Predictive model for inhaled corticosteroid response in hospitalized asthma: implications for precision medicine.

Background: Inhaled corticosteroids (ICS) are effective in reducing symptoms and preventing exacerbations in asthma patients and are therefore used as the first-line therapy for asthma management. However, patients show varying responses to ICS treatment, and some patients may require treatment escalation to systemic corticosteroids (SCS) during the follow-up period. The present study aimed to develop an early prediction model for ICS responsiveness by identifying patients initially treated with ICS who later required treatment escalation to SCS. Additionally, we examined the association between clinical features and ICS responsiveness and provided evidence supporting the potential overuse of SCS. Methods: The study included 5,463 asthma patients (age ≥12 years) who received ICS as their initial treatment between 2006 and 2022 at Xinjiang Uygur Autonomous Region People's Hospital. A subset of patients underwent treatment escalation to SCS during the follow-up period. Baseline predictors, including demographic and clinical features collected at admission, were assessed to identify factors associated with treatment escalation requirements. The data from 2006 to 2018 and 2019 to 2022 were utilized as the derivation and independent temporal validation cohorts, respectively. Machine learning algorithms were used to identify predictive factors and develop the prediction model. Additionally, non-experimental causal inference techniques were applied to evaluate SCS overuse. Results: Of the 5,463 patients, 1,088 (19.9%) patients required treatment escalation to SCS after initial ICS treatment. A random forest model with 40 predictors demonstrated effective discrimination in predicting the necessity for treatment escalation, achieving an area under the curve value of 0.7483 [95% confidence interval (CI): 0.709-0.7876] in the internal validation set and 0.6941 (95% CI: 0.6651-0.7231) in the independent validation cohort. The model achieved sensitivities and specificities of 0.685 and 0.6623 for the internal validation set, and 0.6336 and 0.6153 for the independent validation cohort, respectively. Key predictors for treatment escalation included blood eosinophil counts, neutrophil counts, age, and lymphocyte counts. The analysis of matched sets from the independent validation cohort indicated that some false negatives (patients predicted not to require treatment escalation but who received SCS) potentially represented overuse of SCS, as these patients showed comparable clinical outcomes (e.g., mortality, intensive care unit transfer) despite incurring higher healthcare costs. Conclusions: The developed predictive model provides an accessible, cost-effective tool for the early identification of patients potentially requiring treatment escalation to SCS in a hospitalized setting. Additionally, the findings suggest a potential approach for addressing the overuse of SCS in asthma management.