INTRODUCTION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Approximately 50% of asthmatic subjects develop a late asthmatic response (LAR) in response to inhaled allergen in the laboratory setting (1). The mechanisms involved in the LAR are incompletely understood, but recent studies have revealed important roles for IgE (2) and leukotriene D₄ (3, 4)possibly as orchestrators of the cellular infiltration, edema, bronchospasm, and mucus hypersecretion that contribute to airway narrowing. Clinical investigators often use the allergen challenge test to investigate mechanisms of allergic airway inflammation and to assess new asthma treatments, because the attenuation of the LAR has predicted efficacy in subsequent large clinical trials (5). In clinical experiments that require asthmatic subjects to have a LAR, subjects must be monitored closely for 7 to 10 h after allergen challenge to identify those with a LAR. This is time-consuming, labor-intensive, and expensive. Previous studies that have examined predictors of LAR in asthmatic subjects have not studied large numbers of subjects and have yielded conflicting results about predictors (6) so that no baseline asthma variable is used widely to identify subjects likely to have a LAR.

The aim of this study was to determine whether the development of the LAR in asthmatic subjects could be predicted by examining baseline asthma characterization variables such as FEV₁, provocative concentration of methacholine causing a 20% reduction in FEV₁ (PC₂₀), eosinophil percent in induced sputum, and allergen skin test reactivity. We also examined the predictive value of the allergen-induced decline in FEV₁ at 20 min; this FEV₁ value is of practical utility, because it is the value that is used by investigators to decide whether or not to continue dosing the subject with allergen. To accomplish our aims, we examined a database of 60 asthmatic subjects who had undergone allergen testing in our laboratory over the past 5 yr. We applied a variety of logistic regression techniques and classification tree approaches to analyze the data.

	METHODS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Subjects

We reviewed the data from 60 asthmatic subjects who have undergone allergen challenges for different studies in our laboratory over the past 5 yr (Table 1). All studies had been approved by the Committee of Human Research at the University of California, San Francisco, and all subjects had signed approved informed consent forms. Inclusion criteria for these studies were similar and can be summarized as follows: FEV₁  65% of predicted normal (14), bronchial hyperreactivity to methacholine (PC₂₀  8 mg/ml), and a positive allergen skin prick test to at least one of a panel of aeroallergens. Exclusion criteria were use of any topical or systemic corticosteroid or symptoms of an upper respiratory infection in the previous 6 wk, tobacco use in the previous year (or  10 pack-years lifetime smoking), or a history of significant medical illness other than asthma.

Protocol

Skin reactivity to house dust mites (Dermatophagoides pteronyssinus, D. farinae), cat pelt, ryegrass (Lolium perenne), negative control (Bayer Corporation, Elkhart, IN), and histamine base 1 mg/ml (Histatrol; Center Lab, Port Washington, NY) had been assessed as previously described (15). The allergen provoking the largest wheal was used for skin prick test titration using 2-fold dilutions to determine the smallest concentration causing a positive reaction ( 2 mm wheal with erythema). Bronchial reactivity to methacholine was determined by the dosimeter method. A nebulizer 646 (DeVilbiss, Somerset, PA) was connected to a breath-activated dosimeter whose breath sensor channel has a T piece with a 5-mm orifice which was connected to the nebulizer 646 to limit inspiratory flow to  0.5 L/s (DSM-2; S&M Instrument Company, Doylestown, PA). For each concentration, the subject slowly inhales 5 breaths from functional residual capacity to total lung capacity (mean 5.4 microliter/inhalation at 20 psi and 0.6 s) and then performs spirometry 3 min later. After initial diluent challenge, increasing doubling doses of methacholine (from 0.031 mg/ml to 16 mg/ ml) are delivered every 5 min until FEV₁ falls  20% from postdiluent value. The PC₂₀ is then calculated by semi-log interpolation.

Sputum induction processing and analysis were performed as previously described using an UltraNeb 99 (DeVilbiss) for aerosolization of 3% saline (16). The percentage of eosinophils had been calculated as the percentage of nonsquamous cells in a count of at least 500 nonsquamous cells.

In all protocols allergen challenge had been performed similarly, as previously described (2). In brief, after diluent challenge (calcium- and magnesium-free phosphate buffer saline), allergen dosing started at 4 doubling doses below the concentration predicted to cause a 20% fall in FEV₁ (17). Allergen was delivered using a DeVilbiss 646 nebulizer connected to a breath-activated dosimeter as explained previously (DSM-2; S&M Instrument Company) which delivered an average of 17.5 µl per inhalation (20 psi, 1.6 s). Increasing allergen concentrations were delivered in doubling doses at 12-min intervals; spirometry was performed 10 min after each dose. When the FEV₁ fell between 15% and 20%, FEV₁ was measured every 10 min until it was stable or improving before the next dose was administered. The challenge was stopped when the FEV₁ fell by  20% from postdiluent value. The FEV₁ that led to the decision to stop the challenge was used as a predictor variable (fall in FEV₁ at 20 min). Then, FEV₁ was measured at 20, 30, 45, 60, 90, and 120 min and hourly until 7 h postchallenge. The maximal percent decline in FEV₁ from postdiluent value during the early phase (0 to 2 h) and late phase (3 to 7 h) were recorded.

Statistical Analyses

The following variables were chosen a priori as predictor variables of LAR: age, sex, baseline percent predicted FEV₁ and FEF_25-75%, allergen skin test wheal and flare diameters, methacholine PC₂₀, percent reversibility of FEV₁ after bronchodilator, percent eosinophils in induced sputum, percent decline in FEV₁ after induced sputum, allergen PC₂₀, dose of allergen delivered during challenge (ratio between allergen PD₂₀ over skin test titration expressed as doubling doses), and maximal percent decline in FEV₁ at 20 min of the last allergen dose (used as a surrogate for early asthmatic response [EAR]). In modeling the binary outcome LAR, we used both logistic regression (SAS; SAS Institute Inc., Cary, NC, 1997) and classification tree (S-Plus; StatSci, a division of MathSoft, Inc., Seattle, 1995) analyses. For logistic regression, initial exploratory analyses were used to elicit variable functional form. Subsequently, a variety of stepwise and subset selection strategies were used to arrive at a parsimonious model, from which variable associations with LAR were appraised.

Classification trees (18) are designed to extract patient subgroups that are homogeneous with respect to both outcome and predictor variables (19, 20). This is accomplished by recursively partitioning the data such that at each stage the variable (and its associated cut-point) that best subdivides the data (in terms of optimizing homogeneity) is determined. Cross validation, or an independent test sample within the database, is then used to assess how many such divisions to adopt.

	RESULTS

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

Thirty-four subjects (57%) developed a LAR. The allergen-induced decline in FEV₁ at 20 min in FEV₁ was highly correlated with the EAR (EAR = maximal percent fall in FEV₁ in the first 2 h after allergen challenge, r = 0.87, p < 0.0001 by Spearman rank correlation test).

Logistic regression using forward, backward, or bi-directional approaches for stepwise variable selection, consistently selected or retained methacholine PC₂₀ and the decrease in FEV₁ at 20 min as the only variables associated with the LAR. Surprisingly, baseline percent predicted FEV₁, percent reversibility of FEV₁ after bronchodilator, percent eosinophils in induced sputum, skin test reaction, and dose of allergen in relation to skin test titration (in doubling doses from the skin test titration result to the last dilution of allergen given in the allergen challenge) were not associated with the presence of LAR (Table 2).

The results of classification tree analysis using binary LAR as the outcome variable are presented in Figure 1 (absence of LAR if < 15% decline, or presence of LAR if  15% decline in FEV₁ between 3 and 7 h after allergen challenge). For PC₂₀ the classification tree analysis yielded a split point (threshold) of 0.25 mg/ml. This produced two subgroups with respective LAR incidences of 87% (PC₂₀  0.25 mg/ml) and 38% (PC₂₀ > 0.25 mg/ml). This latter subgroup was further partitioned on the basis of the data for allergen-induced decline in FEV₁ at 20 min, and the classification tree analysis yielded a split point (threshold) of 27%. The resultant groups had LAR incidence rates of 13% (FEV₁ at 20 min  27%) and 57% (FEV₁ at 20 min > 27%). Similar results in terms of split variables and thresholds were obtained using LAR as a continuous outcome (data not shown).

View larger version (15K): [in this window] [in a new window]   Figure 1.   Classification tree analysis of predictors of LAR showing cutoff values for PC20 methacholine (Met-PC20) and percent fall in FEV1 at 20 min after allergen challenge. The first split is performed on PC20 methacholine with a split point (threshold) of 0.25 mg/ml. This produces two subgroups with respective LAR incidences of 87% and 38%. This latter subgroup is further partitioned on the basis of the immediate decrease in FEV1 with a split point of 27%. The resultant groups have LAR incidence rates of 13% and 57%.

	DISCUSSION

TOP ABSTRACT INTRODUCTION METHODS RESULTS DISCUSSION REFERENCES

The main findings of our study are that very low values for PC₂₀ methacholine and large declines in FEV₁ at 20 min of allergen challenge predict the development of a late asthmatic response to inhaled allergen in asthmatic subjects. This information can be used to improve the efficiency of screening for subjects likely to have a LAR in allergen challenge studies.

Of the 60 asthmatic subjects who underwent aerosolized allergen challenge in our laboratory, 57% developed a LAR; this percentage is consistent with data from previous studies (1). Using logistic regression with LAR as a binary outcome, the variables most predictive of LAR were methacholine PC₂₀ and the decline in FEV₁ at 20 min of allergen challenge. The same variables emerged from the classification tree analysis (see Figure 1) which provides a complementary approach to data analysis. Such tree-structured methods afford a flexible nonparametric tool for performing exploratory data analysis and constructing simple, interpretable classification schemes. The tree paradigm entails subdividing or splitting the sample into homogeneous subgroups. The divisions are based on the predictor variables and are formulated such that subjects with like predictor variable values are placed in the same subgroup. Thus, subgroups having prognostic significance are created. Such prognostic groupings are not readily extracted from logistic regression or discriminant analyses (20). Additionally, a computationally intensive search for optimal subdivisions is undertaken whereby every allowable binary subdivision of every covariate is examined before implementing a particular split. The whole procedure is then reinitiated on the two resultant subgroups. Allowable splits are those that preserve the ordering of ordered variables (for continuous or categorical variables) or are otherwise unrestricted (for unordered categorical variables). Details of this kind of analysis are provided elsewhere (18).

In the case of PC₂₀ methacholine, the classification tree analysis revealed that a PC₂₀ value of 0.25 mg/ml represents the best cutoff value in terms of increasing the yield of subjects with a LAR. Thus, the incidence of LAR was 87% in the subgroup of subjects whose PC₂₀ was  0.25 mg/ml and 38% in the subgroup whose PC₂₀ was > 0.25 mg/ml. These data indicate that the PC₂₀ could be used to improve the efficiency of screening asthmatic subjects for a LAR. For example, using traditional methods of screening that are not based on the results of baseline PC₂₀ (except to document hyperreactivity), it is necessary to screen 17 to 18 subjects to find 10 with a LAR. However, if subjects are first screened to document that they have a PC₂₀ methacholine  0.25 mg/ml, then it is only necessary to screen 11 to 12 asthmatics to find 10 with a LAR. These examples are based on data from a population of subjects with mild to moderate asthma who required rescue bronchodilators only for asthma control. In addition, although we used a standard method for measuring PC₂₀ methacholine, it is possible that studies which use a different protocol for measuring methacholine reactivity could identify a threshold value for PC₂₀ different from ours.

The classification tree paradigm was applied again to the two subgroups resulting from the split based on PC₂₀ methacholine. The next outcome tested was the decline in FEV₁ at 20 min after allergen challenge, because our analyses had shown that this outcome was associated with the development of a LAR. Here the classification tree analysis revealed that a threshold value of 27% represents the best cutoff value in terms of increasing the yield of subjects with a LAR when the PC₂₀ is > 0.25 mg/ml. Thus, for those subjects with PC₂₀ > 0.25 mg/ml, the incidence of LAR was only 13% if the decline in FEV₁ at 20 min was  27%, whereas the incidence of LAR increased to 57% if the decline was > 27%. Although it may be tempting to apply this result by targeting a decline in FEV₁ of  27% during allergen challenge, the risks of such an approach may outweigh any advantage in terms of yield of LAR.

Our results confirm and extend previous studies in which the development of a LAR in allergic asthmatic subjects has been found to be associated with increased bronchial responsiveness to histamine or methacholine (11, 12, 21) and with the early asthmatic response (9, 10). Although Machado and Stalenheim did not find a correlation between LAR and baseline bronchial reactivity to methacholine (8), this study included smokers and subjects with borderline hyperresponsiveness (6 of 24 subjects had methacholine PC₂₀ higher than 8 mg/ml). In addition, although Killian and coworkers did not find a significant correlation between EAR and LAR, their study had only nine subjects (11). We are unaware of any other study that has examined the relationship between LAR in asthma and baseline sputum eosinophilia. Our analysis showed no association between these two variables. We also found no significant association between LAR and baseline FEV₁.

In summary, we found that very low baseline values for PC₂₀ methacholine and large declines in FEV₁ at 20 min after allergen challenge are predictors of a LAR in asthmatic subjects. Our data suggest that investigators could select asthmatic subjects with a PC₂₀ < 0.25 mg/ml in order to improve the efficiency of identifying subjects with a LAR.