Tidal breathing flow-volume loop for early asthma identification in young children: An uncertainty-aware machine learning approach

Conventional tidal breathing parameters are often insensitive to mild early airway obstruction. In addition, existing artificial intelligence (AI) models may generate overconfident errors in clinically heterogeneous cases.

This retrospective study (2022-2025) included 1121 children aged 1-3 years from West China Second University Hospital, Sichuan University. Data from the Jinjiang campus were used for model development (n = 805) and data from the Huaxi campus served as an independent spatial validation cohort (n = 316). We extracted novel geometric morphological and derived functional features from TBFVL and used machine learning model to evaluate a conventional-feature baseline model versus an enhanced model incorporating novel features. We further proposed and evaluated the Adaptive Conformal Three-way Decision framework (ACT-Decision) to separate high-uncertainty cases from automatic decision-making.

The final feature set comprised 20 newly derived features, 8 conventional parameters and 1 demographic variable. In the Huaxi validation cohort, the enhanced model using combined features outperformed the baseline model using only conventional features, achieving an accuracy of 0.788 (95% CI: 0.741, 0.829), ROC-AUC of 0.867 (95% CI: 0.827, 0.902) and F1-score of 0.798 (95% CI: 0.744, 0.841). SHAP analysis confirmed contributions from both conventional and newly derived features. In the Huaxi test subset, ACT-Decision automatically classified 74.78% of children with an accuracy of 0.849 (95% CI: 0.787, 0.896).

TBFVL-derived features improved discrimination between children aged 1-3 years with clinically diagnosed asthma and healthy controls. ACT-Decision framework enabled selective deferral of higher-uncertainty predictions for clinician review.