Abstract / Summary
Abstract Background Small airway dysfunction (SAD) is an early stage of chronic obstructive pulmonary disease (COPD), but conventional spirometry lacks sensitivity for its detection. Impulse oscillometry (IOS) can assess respiratory mechanics without effort dependence; however, the diagnostic performance of single IOS parameters is limited. This study aimed to evaluate the value of IOS parameters in early SAD screening and to construct a prediction model based on IOS parameters using machine learning algorithms. Methods This cross-sectional study included 237 subjects (144 healthy controls, 44 SAD patients, and 49 COPD patients). All participants underwent IOS and spirometry. Analysis of covariance (ANCOVA) adjusted for age, sex, and body mass index was used to compare IOS parameters among groups. The Jonckheere-Terpstra test assessed monotonic trends across disease stages. Spearman partial correlation analyzed relationships between IOS parameters and small airway indices (MMEF%, FEF50%, and FEF75%). Within the COPD subgroup, GOLD grade comparisons and multiple linear regression were performed. Machine learning models based on IOS parameters were developed for SAD screening, and their diagnostic performance was evaluated using receiver operating characteristic curves. Results After adjusting for confounders, R5-R20, (R5-R20)/R5, Fres, and AX differed significantly among the three groups (healthy controls, SAD patients, and COPD patients) (all P < 0.001), with significant monotonic increasing trends from healthy to SAD to COPD (all P < 0.001). In the SAD stage, AX (|Z| = 0.76) and Fres (|Z| = 0.97) already showed marked deviations from normal, while FEV1% remained at the borderline of normal (|Z| = 0.64). Within the COPD subgroup, AX showed a negative correlation with FEV1% ( r = -0.383, P = 0.007). For distinguishing healthy from SAD, Fres showed the highest AUC (0.692, sensitivity 65.9%, specificity 66.7%), while AX demonstrated higher sensitivity (70.5%, specificity 58.3%, AUC 0.662) at its optimal cutoff. Among the machine learning models, AdaBoost achieved the highest AUC (0.855), with a sensitivity of 76.9% and specificity of 84.1%, while SVM showed comparable performance (AUC 0.843, sensitivity 76.9%, specificity 88.6%). All machine learning models outperformed the single AX parameter (AUC = 0.662) in terms of AUC. Conclusions IOS parameters sensitively reflect the progressive deterioration of small airway function. AX demonstrated favorable sensitivity for early SAD screening, while Fres showed the highest overall diagnostic accuracy. The machine learning model based on IOS parameters achieved good diagnostic performance for SAD screening. The AdaBoost model demonstrated the best overall performance, while SVM showed similar results. These findings suggest that IOS parameters alone, particularly Fres, AX, and R5-R20, may serve as sensitive, non-invasive markers for SAD detection, and were identified as the key physiological drivers of model performance. Clinical trial number not applicable.