Pilot screening multivariable prediction model for late diagnosed cystic fibrosis

Abstract Background Diagnosing cystic fibrosis (CF) in adults is challenging. Patients often present with atypical and varied symptoms. The aim of this study is to develop a pilot prediction model to aid clinicians in identifying patients who should undergo appropriate CF diagnostic testing. Methods This is a retrospective, single-center, observational cohort study of adult (≥18 yrs) patients who underwent sweat testing for evaluation of clinically suspected CF. Dataset 1 includes 171 patients with CF-supportive conditions from 2005-2021 and was used both to develop the screening model and to subject it to internal validation. Dataset 2 includes 87 patients with CF-supportive conditions from 2021-2023 and was used to subject the prediction model to external validation. Based on CF Foundation guidelines, dataset patients were each assigned one of three CF classifications: (1) CF likely (Type A ), (2) CF not resolved (Type B ), or (3) CF unlikely (Type C ). Clinical predictor variables indicate the presence (value=1) or absence (value=0) of 7 CF-related clinically observable patient conditions. The screening model decision rule recommends a patient for CF testing if either of two conditions holds: (1) A numerical patient score S A calculated using multivariate logistic regression (LR) equals or exceeds a threshold value C 0 determined using ROC curve analysis, or (2) the patient diagnosis is positive for bronchiectasis/tree-in-bud and negative or unknown for each of 6 other clinical conditions. Results For each dataset, the observed sensitivity value for recommending CF testing was 1.00 for patients that CF testing classified as Type A or Type B . Further, for the two datasets in combination, 2/128=1.56% of the Type C patients were not recommended for CF testing. Conclusions When applied separately to each dataset, the screening model matched the predictive accuracy of clinical judgment both for Type A and Type B patients. A similar predictive accuracy comparison for Type C patients is not possible because the CF types are unknown for patients not recommended for CF testing.