Publication Summary
Background: Respiratory symptoms are considered to provide limited diagnostic value in the assessment of exercise-induced bronchoconstriction (EIB) (Price et al. Resp Med 2016; 120, 36-43). Aim: To evaluate perceived dyspnoea in athletes applying machine learning methods. Method: Cross-sectional evaluation of athletes (n = 65; male: n = 39) reporting exertional symptoms. All athletes completed the Dyspnoea-12 questionnaire (D-12) and performed spirometry pre-and-post a eucapnic voluntary hyperpnoea challenge. D-12 (predictor variable: 1-12 and response variables: 0=nil; 1=mild; 2=moderate; 3=severe) were evaluated against objective evidence of EIB (≥15% fall in FEV1 cut-off) (Price et al. Am J Respiratory Crit Care Med 2016; 193:1178-1180) using partial least square correlation analysis (PLSCA). D-12 items considered to have diagnostic relevance were subsequently analysed using principal component analysis (PCA) coupled with receiver operating characteristics-area under the curve (ROC-AUC). Results: Eleven athletes (17%) were diagnosed with EIB (average fall in FEV1: -29 ± 10%). D-12 (physical +/- affective components) failed to differentiate between EIB+ and EIB- (p>0.05), however, PLSCA and PCA identified Q4 and Q11 as the most influential questions to rule-in a diagnosis: ROC-AUC = 75%; sensitivity = 73%; specificity = 73% (p Conclusions: Although D-12 should not be employed to confirm EIB, it is possible to quantify the diagnostic relevance of symptom-based questions in this context by applying machine learning algorithms. The development and validation of an EIB specific questionnaire to aid the assessment of athletes reporting breathing difficulty remains a priority.
