Diagnostic accuracy of single variables used for predicting BHR
We created ROC curves to evaluate the ability of many of the variables to predict positive BHR. In the spirometry measurements, the 2 largest AUCs for a positive BHR diagnosis were 0.763 for FEF25%-75% and 0.762 for FEF50%(Table 3). The FEV1/FVC, EOS%, EOS, and IOS measurements did not give high AUCs for positive BHR diagnosis.
In patients with chest tightness, the AUCs for positive BHR diagnosis were FEF50% 0.751 (95% CI, 0.637 - 0.864), FEF75% 0.812 (95% CI, 0.708 - 0.916), FEF25%-75% 0.763 (95% CI, 0.651 - 0.875), FENO 0.731 (95% CI, 0.607 - 0.855), and EOS counts 0.706 (95% CI, 0.580 - 0.832).
Diagnostic accuracy of small-airway function variables combined with FENO andinternal cross-validation of the final models
To determine whether combining measurements would improve BHR prediction, we repeated the ROC analyses for spirometry measurements combined with FENO. The AUC of FEF50% combined with FENO was 0.845 (95% CI, 0.812-0.878), which was significantly higher than the AUC of univariate FEF50% (P < .001) (Table 4 and Figure 1). Similarly, the other spirometry measurements also had higher AUCs when combined with FENO than they did alone (Table 4) NPV was ≥85.45% for all of the combinations.
We then transformed the continuous variables into binary variables according to the cut-off values shown in Tables and reanalyzed the mentioned above ROC curves. The AUCs of FEF50% and FEF25%-75% combined with FENO remained high (Figure 2).
In patients with chest tightness, the AUCs of FEF50%, FEF75%, and FEF25%-75% combined with FENO were 0.880 (95% CI, 0.806-0.954), 0.892 (95% CI, 0.812-0.972), and 0.884 (95% CI, 0.805-0.934), respectively (Table 5).
The error rates between the average AUC of 5 different cross-validation models and the whole-model AUC using the entire data set were lower than 0.05 for all chosen variables, indicating that the data model has stable predictive ability for different data sets.