Abstract

Based on available resources, diagnostic tests are used in a stepwise approach starting with a detailed medical and occupational history; assessment of nonspecific bronchial hyperresponsiveness (NSBHR); and immunological sensitization with skin-prick tests (SPT) or specific immunoglobulin E to workplace agent when available. The second step includes serial assessments of NSBHR and peak expiratory flow at work and off-work, or specific inhalation challenge (SIC). As the reference test for diagnosing OA, SIC is only available in a few centers around the world. Therefore, Suarthana and colleagues developed non–SIC-based diagnostic models for OA using Quebec data of 160 symptomatic subjects who completed an SIC procedure (ie, development set).These subjects were exposed to high-molecular-weight (HMW) protein agents such as flour, laboratory animal allergens, and latex. OA was defined as a positive SIC, namely a sustained fall in forced expiratory volume in one second (FEV1) >20% from baseline value after exposure to the suspected occupational agent.
The final predictive model included age (>40 vs. =40 years); agent type (flour vs. other HMW agents); the presence of work-related rhinoconjunctivitis (yes/no); inhaled corticosteroid use (yes/no); positive SPT reaction to the specific occupational agent(s) (yes/no); and the presence of NSBHR (defined as the provocative concentration of histamine/methacholine causing a 20% decrease in FEV1 (PC20) =16 mg/mL). This final model demonstrated a good accuracy and internal validity. To facilitate its use, it was transformed into clinical scores equipped with the corresponding predicted probability of OA. Sum scores = 150 with a corresponding predicted probability of OA > 0.25 were recommended for referral cutoff with a sensitivity of 96%, specificity 59%, positive predicted value 72%, and negative predicted value 94%.
To externally validate the model, first, we calculated individual probabilities in European data using the equation from the Canadian model without any adjustments (no update method). The model demonstrated good discrimination in European data although lower than in Canadian (AUC 0.83 vs. 0.89, respectively). The Brier score was also higher in European data (0.171 vs. 0.121), but the model was still informative (ie, the maximum score for a noninformative model in a population with a 57.5% prevalence of OA was 0.24). Second, we recalibrated the intercept of the model. Finally, the intercept and coefficients were re-estimated (refitting method). Both methods did not improve the discriminative ability of the model: The AUC remained the same, although the Brier score slightly improved.

Full Text

Comments