ARTICLE IN PRESS

Respiratory Medicine (2007) 101, 2145–2152

0954-6111/$ - see frodoi:10.1016/j.rmed.

�Corresponding auF-75679 Paris Cedex

E-mail address: a

Patients’ perception of asthma severity

Alain Luriea,�, Christophe Marsalab, Sarah Hartleyc,Bernadette Bouchon-Meunierb, Daniel Dussera

aService de Pneumologie, Hopital Cochin, Assistance Publique-Hopitaux de Paris, 27 rue du Faubourg Saint-Jacques,F-75679 Paris Cedex 14, FrancebUniversite Pierre et Marie Curie-Paris6, UMR 7606, LIP6, 104 avenue du President Kennedy,F-75016 Paris, FrancecLaboratoire d’Explorations Fonctionnelles Respiratoires et du Sommeil, Hertford British Hospital,3 rue Barbes, F-92300 Levallois-Perret, France

Received 6 October 2006; accepted 13 May 2007Available online 23 July 2007

KEYWORDSAsthma;Asthma severity;Guidelines;Decision-makinganalysis;Fuzzy logic

nt matter & 20072007.05.027

thor. Service de Pn14, France. Tel.:

lain.lurie@laposte

SummaryObjectives: To identify variables patients use to determine the severity of their asthma,the perceived severity (PS), using a fuzzy decision-making analysis (FDMA). To comparethese variables with those involved in the assessment of asthma severity according to theglobal initiative for asthma (GINA) guidelines, the objective severity (OS).Patients: Outpatients (51 men, 62 women), aged (m7SD) 42.9716.3 years with(% patients) mild intermittent (6.2), mild persistent (15.9), moderate (65.5) and severe(12.4) asthma.Design: Cross sectional, observational study.Methods: Both OS (rated by doctors) and PS (rated by patients) were rated as mildintermittent, mild persistent, moderate, or severe. Variables involved in OS assessment,variables self-assessed by patients (dyspnea, perceived treatment efficacy, asthma-relatedquality of life questionnaire [AQLQ]), patients’ sociodemographic characteristics, andasthma characteristics, were evaluated with questionnaires. These variables were pooled,and considered as potential variables patients might use to determine their PS. They weretested against the PS measurement using FDMA. This identified variables patients actuallyused to determine PS.Results: On the day of consultation, 68.1% of patients classed their asthma as mildintermittent or mild persistent, 23.9% as moderate persistent, and 8.0% as severepersistent. There was a significant discrepancy (po0.01) between PS and OS with a clearpatient tendency to underestimate asthma severity as compared to OS. Patientsdetermined PS level according to variables assessing their asthma perception, i.e., AQLQmeasures and dyspnea, but not variables involved in OS assessment, such as symptom

Published by Elsevier Ltd.

eumologie, Hospital Cochin, Assistance Publique-Hopitaux de Paris, 27 rue du Faubourg Saint-Jacques,+33 6 60 56 86 86; fax: +33 1 42 88 53 35.

.net (A. Lurie).

ARTICLE IN PRESS

A. Lurie et al.2146

frequency or knowledge of their peak flow rates. Duration of asthma and treatmentcharacteristics were also involved.Conclusion: FDMA identified variables patients used to determine PS. It highlighted adiscrepancy between patients’ and doctors’ perceptions of asthma severity, suggestingthat assessment of asthma severity should consider both patients’ and doctors’perceptions of the disease and includes an AQLQ measure.& 2007 Published by Elsevier Ltd.

Introduction

The global initiative for asthma (GINA) guidelines recognizefour distinct levels of asthma severity (objective severity(OS)).1,2 OS is assessed by rating a combination of symptoms,lung function tests, and medication use, and is the key tosuccessful management using stepped care pharmacother-apy. Although OS has been thoroughly studied, littleresearch has investigated patients’ perception of asthmaseverity (perceived severity (PS)).3–12 Variables which havebeen identified to influence patients’ perception of theirasthma burden include symptom frequency, symptombother, activity limitation, acute exacerbation,13 numbersof missed workdays, overall health status, and prescriptionsfor inhaled corticosteroids and long acting bronchodila-tors.14 The correlation between self-PS of asthma andobjective assessment of severity on the basis of GINAcriteria15 or on the National Asthma Education and Preven-tion Program Expert Panel II Management Guidelines hasbeen shown to be consistently poor.14 Patients with asthmatend to underestimate their asthma severity as comparedwith asthma severity assessed by physicians.15 In anepidemiological study, patients with mild asthma have beenshown to be less prone to report their asthma than patientswith moderate or severe asthma.16 It seems, therefore, thatpatients and physicians use different variables to defineasthma severity.

To our knowledge, the variables patients take intoconsideration when rating the severity of their own asthma(i.e., how they make the decision to rate the severity oftheir own asthma as, for example, mild or severe) are notknown. Several kinds of variables may be involved in adecision-making process. For numerical variables (e.g., peakexpiratory flow rates, PEFR), a decision may vary dependingon whether the value is above or below a specified limit (thethreshold). For example, patients with usually normal PEFRmay be instructed to increase their dose of inhaledcorticosteroids when their PEFR is p60%. In reality, adecision may depend on a fuzzy threshold (it may be 65%or even ‘‘about 60%’’) or on intuitive, non-explicit variables(as in subjective judgments). Methods using fuzzy logic aresuited to managing these variables, when a precisemathematical description of a process is not possible.17–19

Fuzzy logic has been used to help decision-making in severalmedical specialities.19–22 The methods are derived from thefuzzy set theory that was introduced in 1965 by Zadeh23 tosimulate the process of human reasoning.17,19 A fuzzy set isan extension of a classical set when the boundaries of theset cannot be clearly defined (Scheme 1).

Inductive learning analysis is used extensively whenperforming data analysis to describe and to analyze data,

in order to help a decision-making process.19–22,24,25 Itenables us to highlight dependence between variables.When coupled with fuzzy logic (i.e., fuzzy decision-makinganalysis, FDMA), it can provide a good simulation of‘‘human’’ decision-making processes.17,19,23 It is a meansto identify and rank variables used in an individual’sdecision-making process.26,27 It shows the most probablerelationship between the identified variables and thedecision: for a given individual, the values of theseidentified variables enable the result of their decision tobe predicted.

First, we used an FDMA to explore and rank by order ofimportance the variables patients took into account todetermine their PS level. Next, we compared thesevariables with those involved in the assessment of asthmaseverity according to the GINA guidelines.

Methods

Patients

Asthmatic patients attending routine follow-up appoint-ments were sequentially recruited from two respiratoryoutpatient clinics, at Cochin University Hospital, Paris, andthe Hertford British Hospital, Levallois-Perret, France. Atthe time that the study took place, the French Health CareSystem allowed a patient to consult a specialist directlywithout having previously consulted a general practitioner.Exclusion criteria were co-existing chronic disease (such aspsychiatric or rheumatologic disorders), pregnancy, orhaving given birth less than 3 months prior to inclusion.The study was approved by the Cochin University HospitalEthics Committee, and all patients gave written informedconsent.

Design and description of the study

This study was conducted in naturalistic conditions, during asingle outpatient consultation.

Outcome measures

Variables assessed by patientsAssessment of asthma severity according to patients—PS.PS on the day of consultation and over the past 6 weeks wereself-assessed by the patient using a 4-point Likert scale(mild intermittent, mild persistent, moderate persistent,and severe persistent).

Assessment of dyspnea, perceived response to treatmentsuccess, and asthma-related quality of life. Dyspnea was

ARTICLE IN PRESS

30 500

Set m

embe

rshi

p

0

1

Set of patients with ‘Perceived

moderate asthma’

Overlap

Variable ‘Var’

Set of patientswith ‘Perceived

severeasthma’

Scheme 1 Overlap of fuzzy sets. To illustrate the overlap of fuzzy sets, let us consider the set of patients that perceive their asthmaas moderate (the ‘‘Perceived moderate’’ set) and the set of patients that perceive their asthma as severe (the ‘‘Perceived severe’’

set). Patients perceive their asthma asmoderate or severe depending on whether a given variable (labeled here ‘‘Var’’) is above orbelow a threshold value. As the perception is subjective, a unique threshold value that separates the values of Var into 2 subsets doesnot exist: the frontier between the ‘‘Perceived moderate’’ set and the ‘‘Perceived severe’’ set is fuzzy. In the fuzzy set theory, tobetter reflect real life decision-making processes, the limit between two sets is not considered as a single value but is associatedwith an interval of values. For example, when Var is p30, patients unquestionably perceive their asthma as moderate and, when Varis X50, patients unquestionably perceive their asthma as severe. However, when Var ranges from 30 to 50, patients may perceivetheir asthma as more or less severe depending on the value of Var.

Patients’ perception of asthma severity 2147

assessed on the day of consultation and over the past6 weeks using a 100mm horizontal visual analog scale,from ‘‘not at all breathless’’ to ‘‘extremely breathless’’.Perceived response to treatment was measured using a5-point Likert scale (1 ¼ not at all effective to 5 ¼ veryeffective). Asthma-related quality of life was measuredusing the validated French version of the Marks—asthmaquality of life questionnaire (AQLQ),28,29 a 20-item ques-tionnaire with Likert scale responses. Sub-scores for breath-lessness, mood disturbance, social disruption, and concernfor health, were calculated.

Variables assessed by doctorsAssessment of asthma severity according to the GINAguidelines—OS. Doctors assessed OS by answering a ques-tionnaire, while referring to the patients’ case report forms.Overall asthma severity, diurnal symptom severity, noctur-nal symptom severity, exacerbation severity, and respiratoryfunction severity, were rated as mild intermittent, mildpersistent, moderate persistent or severe persistent.1,2

Symptoms mentioned on the case report forms includedcough, wheezing, dyspnea, and chest tightness. Treatment‘‘severity’’ was similarly categorized according to the levelof pharmacotherapy used.1,2

Respiratory function tests. Forced expiratory volume(FEV1) was determined with a Dyn’Rs spirometer (Dyn’R,Le Muret, France). PEFR was measured with a MiniWrightpeak flow meter (Airmeds, London, UK). PEFR and FEV1were expressed as the percentage of the predicted value.

Data analysis

FDMAFuzzy logic was applied to an inductive learning analysis toproduce a testable model of the variables patients took intoaccount to determine their PS level. The FDMA process isdescribed in Appendix. As about 40% of patients with mild PS

seemed to have difficulties in discriminating mild inter-mittent PS from mild persistent PS, these two categorieswere collapsed into one category (mild PS). Thus, PS wasdivided into three classes: mild, moderate, and severe PS.FDMA identified and ranked variables patients may use todetermine their PS level. The variable ranked ]1, the root ofthe decision-making process, represents the variable that isthe most important to all patients in determining their PSlevel. The higher the rank (2nd, 3rd, etc.), the lessimportant the variable is in the decision making process.Variables ranked greater than five were not considered asthey are associated with a too small number of patients tobe considered as important. As the same variable can beused in different ranks of the process, it may appear twice.

Statistical analysisCategorical variables were compared by w2 tests; p valuesless than 0.05 were considered significant. Analyses werecarried out using GB-STAT software (Dynamic Microsystems,Silver Spring, MD).

Results

Study population and asthma characteristics. One hundredand twenty five patients with asthma were asked toparticipate in the study. Because the study simply requiredthe patients to complete a questionnaire while attending aroutine consultation, participation rate was high. Twelvepatients approached refused to participate. Data was notcollected on non-participants. One hundred and thirteenadult outpatients with asthma (51 men, 62 women) wererecruited, with an average age (7SD) of 42.9716.3 years(Table 1). Most patients were well educated, with 42.5%having had tertiary (university) education (Table 2). Severityaccording to GINA guidelines was (% of the total number ofpatients): mild intermittent (6.2), mild persistent (15.9),moderate persistent (65.5) and severe persistent (12.4).

ARTICLE IN PRESS

Table 1 Characteristics of patients’ asthma.

Asthma history (mean7SD)Duration of asthma (years) 16.8714.1Tobacco consumption (pack-years) 5.172.9Number of visits to the doctora 4.374.3Number of days missed at worka 2.379.9

Respiratory function (mean7SD)FEV1 (% predicted) 83.6718.7PEFR (% predicted) 81.5719.6

Asthma treatment (% patients)Inhaled corticosteroid 91.1Inhaled long-acting beta-2 agonist 36.3Long-acting theophylline 7.9Inhaled anticholinergic 3.5Inhaled anti-allergic 3.5Inhaled short-acting beta-2 agonist as

required75.2

FEV1: Forced expiratory volume in 1 s; PEFR: peak expiratoryflow rate.

aDuring the last 12 months.

Table 2 Socioeconomic characteristics of patients.

Education (% of patients)Primary education only 4.4Secondary education 8.0Secondary technical education 16.0Baccalaureat (high school leaving certificate) 23.8Technical college 5.3University degree 42.5

Family situation (% patients)Single (never married) 29.3Married/cohabiting 56.6Divorced 9.7Widowed 4.4

Occupational profile (% patients)Full-time work 46.0Part-time work 8.0Work at home 6.2Student 14.2Retired 16.8Unemployed 3.5Partial disability 5.3

Table 3 Dyspnea, treatment efficacy, and quality oflife measurements as perceived by patients.

Variables studied Results

Dyspneaa (mean7SD, mm)On the day of the study 23.3724.7During the previous 6 weeks 31.1725.1

Perception of treatment efficacy (% of patients)Not at all effective 1.8A little bit effective 5.3Moderately effective 12.4Effective 38.0Very effective 42.5

AQLQb (mean7SD)Total 2.372.0Breathlessness 2.972.4Mood disturbance 2.272.2Concerns for health 2.272.1Social disruption 1.972.1

aDyspnea was assessed using a 100mm horizontal visualanalog scale. The higher the score, the more breathless thepatient feels.

bAQLQ: asthma quality of life questionnaire. Total scoreand sub-scores for breathlessness, mood disturbance, con-cerns for health, and social disruption were calculated. Totaland subscale scores range from 0 to 10 (higher scoresrepresent a greater impact of asthma on quality of life).

A. Lurie et al.2148

PS. On the day of consultation, 68.1% of patients classedtheir asthma as mild, 23.9% as moderate persistent, and8.0% as severe persistent. Over the preceding 6 weeks,65.2% of patients classed their asthma as mild, 24.8% asmoderate persistent, and 10.0% as severe persistent. PS onthe day of the study and during the previous 6 weeks werenot statistically different (p40.05). There was a significantdiscrepancy (po0.01) between PS on the day of consultationand OS with a clear patients’ tendency to underestimatetheir asthma severity as compared to OS.

Dyspnea, treatment efficacy, and quality of life measure-ments as perceived by patients are reported in Table 3.

Variables found to be used by patients to determine PSlevel. These are reported in Table 4. Variables assessing thepatients’ perceptions of asthma (e.g., quality of lifemeasures, and dyspnea) were the most important, occupy-ing the first four ranks and accounting for six of the 13variables. Variables assessing some aspects of treatmentcharacteristics were involved and ranked 3rd, 4th, and 5th,i.e., ‘‘Oral corticosteroid use during the last 12 months’’(rank ]3), ‘Total number of puffs’ and ‘‘Number of puffs ofinhaled corticosteroids’’ (rank ]4), and ‘‘Long-acting theo-phylline use’’ (rank ]5). The PS also depended on patients’demographic characteristics and asthma characteristics(i.e., duration of asthma). However, variables assessingasthma severity according to GINA guidelines, such assymptom frequency or patient knowledge of their PEFR,were not considered by patients to influence their PS.

Cross-validation tests showed a mean rate of 7377%,indicating an accuracy of 73% (a decision taken from thismodel can be considered correct in about three out of fourcases) with a high robustness (7%).

Discussion

Our study shows (i) that FDMA identified and ranked thevariables actually used by patients to determine the severityof their asthma, (ii) that these variables were very differentfrom those used by doctors to assess asthma severity, and

ARTICLE IN PRESS

Table 4 Variables found to be used by patients todetermine their asthma severity level, in decreasingorder of importance.

Rank oforder

Variables

Rank ]1 AQLQ concern for healtha,b

Rank ]2 Duration of asthmab

Dyspnea on the day of the studyb

Rank ]3 AQLQ concern for healthb

Duration of asthmab

Short course of oral corticosteroids during thelast 12 monthsPerceived severity during the previous 6 weeks

Rank ]4 AgeTotal number of puffs taken dailyInhaled corticosteroids (number of puffs)AQLQ breathlessnessDyspnea on the day of the studyb

Rank ]5 Long-acting theophylline use

aAQLQ concern for health was found to be the root of thedecision-making process, which represented the variablewhich was important to all patients in determining their PSlevel. The other ranks show the other variables patientsused, in decreasing order of importance.

bThe same variable can be used at different level of adecision-making process, and can thus appear at differentranks. AQLQ: asthma quality of life questionnaire. Therelevant sub dimension of the AQLQ is shown in brackets.

Patients’ perception of asthma severity 2149

(iii) the importance of AQLQ measures for patients indetermining their PS level.

Up to now, studies have used FDMA to develop decision-making models.19–22 To our knowledge, our study is the firstto use an FDMA to explore a patient’s decision-makingprocess. FDMA identified key variables patients used toassess their PS level. It was created without prior modelspecification, unlike systems created solely by experts usingselected variables,30 which may ignore variables that areimportant to patients. FDMA and statistics are two com-plementary methods but rather different. In statistics, theaim is to tackle problems associated with laws of probabilitythat govern them. In FDMA, no such hypothesis is assumedand the aim is to tackle problems even if they are notgoverned by an underlying law of probability. In our study,the decision-making process is related to very subjectiveknowledge (perception by the patients). We believe that alaw of probability explaining such subjective relations doesnot exist, or, if such an underlying law does exist, that noassumption about its properties can be made. Testing ourmodel against a ranking derived from statistical tests has thelimitation that the statistical tests are not appropriate forthese data. Moreover, FDMA has shown the most probablerelationship between these variables and the level of PS,whereas studies that have used conventional statistics ascorrelations have only shown that given variables areassociated to PS.3–12

Patients in this study tended to underestimate theirasthma severity. This was also remarkably and consistentlyreported in an international survey in 29 countries assessing

international variations in the severity and management ofasthma.15 The discrepancy between patient assessed sever-ity and the severity according to the GINA guidelines isinteresting but may not be surprising. Firstly, there is noagreed ‘‘gold standard’’ for classifying asthma severity, andthe lack of consensus between guidelines leads to discre-pancies in severity assessment.1,2,8,31–35 Secondly, thevariables considered to be important to patients are notcentral to GINA guidelines assessment of severity: (i) Wefound that the way in which asthma affected the patient’slife (i.e., quality of life and symptom perception) was moreimportant in their judgment of severity than symptomfrequency. The importance of quality of life in severityassessment has been confirmed by previous studies.6,14,36,37

Moreover, quality of life is only weakly correlated withasthma symptoms, and is weakly or not at all correlatedwith lung function.6,37–41 In the validation process of ahealth-related quality of life measure, the face and contentvalidations take into account the perception of patients anddoctors, respectively. The use of AQLQ measure maytherefore improve asthma severity assessment from thepatients’ point of view. On the other hand, Schatz et al.13

identified that symptom frequency influenced the patients’perception of their asthma burden, independently ofvariables such as symptom bother, activity limitation, andacute exacerbation. (ii) The importance accorded to asthmaduration by our patients has also been found in children, inwhom asthma duration is associated with objective mea-sures of severity.42 (iii) The guidelines give weight to theworst individual variable, which emphasize the importanceof nocturnal symptoms.4,7,43 We confirmed the study byOsman et al.44 who found that these are less important topatients than daytime symptoms such as breathlessness. (iv)Measurements of lung function heavily influence GINAguidelines assessed severity, and these are known tocorrelate poorly with patient reported symptoms.5,41,45 Thiswas also confirmed by our study, which found that lungfunction measurements (patient knowledge of their PEFR)did not influence patients’ estimates of their diseaseseverity.

We cannot rule out that some patients might haveconfused severity and control of their asthma. However,the FDMA identified variables (e.g., those related to asthmatreatment) that GINA guidelines recommend to be used toassess asthma severity.1,2 Measures that would distinguishthe construct of severity from that of control remain to beidentified.13 In fact, measures of control and severity arenot independent as shown by Combescure et al.46: asthmacontrol depends on the severity level of the disease(this study was performed in naturalistic conditions whichwere similar to those of our study). This may be related, inpart, to the fact that assessment tools for severity andcontrol of asthma use similar variables, e.g., symptom andexacerbation frequency, and lung function.47

The results that we have do not permit us to explore thedifferences between patients’ and doctors’ views ofseverity. Like Yawn et al.,14 we can only conclude thatvariables used by doctors to assess asthma severity are notappropriate to assess the patients’ perception of theirasthma severity. Doctors need to develop a tool to assessseverity from both the physicians’ and the patients’perspectives, used along with objective lung function tests

ARTICLE IN PRESS

A. Lurie et al.2150

when available, to more accurately assess asthma severity.Our findings support the suggestion that the clinicalparameters of the GINA guidelines and variables such ashealth-related quality of life represent distinct dimensionsin the measurement of asthma severity, both of which arevaluable in determining the severity of the disease.37,48

The patients were recruited sequentially from two publichospitals without evident recruitment bias. However, theeducation level of our population was high, and there was aslight predominance of women. When compared to malegender, female gender is associated with a higher FEV1, acritical determinant of the severity assessed with GINAguidelines.37,49 Thus, we might expect that our patientswould have less severe asthma when compared with the‘‘general population of patients with asthma’’. At the timethat the study took place, the French Health Care systemallowed patients to consult a specialist directly for care oftheir asthma and the distribution of adult asthma severity insecondary care was not known. The effect of our populationon quality of life scores is difficult to determine, as theincreased health-related quality of life scores associatedwith higher educational level3 may have been negated bythe reduced health-related quality of life scores associatedwith female gender.37,49

In conclusion, we report, using a FDMA, a discrepancybetween patients’ and doctors’ perceptions of asthmaseverity. Future studies assessing asthma severity shouldtake into consideration both patients’ and doctors’ percep-tions of asthma severity and should include an AQLQmeasure. We believe that FDMA might be also applied invarious other domains in order to assess patients’ decision-making processing or behaviors.

Acknowledgments

We thank Professor E. Sanchez for his helpful commentsregarding this manuscript, and Anne Compagnon, Ph.D.(AstraZeneca Laboratories, France) for her support. Wegratefully acknowledge the financial support received fromAstraZeneca Laboratories, France, and the French ForeignAffairs Department.

Appendix. The FDMA

How the FDMA was performed using the set ofpatients studied

Each patient is characterized by their PS level and the set ofvalues of all variables studied. The values of the variable‘‘PS’’ (i.e., the PS level) indicate the result of a decisionprocess made by each of the patients to determine theseverity level of their own asthma. PS is divided into threelevels, i.e., mild, moderate and severe. The set of values ofall the variables studied included those involved in OSassessment, variables self-assessed by patients (dyspnea,perceived treatment efficacy, asthma-related quality of lifequestionnaire [AQLQ]), patients’ socio-demographic char-acteristics, and asthma characteristics. A given patient wasdescribed by means of a value of each variable, and wasassociated with a given level of PS. For example, for a givenpatient, ‘‘AQLQ concern for health ¼ 2’’, ‘‘Duration of

Asthma ¼ 15 years’’, ‘‘Dyspnea ¼ 30mm’’, ‘‘Age ¼ 37years’’, ‘‘Oral CS ¼ yes’’, etc., were associated with ‘‘mildPS’’. These variables were pooled, and considered aspotential variables patients might use to determine theirPS. They were tested against the PS measurement usingFDMA. This identified and ranked variables patients actuallyused to determine PS, without making a priori assumptionsabout relationships between variables.

How variables used by patients to determine theirPS level were identified and ranked

The analysis was performed by means of the so-called TopDown Induction of Decision Tree algorithm.50 The reasons wechose this approach among others are explained below. Theanalysis process was performed, first, by determining thevariable ranked ]1, the root of the decision-making process,which represents the variable that is the most important toall patients in determining their PS level. Then, variables ofhigher ranks were identified.

Identification and ranking of variables of higher rankswere done by successive splits of the set into subsetsaccording to the predictive values of variables:

�

The predictive value of variables is such that: ‘‘knowingvalues of important (i.e., informative) variables leads tothe knowledge of the PS level’’. The Shannon entropy, ameasure from the information theory, was used to assessthese predictive values.51 The entropy measures the linkexisting between variables and PS. It is composed of theprobabilities of PS levels related to variables used bypatients to determine their PS level. The lower theShannon entropy, the more predictable the decision. Forexample, a Shannon entropy is zero when an outcome isinescapable (probability ¼ 1). � The splits allowed the identification of variables taken

into account by patients to determine their PS level, astheir PS level is different depending on whether theconsidered variable is below or above a given split value.In the case of FDMA, these values are fuzzy in order tosimulate the process of human reasoning, and to mirrorthe non-explicit nature of decision-making.17,19

The set of patients was split according to the values ofeach variable into as many subsets as values for thesevariables. For instance, if the variable ‘‘Short course of oralcorticosteroids during the last 12 months’’ was selected, itled to two subsets: the subset of patients with ‘‘Short courseof oral corticosteroids during the last 12 months ¼ yes’’ andthe subset of patients with ‘‘Short course of oral corticos-teroids during the last 12 months ¼ no’’. For each subset,two exclusive situations could occur: (i) if all patients fromthe subset (or at least a significant majority of them) are ofthe same PS level, the variable is then identified as used bypatients to determine the PS level, (ii) if not, the process ofsplitting is repeated on this subset. To do this, the Shannonentropy is calculated allowing identification and ranking ofthe variables according to their predictive value. The bestdecision-making process uses the minimum amount ofvariables needed by patients to determine their PS

ARTICLE IN PRESS

Patients’ perception of asthma severity 2151

level.17,52,53 The splitting values for continuous variableswere determined by a similar process.53

Cross-validation

In statistics, an a priori hypothesis is made before perform-ing an analysis. The ‘‘sample size’’ and ‘‘statistical power’’depend on the work hypothesis. On the contrary, FDMA doesnot depend on an a priori hypothesis on the way patientsdetermine the severity of their asthma. The greater thenumber of patients, the better (i.e., the more robust) themodel developed by the FDMA.

The model we obtained was validated by means of cross-validation. This evaluates the accuracy and the robustness ofthe model.54 The accuracy of the model to correctly predictPS was tested by a so-called four-fold cross-validation test.The whole set of patients was divided into four distinctsubsets, S1, S2, S3 and S4. Three of these subsets weremerged to perform a FDMA. The PS level of each patient fromthe remaining subset was compared to that predicted by theFDMA performed with the three merged subsets. Thepercentage of good PS prediction measures the accuracy ofthe FDMA. The greater the accuracy, the more reliable theidentification and ranking of the variables patients may use todetermine their PS level. We aimed to perform a FDMA withaccuracy greater than 60%. This process was conducted fourtimes. Thus, four accuracies were obtained: one from testingS4 with the model generated from S1+S2+S3, one from testingS3 with the model generated from S1+S2+S4, another fromtesting S2 with the model generated from S1+S3+S4 andfinally, one from testing S1 with the model generated fromS2+S3+S4. The statistical mean of these four accuracieswas used to assess the accuracy of the FDMA performedwith the whole set of patients. The standard deviation (SD) ofthis mean was used to assess the robustness of the model(the lower the SD, the more robust the model). The morerobust the model, the more confident we could be with theaccuracy of the FDMA. Analyses were carried out using theSalammbo software (LIP6, Paris, France).52

Several fuzzy models were considered for this study:

�

Models that classify patients, e.g., the fuzzy neuralnetwork model. Knowing values of all variables related toa given patient enables the determination of their PSlevel. The Neural network models (also called ‘‘blackboxes’’) do not offer a fully interpretable model. Theyclassify patients according to their PS level, but do notidentify nor rank variables. Such a model could not beuseful in our study. � Models that identify groups of patients. These models

enable the identification of groups of patients withsimilar PS levels according to the values of variables.J Fuzzy clustering based approaches. The way of

measuring such a similarity between patients is notpossible without knowing the predictive value (i.e.,the informative power) of each variable and theircombination. As this was the aim of our study, wecould not use this model.

J Bottom-up based ‘‘ifythen’’ rule algorithms. Thesemodels build rules by merging patients one by one

until no more merging is possible, thus forming thegroups of patients. However, in the literature, thismodel is generally only empirically explained, whichmakes the interpretation of the rule constructiondifficult. Consequently, it is more difficult to explainthe resulting sets, and we decided that such a modelcould not be used in our study.

J The top-down based ‘‘ifythen’’ rule algorithms. Thisis the FDMA model we used, as described above. Itsmain advantage is that the interpretation of this FDMAdoes not require specific knowledge in computerscience. One only has to understand the significanceof variables used by patients. The variables identifiedare ranked according to their importance in thedecision-making process. It is this relative simplicitythat explains our choice of this FDMA. In addition, thenovelty of our study relies on the use of the FDMA thatsimulates the process of human reasoning, mirroringthe non-explicit nature of decision-making. FDMA canbe applied in various other domains in order to assesspatients’ decision-making processing or behaviors.

References

1. National Heart Lung and Blood Institute. National Institutes ofHealth. Global initiative for asthma; global strategy for asthmamanagement and prevention. NHLBI/WHO workshop report(02-3659), 2003. p. 1–176.

2. National Heart Lung and Blood Institute. National Institutes ofHealth. Global initiative for asthma; global strategy for asthmamanagement and prevention. NHLBI/WHO workshop report(02-3659), Updated 2005 from the 2004 document, 2005.p. 1–200.

3. Apter AJ, ZuWallack RL, Clive J. Common measures of asthmaseverity lack association for describing its clinical course.J Allergy Clin Immunol 1994;94:732–7.

4. Colice GL, Burgt JV, Song J, Stampone P, Thompson PJ.Categorizing asthma severity. Am J Respir Crit Care Med 1999;160:1962–7.

5. Kendrick AH, Higgs CM, Whitfield MJ, Laszlo G. Accuracy ofperception of severity of asthma: patients treated in generalpractice. Brit Med J 1993;307:422–4.

6. Moy ML, Israel E, Weiss ST, Juniper EF, Dube L, Drazen JM.Clinical predictors of health-related quality of life depend onasthma severity. Am J Respir Crit Care Med 2001;163:924–9.

7. Colice GL. Categorizing asthma severity: an overview ofnational guidelines. Clin Med Res 2004;2:155–63.

8. Miller MK, Johnson C, Miller DP, Deniz Y, Bleecker ER, Wenzel SE.Severity assessment in asthma: an evolving concept. J AllergyClin Immunol 2005;116:990–5.

9. Stoloff SW, Boushey HA. Severity, control, and responsiveness inasthma. J Allergy Clin Immunol 2006;117:544–8.

10. Van Ganse E, Antonicelli L, Zhang Q, Laforest L, Yin DD, NoceaG, et al. Asthma-related resource use and cost by GINAclassification of severity in three European countries. RespirMed 2006;100:140–7.

11. Yawn BP, van der Molen T, Humbert M. Asthma management: areGINA guidelines appropriate for daily clinical practice? PrimCare Respir J 2005;14:294–302.

12. Yawn BP, Yawn RA. Measuring asthma quality in primary care:can we develop better measures? Respir Med 2006;100:26–33.

13. Schatz M, Mosen D, Apter AJ, Zeiger RS, Vollmer WM, Stibolt TB,et al. Relationships among quality of life, severity, and controlmeasures in asthma: an evaluation using factor analysis.J Allergy Clin Immunol 2005;115:1049–55.

ARTICLE IN PRESS

A. Lurie et al.2152

14. Yawn BP, Fryer GE, Lanier D. Asthma severity: the patient’sperspective. J Asthma 2004;41:623–30.

15. Rabe KF, Adachi M, Lai CK, Soriano JB, Vermeire PA, Weiss KB,et al. Worldwide severity and control of asthma in children andadults: the global asthma insights and reality surveys. J AllergyClin Immunol 2004;114:40–7.

16. Toren K, Palmqvist M, Lowhagen O, Balder B, Tunsater A. Self-reported asthma was biased in relation to disease severity whilereported year of asthma onset was accurate. J Clin Epidemiol2006;59:90–3.

17. Bouchon-Meunier B. Uncertainty management in medicalapplications. In: Akay M, editor. Nonlinear biomedical signalprocessing fuzzy logic, neural networks, and new algorithms.New York: Wiley; 2000. p. 5–11.

18. Hanson CW3rd, Marshall BE. Artificial intelligence applica-tions in the intensive care unit. Crit Care Med 2001;29:427–35.

19. Bates JH, Young MP. Applying fuzzy logic to medical decisionmaking in the intensive care unit. Am J Respir Crit Care Med2003;167:948–52.

20. Bates JH, Hatzakis GE, Olivenstein R. Fuzzy logic and mechan-ical ventilation. Respir Care Clin N Am 2001;7:363–77.

21. Schuster A, Adamson K, Bell DA. Fuzzy logic in a decisionsupport system in the domain of coronary heart disease riskassessment. In: Barro S, Marin R, editors. Fuzzy logic inmedicine. Heidelberg: Springer; 2001. p. 177–204.

22. Zhang XS, Huang JW, Roy RJ. Depth of anesthesia control withfuzzy logic. In: Barro S, Marin R, editors. Fuzzy logic inmedicine. Heidelberg: Springer; 2001. p. 83–112.

23. Zadeh LA. ‘‘Fuzzy’’ sets. Inform Control 1965;8:338–53.24. Hampel U, Schleicher E, Freyer R. Volume image reconstruction

for diffuse optical tomography. Appl Opt 2002;41:3816–26.25. Hsu CH. Fuzzy logic automatic control of the Phoenix-7 total

artificial heart. J Artif Organs 2004;7:69–76.26. Kitsantas P, Hollander M, Li L. Using classification trees to

assess low birth weight outcomes. Artif Intell Med 2006;38:275–89.

27. Perner P. Intelligent data analysis in medicine-recent advances.Artif Intell Med 2006;37:1–5.

28. Marks GB, Dunn SM, Woolcock AJ. An evaluation of an asthmaquality of life questionnaire as a measure of change in adultswith asthma. J Clin Epidemiol 1993;46:1103–11.

29. Lurie A, Guillemin F, Djaballah K, Marks G, Woolcock A, DusserD. Quality of life in ambulatory care patients with asthma:validation and reliability of the French version of the Universityof Sydney asthma quality of life questionnaire [abstract]. Am JRespir Crit Care Med 1997;155:A721.

30. Redier H, Daures JP, Michel C, Proudhon H, Vervloet D, CharpinD, et al. Assessment of the severity of asthma by an expertsystem. Description and evaluation. Am J Respir Crit Care Med1995;151:345–52.

31. National Institutes of Health. National Heart Lung and BloodInstitute. Guidelines for the diagnosis and the management ofasthma. Expert Panel Report II, Bethesda, MD. Public HealthService; 1997, Publication No. 97-4051.

32. Boulet LP, Becker A, Berube D, Beveridge R, Ernst P. Canadianasthma consensus report, 1999. Canadian asthma consensusgroup. Can Med Assoc J 1999;161:S1–S61.

33. Baker KM, Brand DA, Hen Jr. J. Classifying asthma: disagree-ment among specialists. Chest 2003;124:2156–63.

34. Colice GL. The seduction of asthma severity categorization.Chest 2003;124:2054–6.

35. Clancy K. British guidelines on the management of asthma.Thorax 2004;59:81–2.

36. Barley EA, Jones PW. A comparison of global questions versushealth status questionnaires as measures of the severity andimpact of asthma. Eur Respir J 1999;14:591–6.

37. Wijnhoven HA, Kriegsman DM, Hesselink AE, Penninx BW, deHaan M. Determinants of different dimensions of diseaseseverity in asthma and COPD : pulmonary function and health-related quality of life. Chest 2001;119:1034–42.

38. Juniper EF, Guyatt GH, Ferrie PJ, Griffith LE. Measuring qualityof life in asthma. Am Rev Respir Dis 1993;147:832–8.

39. Juniper EF, Johnston PR, Borkhoff CM, Guyatt GH, Boulet LP,Haukioja A. Quality of life in asthma clinical trials: comparisonof salmeterol and salbutamol. Am J Respir Crit Care Med 1995;151:66–70.

40. Leroyer C, Lebrun T, Proust A, Lenne X, Lucas E, Rio G, et al.Knowledge, self-management, compliance and quality of life inasthma: a cross-sectional study of the French version of theasthma quality of life questionnaire. Qual Life Res 1998;7:267–72.

41. Benayoun L, Druilhe A, Dombret MC, Aubier M, Pretolani M.Airway structural alterations selectively associated with severeasthma. Am J Respir Crit Care Med 2003;167:1360–8.

42. Zeiger RS, Dawson C, Weiss S. Relationships between duration ofasthma and asthma severity among children in the childhoodasthma management program (CAMP). J Allergy Clin Immunol1999;103:376–87.

43. Fuhlbrigge AL, Adams RJ, Guilbert TW, Grant E, Lozano P,Janson SL, et al. The burden of asthma in the United States:level and distribution are dependent on interpretation of thenational asthma education and prevention program guidelines.Am J Respir Crit Care Med 2002;166:1044–9.

44. Osman LM, McKenzie L, Cairns J, Friend JA, Godden DJ, LeggeJS, et al. Patient weighting of importance of asthma symptoms.Thorax 2001;56:138–42.

45. Teeter JG, Bleecker ER. Relationship between airway obstruc-tion and respiratory symptoms in adult asthmatics. Chest1998;113:272–7.

46. Combescure C, Chanez P, Saint-Pierre P, Daures JP, Proudhon H,Godard P. Assessment of variations in control of asthma overtime. Eur Respir J 2003;22:298–304.

47. Juniper EF, O’Byrne PM, Guyatt GH, Ferrie PJ, King DR.Development and validation of a questionnaire to measureasthma control. Eur Respir J 1999;14:902–7.

48. Juniper EF, Wisniewski ME, Cox FM, Emmett AH, Nielsen KE,O’Byrne PM. Relationship between quality of life and clinicalstatus in asthma: a factor analysis. Eur Respir J 2004;23:287–91.

49. Osborne ML, Vollmer WM, Linton KL, Buist AS. Characteristics ofpatients with asthma within a large HMO: a comparison by ageand gender. Am J Respir Crit Care Med 1998;157:123–8.

50. Quinlan JR. Induction of decision trees. Mach Learn 1986;1:86–106.

51. Shannon CE. The mathematical theory of communication.Urbana: University of Illinois Press; 1948.

52. Bouchon-Meunier B, Marsala C. Learning fuzzy decision rules.In: Bezdek J, Dubois D, Prade H, editors. Fuzzy sets inapproximate reasoning and information systems hand book offuzzy sets. Norwell: Kluwer Academic; 1999. p. 279–304.

53. Marsala C. Fuzzy partitioning methods. In: Pedrycz W, editor.Granular computing: an emerging paradigm: studies in fuzzi-ness and soft computing. New York: Springer; 2001. p. 163–86.

54. Kohavi R. A study of cross-validation and bootstrap for accuracyestimation and model selection. In: Proceeding of the fourteeninternational joint conference on artificial intelligence 1995,San Matteo, CA: Morgan Kaufmann; 1995. p. 1137–43.