RESEARCH ARTICLE OpenAccess Optimalsequenceoftestsforthe .

1y ago
5 Views
3 Downloads
1.64 MB
14 Pages
Last View : 1m ago
Last Download : 2m ago
Upload by : Lucca Devoe
Transcription

Luque et al. BMC Medical Informatics and Decision Making (2016) 16:9DOI 10.1186/s12911-016-0246-yRESEARCH ARTICLEOpen AccessOptimal sequence of tests for themediastinal staging of non-small cell lungcancerManuel Luque1* , Francisco Javier Díez1 and Carlos Disdier2AbstractBackground: Non-small cell lung cancer (NSCLC) is the most prevalent type of lung cancer and the most difficult topredict. When there are no distant metastases, the optimal therapy depends mainly on whether there are malignantlymph nodes in the mediastinum. Given the vigorous debate among specialists about which tests should be used, ourgoal was to determine the optimal sequence of tests for each patient.Methods: We have built an influence diagram (ID) that represents the possible tests, their costs, and their outcomes.This model is equivalent to a decision tree containing millions of branches. In the first evaluation, we only took intoaccount the clinical outcomes (effectiveness). In the second, we used a willingness-to-pay of e 30,000 per qualityadjusted life year (QALY) to convert economic costs into effectiveness. We assigned a second-order probabilitydistribution to each parameter in order to conduct several types of sensitivity analysis.Results: Two strategies were obtained using two different criteria. When considering only effectiveness, a positivecomputed tomography (CT) scan must be followed by a transbronchial needle aspiration (TBNA), an endobronchialultrasound (EBUS), and an endoscopic ultrasound (EUS). When the CT scan is negative, a positron emissiontomography (PET), EBUS, and EUS are performed. If the TBNA or the PET is positive, then a mediastinoscopy isperformed only if the EBUS and EUS are negative. If the TBNA or the PET is negative, then a mediastinoscopy isperformed only if the EBUS and the EUS give contradictory results. When taking into account economic costs, apositive CT scan is followed by a TBNA; an EBUS is done only when the CT scan or the TBNA is negative.This recommendation of performing a TBNA in certain cases should be discussed by the pneumology communitybecause TBNA is a cheap technique that could avoid an EBUS, an expensive test, for many patients.Conclusions: We have determined the optimal sequence of tests for the mediastinal staging of NSCLC byconsidering sensitivity, specificity, and the economic cost of each test. The main novelty of our study is therecommendation of performing TBNA whenever the CT scan is positive. Our model is publicly available so thatdifferent experts can populate it with their own parameters and re-examine its conclusions. It is therefore proposed asan evidence-based instrument for reaching a consensus.Keywords: Decision making under uncertainty, Cost-effectiveness analysis in medicine, Probabilistic graphicalmodels, Influence diagrams, Bayesian networks*Correspondence: mluque@dia.uned.es1 Dept. Artificial Intelligence, UNED, Juan del Rosal, 16, 28040 Madrid, SpainFull list of author information is available at the end of the article 2016 Luque et al. Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, andreproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to theCreative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication o/1.0/) applies to the data made available in this article, unless otherwise stated.

Luque et al. BMC Medical Informatics and Decision Making (2016) 16:9Page 2 of 14BackgroundMethodsLung cancer is a very common tumor in the developedworld and the leading cause of cancer death. Lung cancer can be classified into two major types: small-cell lungcancer and NSCLC. The former, which accounts for 20 %of cases [1], is usually inoperable and treatable only withchemotherapy or chemoradiotherapy. In contrast, whenit is limited to the lung, to certain adjacent structures,and to lymph nodes proximal to the lung, NSCLC can betreated with surgical resection. However, more than 80 %of NSCLC patients cannot be treated with surgery becausethe disease is out of control due to a metastasis [2]. A disappointing fact is that a high percentage of patients thatmay benefit from surgery die of lung cancer. A correctassessment at an early stage of the disease—the stagingphase—would help to determine which patients may benefit from surgery and, in turn, to avoid dangerous, painful,and unnecessary surgery when metastasis has alreadyoccurred.When there are no distant metastases, mediastinalstaging, i.e., determining whether malignant mediastinal lymph nodes are present or absent, is the mostimportant prognostic factor in patients with NSCLCand, consequently, determines the therapeutic strategy.Various techniques are available to study the mediastinum, such as non-invasive imaging techniques (CTscan and PET) and minimally invasive endoscopic techniques (TBNA, EBUS, EUS), involving varying degreesof sensitivity and specificity; more invasive surgical techniques include mediastinoscopy. The main treatmentoptions for lung cancer include surgery, chemotherapy,radiation therapy, chemoradiotherapy, palliative and supportive care, and no treatment. The applicability ofeach treatment depends on the stage of the tumor.Due to the variety of available tests and treatments forNSCLC, each one with its pros and cons and different economic costs, there is a vigorous debate amongspecialists about which tests and treatments shouldbe used to strike a balance between effectiveness andcosts [3, 4].In an attempt to clarify this controversy using anevidence-based approach [5–7], we have built an IDfor this problem from the perspective of the Spanishpublic health system. The ID was evaluated twice,first without considering economic costs, and then byconverting costs into effectiveness using a willingnessto-pay of e30,000 per QALY, the shadow thresholdestimated for that health system [8, 9]. We performedseveral types of sensitivity analysis to study the effectof the uncertainty in the numerical parameters of themodel.This paper has been written following the Consolidated Health Economic Evaluation ReportingStandards [10].PreliminariesThis section describes IDs, the explanation capabilitiesavailable in the software tool used to build the model,and the basic principles of cost-effectiveness analysis inmedicine.Influence diagramsDecision trees [11] are a traditional framework formodeling decision problems in medicine. Since decisiontrees explicitly represent all the possible decision scenarios, the size of the model grows exponentially with thenumber of variables. That combinatorial explosion makesthe use of decision trees prohibitive for medium or largeproblems.IDs [12, 13] arose as an alternative to decision trees.Their compactness, based on a causal graph, eases communication with experts, simplifies the solution anddebugging, and thus makes IDs appropriate for muchlarger decision problems.We start by considering an example of a medicalID. A physician has to decide whether to treat ornot a patient, who may suffer from a disease (X).Before deciding how to treat the patient (D), the physician can perform a test (decision T), whose result(Y ) will help determine whether the patient suffersfrom the disease. The overall effectiveness results fromsubstracting the morbidity of the test (U1 ) from thequality of life that results from treating the patient(U2 ).Formally, an ID consists of an acyclic directed graphhaving three disjoint sets of nodes: decision nodes VD(graphically represented by squares or rectangles), chancenodes VC (circles or ovals), and utility nodes VU (diamonds or hexagons). Decision nodes represent the actionsunder the direct control of the decision maker. Chancenodes represent uncertain events. In medical IDs, utility nodes represent medical outcomes and costs (qualityof life, morbidity, mortality, economic cost.). Here, twotypes of utility nodes are distinguished: ordinary, havingparents that are chance and decision nodes, and supervalue (SVN), having parents that are other utility nodes[14]. Given that each node represents a variable, we willuse the concepts of node and variable interchangeably.We assume that all the chance and decision variables arediscrete.IDs contain three types of arcs, depending on the typeof node they go into. Arcs into chance nodes representprobabilistic dependencies. Arcs into decision nodes represent availability of information or precedence relationsbetween decisions. Arcs into ordinary utility nodes indicate the domain of the associated utility function; arcs intoa SVN U indicate that the associated utility function is acombination of the utility functions of the parents of U.

Luque et al. BMC Medical Informatics and Decision Making (2016) 16:9Page 3 of 14For example, in Fig. 1 X and Y are chance nodes, T andD are decision nodes, U1 and U2 are utility nodes, and thechild of U1 and U2 is a SVN of type sum. Node X has acausal and probabilistic relationship with node Y. VariableX is not observable, is unknown when making decision Dand there is thus no arc from node X to node D. However, variable Y is observable, its values are known whenmaking decision D, and it can therefore be observed bythe decision maker at that moment. This explains the arcpointing Y to node D.We assume a path connects all the decision nodes, indicating the order in which decisions are made. Let n be thenumber of decisions in the ID. The total order of decisions{D0 , D1 , . . . , Dn 1 } partitions the set of chance variablesinto the collection of sets {C0 , C1 , . . . , Cn }, where Ci , 0 i n, is the subset of chance variables known for Di butunknown for any previous decision, and Cn is the set ofunknown chance variables. Furthermore, the no-forgettinghypothesis [15] is assumed, which states that the decisionmaker recalls all the previous decisions and observations.For example, in Fig. 1 decision T precedes decision D,i.e., D0 T and D1 D. This order partitions the set ofchance variables in the following way: C0 , C1 {Y }and C2 {X}. This partition has a simple interpretation:no chance variable is observed before deciding on T, variable Y is observed after deciding on T but before decidingon D, and variable T is unknown or observed after deciding on D. Thus, when deciding on D, the decision makerknows the values of T and Y.The quantitative information that defines an ID isgiven by (1) assigning to each chance node a conditionalprobability, (2) assigning to each ordinary utility node areal-valued function, and (3) assigning to each SVN autility-combination function. Conditional probabilitiesand utility functions of ordinary utility nodes are represented as tables. In the example, the quantitative information consists of the conditional probabilities P(x) andP(x y) and the utility functions ψ1 (x, d) and ψ2 (t).The optimal policy for a decision D is a function thatmaps each configuration of the variables known in D ontothe option of D that maximizes the expected utility. Thepurpose of evaluating an ID is to compute an optimalstrategy composed of a set of optimal policies, one for eachdecision in the ID.It is well known that inference in probabilistic graphicalmodels, such as Bayesian networks and IDs, is an NP-hardproblem [16, 17]. Although the set of decisions is totallyordered in IDs, the search space of an ID solution algorithm grows exponentially with the number of variables.However, there are algorithms and software packages thatcan evaluate BNs and IDs for many real-world problemsin less than a second.In this paper, the ID was built using Elvira and OpenMarkov, two software tools for probabilistic graphicalmodels [18, 19]. Debugging a probabilistic expert systembased on an ID is very difficult because the output of theevaluation of a medium or large problem may be the resultof thousands of mathematical operations such as sums,products, divisions, and maximizations. Explaining thereasoning is a key factor in the acceptance of expert systems in real-world domains like medicine. One exampleconsists of explaining why the system recommends oneaction instead of another. Lacave et al. [20, 21] describedhow the explanation capabilities of Elvira were useful forbuilding probabilistic medical models and how such capabilities can help make the probabilistic reasoning moreunderstandable to human users. However, it is a very difficult task that has not yet been completely solved in thefield of IDs [20].Cost-effectiveness analysis in medicineFig. 1 Example of medical influence diagram. A patient may sufferfrom a disease (X). Before deciding how to treat the patient (D), thephysician can decide to perform a test (T). This test will produce thetest result (Y), which would help to determine whether the patientsuffers from the disease. The doctor has to select a strategy takinginto consideration the morbidities associated to the test (U1 ) and thehealth state of the patient after treating him (U2 )If economic costs are included in a medical decision problem, the result is a problem with two criteria to optimize:effectiveness, measured in clinical units, usually QALYs,which we want to maximize, and cost, measured in monetary units, which we want to minimize1 . In medicine,this problem is typically solved with cost-effectivenessanalysis [22].The goal of cost-effectiveness analysis is to maximizethe net health benefit (NHB) [23], which is defined asfollows:NHB E C/λ ,(1)

Luque et al. BMC Medical Informatics and Decision Making (2016) 16:9Page 4 of 14where E is effectiveness, C is cost2 , and λ is used to convert effectiveness into cost or vice versa. Parameter λ issometimes called willingness to pay and represents themaximum amount of money an individual is willing tosacrifice to gain a unit of effectiveness (health benefit).The value of λ is always positive but depends on eachdecision maker.Decision variables Each decision on whether to performa laboratory test has been represented by a variable withthe prefix Dec on its name. A node Dec CT scan wasnot included because a CT scan is always done. The decision maker can perform the EBUS and EUS separately,or both at the same time; this is represented by nodeDec EBUS EUS. These decisions forced the addition of anew state no result to the variables TBNA, PET, EBUS,EUS, and MED, to reflect the fact that when a test is notperformed, its result is not available. Variable Treatmentrepresents the set of possible treatments; its states arethoracotomy, chemoradiotherapy, and no treatment.We included a node Treatment because the effectiveness of the tests is due to the guide they offer on whichtreatment to apply.Construction of the modelThis section describes the construction of MEDIASTINET,an ID for determining the optimal sequence of tests for themediastinal staging of NSCLC. The model was developedwith the help of a pneumologist during several oral interviews; the pneumologist is the third author of this paper.The perspective adopted in the model was that from theSpanish public health system. The primary decision makeris the Spanish Ministry of Health. Our model applies topatients that have lung cancer, presumably operable, withno distant metastases.The model assumes that a CT scan is always performed.Silvestri et al. [24] state that a “CT scan is clearly an imperfect means of staging of the mediastinum, but it remainsthe best overall anatomic technique for studying the thorax”. Additionally, the authors highlight the importanceof CT scan, a non-invasive test, for guiding the choice ofnodes for the most invasive techniques.Structure of the graphThe graph structure of the ID (see Fig. 2) was built manually following the expert’s knowledge of the variablesinvolved in the problem and the causal relations betweenthem.Chance variables The TNM classification uses three factors (T, N, and M) to describe the extent of a cancer.The N factor indicates whether regional lymph nodes areaffected. Given that the objective is the mediastinal staging of NSCLC, the value of N [1] has been representedby the chance variable N2 N3. Even though N takes onfour possible values, from N0 to N3, it has been modeled here as a binary variable because cancers are operablefor groups N0 and N1, and inoperable for most of N2and all of N3. The laboratory tests that can be performedare represented by binary variables CT scan, TBNA, PET,EBUS, EUS, and MED (mediastinoscopy). Binary variableMED Sv represents whether the patient survives mediastinoscopy.Nodes representing test results (CT scan, TBNA, PET,EBUS, EUS, EUS, and MED) have a causal and probabilistic relationship with node N2 N3. This justifies thearcs pointing from N2 N3. Moreover, we have drawn arcsfrom CT scan to the other test result variables becausethe CT scan result can influence their sensitivity andspecificity.Utility nodes Utility nodes in the ID of Fig. 2 can begrouped into three sets, each surrounded by a dashedrectangle.The first group represents effectiveness, measured inQALYs. The node Total QALE accumulates the overallquality adjusted life expectancy (QALE) of patients [25].This node sums the morbidities due to medical tests,and the utility function of Net QALE. This last node represents the QALE of patients considering that they candie due to the treatment or to the mediastinoscopy, andis the product of three nodes: Surv QALE representsthe QALE of the survivors of the tests and treatments;MED Survival indicates whether the patient survives themediastinoscopy; and Immediate Survival represents theimmediate survival rate after the treatment.The second group of utility nodes represents cost.Total Economic Cost is the total cost. Its parents in thegraph represent the costs of tests and treatments.Finally, the third group relates cost and effectiveness: λ 1 represents (the inverse of ) the willingnessto pay, Weighted Econ Cost represents λ 1 · C, andNet Health Benefit the NHB (Eq. 1). The reason for usingλ 1 instead of λ is to be able to obtain the optimal policyregardless of cost, as shown below.Table 1 presents a list of all the variables, along with thetype of variable (chance, decision or utility), the domain(discrete or continuous), and the set of possible values(domain).Elicitation of probabilities and utilitiesThere are two types of uncertainty in a model. First-orderuncertainty reflects that the outcomes of some variablesare not under the control of the decision maker. Forexample, even if a doctor knows that the prevalence of adisease is 0.05, s/he cannot know with certainty whethera person randomly chosen has the disease. Second-orderuncertainty reflects that the parameters of the probability distribution are not known with certainty [26, 27]. For

Luque et al. BMC Medical Informatics and Decision Making (2016) 16:9Page 5 of 14Fig. 2 Influence diagram MEDIASTINET. Chance nodes (ovals), except N2 N3 and MED Sv, correspond to the laboratory tests that can be performed.Decision nodes (rectangles) correspond to the decision on the treatment and on whether to perform each laboratory test. Utility nodes (hexagonsor diamonds) have been grouped into three sets, each one surrounded by an orange rectangle background. The first group representseffectiveness, measured in QALYs. The second group of utility nodes represents cost, measured in e. The third group relates cost and effectiveness.Node λ 1 represents (the inverse of) the willingness to payexample, a doctor may not know with certainty the valueof the prevalence of a disease, but s/he could assume thatits distribution follows a Beta with parameters α 5 andβ 95. In this example, the prevalence of the disease is aparameter of the model, while α and β are the parametersof the associated second-order distribution.The next step in the construction of the model was tocomplete the quantitative part of the ID, consisting of a setof probability and utility potentials. There were 46 independent parameters. Each parameter of the model, exceptcosts, had an associated second-order probability: Beta distributions were assigned to the non-extremeprobabilities of the prevalence of the disease, thesensitivities and the specificities of tests, and thesurvival rates, as the data that inform each probabilityparameter of the model are binomial (m cases ofinterest are observed from a set of n observations)[27], and the Beta distribution is the conjugate tobinomial data [28]. Uniform distributions were assigned to themorbidities of the tests. As the literature provided uswith the percentage of patients suffering themorbidities and the medical expert said that theeffects only last for a month after the test, wesubjectively assumed that the quality of life ofpatients due to the morbidities follow a uniformdistribution in the interval [0, 1]. We haveaccordingly selected in this case the type ofdistribution with maximum entropy. This estimationof the quality of life is imprecise, but the subsequentsensitivity analysis demonstrated that the calculatedstrategy was not sensitive to the uncertainty in themorbidity parameters. Although λ was known with certainty, we attached ita second-order distribution in order to analyze its

Luque et al. BMC Medical Informatics and Decision Making (2016) 16:9Page 6 of 14Table 1 Variables of the influence diagram, along with their type,the domain type and the values they can takewere used to fit these distributions. In the case of Betadistributions, four different methods were used dependingon the values provided by the literature:NameTypeDomain typeDomainN2 N3ChanceDiscretepositive; negativeCT scanChanceDiscretepositive; negativeTBNAChanceDiscretepositive; negative;no resultPETChanceDiscretepositive; negative;no resultEBUSChanceDiscretepositive; negative;no resultEUSChanceDiscretepositive; negative;no resultMEDChanceDiscretepositive; negative;no resultMED SvChanceDiscreteyes; noDec TBNADecisionDiscreteyes; noDec PETDecisionDiscreteyes; noDec EBUS EUSDecisionDiscreteebus eus;ebus;eus;no testDec MEDDecisionDiscreteyes; herapy;no treatmentEC CT scanUtilityContinuousR EC TBNAUtilityContinuousR EC PETUtilityContinuousR EC EBUSUtilityContinuousR EC EUSUtilityContinuousR EC MEDUtilityContinuousR EC TreatmentUtilityContinuousR TBNA MorbUtilityContinuousR EBUS MorbUtilityContinuousR EUS MorbUtilityContinuousR MED MorbUtilityContinuousR Immediate SurvivalUtilityContinuous[0, 1]MED SurvivalUtilityDiscrete0; 1Surv QALEUtilityContinuousR λ 1UtilityContinuousR variations in the sensitivity analysis phase. A Gammadistribution was assigned to λ, as it is constrained inthe interval [0, ) [27] and it can be interpreted asa cost. Briggs et al. [27] suggest the use of gamma forcosts because the count data of costs is often usuallyrepresented by the Poisson distribution, and theGamma is the conjugate to the Poisson.The parameters of the second-order distributionswereestimated from the medical literature. Different methods1. If the source indicated that m cases of interest wereobserved from a set of n observations, then weelicited a Beta with shape parameters α m andβ n.2. If the source only provided mean μ, we set acoefficient of variation k, and we assigned thestandard deviation as σ k · μ. Then by using themethod of moments [27] we calculated the shapeparameters α and β: z (μ · (1 μ) · σ 2 ) 1,α z · μ, β z α. We have taken the valuek 1/5, following Bond et al. [29, 30] in a studycommissioned by the National Institute for Healthand Care Excellence (NICE). If μ was so close to 1that it was inconsistent to have a Beta with suchparameters μ and σ , then we took the valuek 10 m · 1/5 for setting σ , where m is theminimum natural number guaranteeing consistency.3. If the source provided mean μ and the 95 %confidence interval (CI) [l, u] of the parameter, wefound a Beta distribution with mean μ so that theinterval [l, u] accumulated a 95 % of the probabilitymass. Given mean μ and considering α anindependent parameter, we have β α · (1 μ)/μ.Then we used a bisection method to find the value ofα that fulfills F(u) F(l) 0.95, being F thecumulative distributive function of Beta.4. If the source provided mean μ and two values l andu, the latter being the maximum and the minimumvalues respectively, but the authors did not indicatethat l and u referred to the 95 % CI, then byassuming that the interval [l, u] accumulates 99.9 %of the probability mass we elicited the distributionanalogously to point 3. This value is arbitrary but doesnot significantly affect the sensitivity analysis results.In the case of the Gamma distribution of λ, we followed an approach similar to case 3 above: a mean μ wasobtained from the source, and we assigned the standarddeviation σ k · μ, being k the coefficient of variationexplained above.Tables 2, 3, 4, 5 and 6 present all the independentparameters. Subindices attached to the CIs indicate thetype of distribution and the type of elicitation used:numbers 1 to 4 correspond to a point above for theBeta distribution, and number 5 corresponds to uniformdistributions.All the explanation capabilities for IDs that Elvira andOpenMarkov offer were useful in this quantitative phaseof model construction [20, 21]. The display of several evidence cases made it possible to introduce evidence from

Luque et al. BMC Medical Informatics and Decision Making (2016) 16:9Table 2 Sensitivities and specificities of the tests when CT scan ispositiveSensitivityPage 7 of 14Table 4 Morbidities and costs of testsMorbiditySpecificityMeanCostCISource Mean CIMeanCIMeanCISourceCT scanTBNA78[40.9, 98.5]299.5[94.3, 100]2[24]TBNAPET91[86.7, 94.5]478[71.9, 83.6]4[56]PETEBUS92.5[25.5, 100]299.5[94.3, 100]2[57]EBUSEUS90[32.5, 100]299.5[94.3, 100]2[58]EUSMED83[39.9, 99.9]299.9[100, 100]2[24]MED0.000833 [0.000042, 0.001625]6 [69]Source199— [62, 63]80— [65]1290— [62, 63]0.000021 [0.000001, 0.000041]6 [66]620— [67, 68]0.000125 [0.000006, 0.000244]6 [66]620— [67, 68]3000— [65]0.000108 [0.000005, 0.000211]6 [64]Mean and 95 % CI values are given in percentages. Subindices denote theelicitation method (see Section “Elicitation of probabilities and utilities”)Moribidities are given in QALYs and costs are given in emedical tests and to study its effect on the posterior probability of the variable N2 N3. For example, Fig. 3 showstwo evidence cases in OpenMarkov: the first one (colored in red) contains no evidence and the probabilityof each variable is the prior probability; the second case(colored in blue) contains two contradictory findings, asthe CT scan is positive and the TBNA is negative. Thebars in node N2 N3 show that the probability of positive N2 N3 is lower in the second case than the priorprobability. When the pneumologist asked us why themodel recommended not performing a certain test, wecould impose a policy on that test and analyze the outcomes. This phase of debugging was essential for checkingthe external consistency of the model, as recommendedin [31].Two strategies were computed using two differentcriteria: maximizing effectiveness (disregarding costs) andmaximizing the NHB. The former was carried out bysetting λ 1 to 0, which according to Eq. 1 means thatcosts are ignored. In the latter, λ 1 1/(30, 000e/QALY) because the shadow willingness-to-pay threshold for the Spanish public health system was estimatedto be 30,000 e/QALY [8, 9]. This threshold is consistentwith the ones proposed by the World Health Organizationfor determining whether an intervention is consideredcost-effective [32, 33].Computing these two strategies with different criteria helps to fulfill the recommendation of checking the internal consistency of the model [31] byassigning extreme values for some parameters. Inour case, we should expect that the strategy provided by the model when setting λ 1 to 0 wouldtend to select expensive tests that are discarded whenλ 30, 000 e/QALY.These two versions of the ID were evaluated using thevariable elimination algorithm [34]. The decision tablefor each decision grows exponentially with the numberof variables known when making the decision. Thesetables are not an adequate output for a human expert,not only because of their size—for example, the policy table for decision Treatment has 15,552 columns—but also because most of the columns correspond toimpossible scenarios. For this reason, an algorithm [35]was developed to transform the set of decision tablesinto a single strategy tree, as shown in Figs. 4 and 5.ResultsComputation and representation of optimal strategiesThe first step after building the model was to evaluate thereference case of MEDIASTINET, i.e., the version in whichevery parameter of the model is set to the mean of its associated second-order probability distribution. To have anidea of how big is the search space of evaluating MEDIASTINET , we must note that the size of the set of possiblestrategies for the set of test decisions is 2.44 1011 .Table 3 Sensitivities and specificities of the tests when CT scan TBNA4[2.6, 5.7]297.1[89.7, 99.9]1[59]PET75[66.2, 82.9]493[92.4, 93.6]4[56]EBUS69.2[39.3, 91.9]299.5[94.3, 100]2[60, 61]EUS58[34.7, 79.5]299.5[94.3, 100]2[58]MED47[28.9, 65.5]299.9[100, 100]2[24]Mean and 95 % CI values are given in percentages. In the case of the sensitivity andthe specificity of the EBUS, we considered that the mean value of the distributionwas the average of the values given by the two referencesTable 5 QALE depending on the treatment, given in QALYspos. N2N3Mean CIThoracotomyneg. N2N3Source Mean CISource1.17 [0.78, 1.56]5 [70]5.75 [5.36, 6.14]5 [71]Chemoradiotherapy 1.25 [0.86, 1.64]5 [70]2.64 [2.25, 3.03]5 [72]0.42 [0.03, 0.81]5 [73]2.08 [1.69, 2.47]5 [73]No treatment

Luque et al

Luqueetal.BMCMedicalInformaticsandDecisionMaking (2016) 16:9 DOI10.1186/s12911-016-0246-y RESEARCH ARTICLE OpenAccess Optimalsequenceoftestsforthe .

Related Documents:

Amendments to the Louisiana Constitution of 1974 Article I Article II Article III Article IV Article V Article VI Article VII Article VIII Article IX Article X Article XI Article XII Article XIII Article XIV Article I: Declaration of Rights Election Ballot # Author Bill/Act # Amendment Sec. Votes for % For Votes Against %

Progress Software Corporation September 2016 5 OpenAccess SDK Version 8.1 Highlights This section highlights what’s new in OpenAccess SDK 8.1: OpenAccess SDK SQL Engine and IP SDK

www.lenel.com OpenAccess Alliance Program Factory Certified Product (FCP) Interface Document OAAP Member Information Member Company Name SMI Global Ltd. Member Contact Name: Dick Clark Phone: 44 7815158203 Email: dick@smi-global.net OnGuard Information OnGuard Version/Build, Firmware OnGuard 7.1 OnGuard 7.2

Article 27 Article 32 26 37 Journeyman Glazier Wages Article 32, Section A (2) 38 Jurisdiction of Work Article 32, Section L 43 Legality Article 2 3 Mechanical Equipment Article 15, Section B 16 Out-of-Area Employers Article 4, Section B 4 Out-of-Area Work Article 4, Section A 4 Overtime Article 32, Section G 41

Jefferson Starship article 83 Jethro Tull (Ian Anderson) article 78 Steve Marriott article 63, 64 Bill Nelson article 96 Iggy Pop article 81 Ramones article 74 Sparks article 79 Stranglers article 87 Steve Winwood article 61 Roy Wood art

1 ARTICLES CONTENTS Page Article 1 Competition Area. 2 Article 2 Equipment. 4 Article 3 Judo Uniform (Judogi). 6 Article 4 Hygiene. 9 Article 5 Referees and Officials. 9 Article 6 Position and Function of the Referee. 11 Article 7 Position and Function of the Judges. 12 Article 8 Gestures. 14 Article 9 Location (Valid Areas).

article 22, call time 41 article 23, standby time 42 article 24, life insurance 42 article 25, health benefits 43 article 26, work-related injuries 51 article 27, classification 55 article 28, discharge, demotion, suspension, and discipline 58 article 29, sen

Brian Nicholas Harris, CC Citizen and Glazier . Barbara Anne Bear Citizen and Musician Maureen Sutherland Smith a Public Relations Consultancy Chairman 78 Hamilton Terrace, St Johns Wood, Westminster The Rt Hon The Lord Mayor Barbara Anne Bear Citizen and Musician Michael John Henesy a Meat and Fish Specialist 49 Hatton House, St Georges Estate, Stepney, Tower Hamlets Susan Margaret Hughes .