition rate for the neural net. The logic is that if the model functions well under suchstrict conditions, it would illustrate that the association is explicit and solid.The iterations of 50%, 40% and 30% of the data set to test the training results for prediction wereadopted to represent strict, moderate, and loose conditions, respectively. The Auto Classifier model

Healthcare 2017, 5, 86 of 13was used to explore possible classification models other than Neural Network for similar predictionsusing different approaches. The aggregate results are compared to determine the best approach.We set the chronic disease of stroke and heart attack as the target or dependent variables, andall other2017,behavioralhabits variables as the predictor/independent (input) variables. Neural networkHealthcare5, 86 of 13models were run separately for each dependent variable. The six most important predictors for o runthe modelsThehaddatahadrows2907forof the dependentselectedto runthe modelsagain.again.The data2907forrowsanalysis.analysis.Figure the1 showsthe bestfor predictingstroke,with thehighestofaccuracyof 97.6%.Figure 1 showsbest modelfor modelpredictingstroke, withthe highestaccuracy97.6%. Thebest fitThebestfitonemodelhaslayer,one inputlayer, layerone hiddenlayerwithandfouronenodes,andlayer.one outputlayer. Themodelhasinputone hiddenwith fournodes,outputThe partitionratepartitionrate used.of 50–50 was used.of 50–50 wasThetop threethreepredictorspredictorsstrokeunderthis modelareage, workingweekly workinghours,andThe topforforstrokeunderthis modelare age,weeklyhours, andfrequencyfrequencyofsodadrinkingsoda(sugar).is thenumberoneofpredictorof stroke:theperson,older theof drinking(sugar).Ageis the Agenumberonepredictorstroke: theolder thetheperson,higherthethe ofpossibilitya stroke.theSimilarly,theweeklyhigherworkingthe weeklyworkingandhigherthe higherpossibilitya stroke. ofSimilarly,higher thehoursand thehourshigherthethefrequencythefrequencyof drinkingsoda,higher ofthehavingpossibilityof havinga stroke.The otherpredictorsforof drinkingsoda,the higherthe thepossibilitya stroke.The otherpredictorsfor stroke,in order,stroke,in order, areofconsumptionof vegetablesand fruits, consumptionof alcohol,income,offrequencyare consumptionvegetables andfruits, consumptionof alcohol, hasexercise.It has tobeinnotedthatset,in thedataset,a therewasina disparityinandfrequencyexercise. Itto be notedthatthe datatherewasdisparitythe (50records)whencomparedtothosewhoof people who were diagnosed with a stroke (50 records) when compared to those who were neverwerenever withdiagnosedwitha stroke(2857 Givenrecords).Giventhis situation,thewouldmodelhavewouldhavebeendiagnoseda stroke(2857records).thissituation,the modelbeenabletoableto preciselyonlyonetwoof thetwo groups,thethatgroupwasnever diagnosedwithpreciselypredictpredictonly oneof thegroups,namely namelythe groupwasthatneverdiagnosedwith a stroke.aTostroke.To solvethis problem,the ofdatasize ofpeoplenever diagnosedwitha strokewassolve thisproblem,the data sizepeoplewhowere whoneverwerediagnosedwith a strokewasreducedtoreducedto 500.Usingtraining/testingdifferent 0 wasasselected500. Usingdifferentpercentages,the model thewithdata sizewasselectedthe oneasthetheonehighestwith thehighest (89.1%).accuracyThis(89.1%).Thismodelis alsoinshownwithaccuracymodelis alsoshownFigurein1.Figure 1.Figure 1.1. TheThebestbestNeuralNeuralNetworkNetwork modelsmodels withwith datadata sizessizes 29072907 andand 550.550.FigureTheThe toptop threethree predictorspredictors forfor strokestroke withwith thisthis modelmodel areare workingworking hours,hours, maritalmarital status,status, fpeoplewhowerediagnosedwithaconsumption of fruits. The model classified 42.3% of people who were diagnosed with a strokestroke andand100%100% ofof peoplepeople whowho werewere nevernever diagnoseddiagnosed withwith aa stroke.stroke. ForFor thethe trainingtraining data,data, thethe modelmodel predictedpredicted40.741%diagnosedwitha40.741% ofouranalysesa stroke. The prediction accuracy in training was higher than in testing. A summary of our analysesusingwith thethe bestbest modelsmodels highlighted.highlighted.using neuralneural networknetwork isis shownshown inin TableTable 22 The top three predictors for stroke with this model are workingworking hours,hours, maritalmarital status,status, fpeoplewhowerediagnosedwithaconsumption of fruits. The model classified 42.3% of people who were diagnosed with a strokestroke .Forthetrainingdata,themodelpredicted100% of people who were never diagnosed with a stroke. For the training data, the model predicted40.741%diagnosedwitha40.741% naccuracya arysummaryofof ourour iththebestmodelshighlighted.using neural network is shown in Table 2 with the best models highlighted.

Healthcare 2017, 5, 87 of 13Table 2. Summary of Neural Network analysesHealthcare 2017, 5, 8ChronicDiseaseInput101010Heart Attack1010Heart Attack10101010Stroke10Stroke 101010Chronic ngTestingHiddenNodesAccuracyLayersTable 2. Summary of Neural Network analyses.Data Size 290750501Auto 595.0Output Training Testing Hidden LayersNodesAccuracy70301Auto 296.0Data Size 290760401Auto 395.11501295.050 5050(3,8)Auto 5 95.0170301Auto 296.050501695.0160401Auto 395.150 5050(4,5)(3,8)95.01502295.01501195.050 5050Auto6 196.8150295.070 50301Auto(4,5)497.6150501Auto196.860401Auto 296.6170301Auto 497.650501996.8160401Auto 296.650 5050(2,9)996.81501296.81502296.850 5050(9,9)(2,9)96.8150296.850 50502(3,3)(9,9)96.8150502(3,3)96.8Data Size 550Data Size 55050501Auto 289.01501189.070 5030AutoAuto2 2 89.1170301Auto 289.160 60401AutoAuto3 3 87.7140187.750 50509 985.11501185.11502286.850 5050(3,3)(3,3)86.87 of 13Top 3 Predictor Importanceage, work, fruitTop 3 Predictor Importanceage, vegetable, workage, fruit, alcoholage, vegetable,work, fruit incomeage,age, vegetable, workage,fruit,sugarage, fruit, alcoholage,workage, sugar,vegetable,incomeage, fruit,sugar,sugar fruitMarital,age, sugar, workAge,work sugarMarital,sugar,sugarfruitAge,marital,Age, work sugarSugar,work,fruitAge, marital, sugarAge,workalcoholSugar,work,fruitAge, workalcoholSmoke,fruit,workSmoke, fruit, workAge,work, incomeAge, work, incomeWork, marital fruitWork, maritalfruit workVegetable,marital,Vegetable, marital, workAge,marital, workAge, marital, workFruit,sugarFruit, vegetable,vegetable, sugarWork, vegetable,Work,vegetable,incomeincome4.2. ComparisonComparison with Other ModelsAssociation4.2.1. Associationthreshold (minimum(minimum confidence)confidence) waswas setset toto 94%94% in keepingkeeping withwith the high accuracyaccuracyThe thresholdTherewere68 people(12.386%of all records)who weremarried,requirement ewere68 people(12.386%of all records)whoweredid not drinkin thepastandhadandsmokedat least 100cigarettesin their inentireOf life.thismarried,did notdrinkin 30thedays,past 30days,had smokedat least100 cigarettestheirlife.entiregroup,were diagnosedwith a stroke.114marriedpeople whoseis moreOfthis 97.059%group, 97.059%were diagnosedwith a Therestroke.wereTherewere114 marriedpeopleincomewhose incomethan 75,000.this group,96.5%were96.5%diagnoseda stroke.Therewere 134peoplewhoseismorethan Of 75,000.Of thisgroup,were withdiagnosedwitha stroke.Therewere134 incomepeopleis more incomethan 75,000.Of this96.4%with werea stroke.The importantforwhoseis morethangroup, 75,000.Ofwerethis diagnosedgroup, 96.4%diagnosedwith apredictorsstroke. Thestroke usingassociationmaritalstatus,alcohol consumption,smoking,andconsumption,income. The ciationare marital status,alcoholsmoking,shownin Figureandincome.The 2.results are shown in Figure 2.FigureFigure 2.2. Association-interpretingAssociation-interpreting thethe results.results.4.2.2.4.2.2. BayesianBayesian NetworksNetworksAnalysisAnalysis usingusing BayesianBayesian networksnetworks showsshows thatthat thethe importantimportant predictorspredictors areare alcohol,alcohol, income,income, ciseandothervariablesdonotindicatecausalityand marital status. The connections between exercise and other variables do not indicate rrelatedness.Mostofthepeoplewhodidnothavearather conditional dependencies or interrelatedness. Most of the people who did not have a dnotsmokemorethan100cigarettesintheirwere those who exercised regularly and who did not smoke more than 100 cigarettes in their entireentirelives.The ismodelis accurate96.09% accuratein the phase.trainingThephase.Themodelwas lessinaccurateinlives. Themodel96.09%in the trainingmodelwasless accuratepredicting

Healthcare 2017, 5, 8Healthcare 2017, 5, 88 of 138 of 13patients diagnoseda strokewith(76.67%)than(76.67%)in predictingwho wereneverdiagnosedwithpredictingpatients withdiagnoseda strokethan patientsin predictingpatientswhowere nevera stroke (98.4%).modeldid notwell didin thephase(70.9%)diagnosedwith a Thestroke(98.4%).Thedomodelnottestingdo wellin thetesting(Figurephase 3).(70.9%) (Figure 3).In summary,summary, wewe showshow gmakeInsmokingmakeaa higherhigherincomeincomeandand withwithsignificantcontributiontotothelong workingmorelikelyto betodiagnosedwith chronicdiseasesdiseasessuch as stroke.In comparinglongworkinghourshoursarearemorelikelybe diagnosedwith chronicsuch asstroke. Invarious dataminingdatatechniques,see that indataset,only theBayesianmodelcomparingvariousmining wetechniques,wea biasedsee thatin a biaseddataset,onlynetworkthe Bayesianworked wellbecausethewellmajorityof thethepredictorsproject ausethe predictorsmajority ofinthein theproject werecategorical.and associationcould predictonlycouldone ofpredictthe twoonlygroups(in twothis case,thosewhoNeuralnetworktechniquesand associationtechniquesonewellof erstroke).After adjustingto an unbiaseddataset,predictivetechniquescase,who werediagnosedwith stroke).After adjustingto allan theunbiaseddataset,all theworked well.predictivetechniques worked well.FigureFigure 3.3. BayesianBayesianNetworks:Networks: ModelModelsummarysummaryandandpredictorpredictor importance.importance.5.5. ScopeScopeandand LimitationsLimitationsOurFirst,ourourstudyis cross-sectionaland coversthe yearOur s.limitations.First,studyis cross-sectionaland mespan.However,theresearchhasv

2.3. The Neural Network Model A neural network consists of a series of processing elements called neurons that are interlinked to form a network. Each link has a weight associated with it. Each neuron receives stimuli (information) from the surrounding neurons that are linked to it