Machine learning-enhanced HRCT analysis for diagnosis and severity assessment in pediatric asthma

Abstract

Objectives: Chest high-resolution computed tomography (HRCT) is conditionally recommended to rule out conditions that mimic or coexist with severe asthma in children. However, it may provide valuable insights into identifying structural airway changes in pediatric patients. This study aims to develop a machine learning-based chest HRCT image analysis model to aid pediatric pulmonologists in identifying features of severe asthma.

Methods: This retrospective case-control study compared children with severe asthma (as defined by ERS/ATS guidelines) to age- and sex-matched controls without asthma, using chest HRCT scans for detailed imaging analysis. Statistical analysis included classification trees, random forests, and conventional ROC analysis to identify the most significant imaging features that mark severe asthma from controls.

Results: Chest HRCT scans differentiated children with severe asthma from controls. Compared to controls (n = 21, mean age 11.4 years), children with severe asthma (n = 20, mean age 10.4 years) showed significantly greater bronchial thickening (BT) scores (p < 0.001), airway wall thickness percentage (AWT%, p < 0.001), bronchiectasis grading (BG) and bronchiectasis severity (BS) scores (p = 0.016), mucus plugging, and centrilobular emphysema (p = 0.009). Using AWT% as the predictor in conventional ROC analysis, an AWT% ≥ 38.6 emerged as the optimal classifier for discriminating severe asthmatics from controls, with 95% sensitivity, specificity, and overall accuracy.

Conclusion: Our study demonstrates the potential of machine learning-based analysis of chest HRCT scans to accurately identify features associated with severe asthma in children, enhancing diagnostic evaluation and contributing to the development of more targeted treatment approaches.

Keywords: Children; artificial intelligence; chest high‐resolution computed tomography; machine learning; severe asthma.

Figure 1

Best regression tree (according to the cross‐validated accuracy) using the disease status as the outcome and radiological findings as the predictors. AWT%: airway wall thickness percentage RUL: Right upper lobe.

Figure 2

Normalized permutation importance of radiological findings for the prediction of the disease status using the best random forest (according to the cross‐validated accuracy). Dashed line: minimum value associated with a significant adjusted p‐value. AWT%: airway wall thickness percentage BT: Bronchial thickening ML: Middle lobe LLL: Left lower lobe RLL: Right lower lobe L: Lingula LUL: Left upper lobe RUL: Right upper lobe BG: Bronchiectasis grading MP: Mucus plugging BS: Bronchiectasis severity ASC: Air space consolidation BA: Bronchial‐arterial AT: Air trapping IDLA: Inspiratory decreased lung attenuation GG: Ground glass SCO: Centrilobular opacities.

Figure 3

Receiver operating characteristic (ROC) curve for the ability of AWT% to predict the disease status. The best cut‐off (point closest to the top left corner) and the area under the ROC curve (AUC) are superimposed.