Application of machine learning to predict bronchiolitis severity in children: a single-centre retrospective cohort study.

Several severity scores have been developed to assess disease severity in infants with bronchiolitis, but they often lack objectivity and may not reliably reflect clinical outcomes. We performed a single-centre retrospective cohort study including infants younger than 12 months admitted with a clinical diagnosis of bronchiolitis over three epidemic seasons (2021-2024). Cluster analysis was applied to identify groups of infants with similar clinical patterns, which were subsequently mapped to severity levels. A machine learning (ML) model was then trained to predict these severity levels using anamnestic and clinical variables available at admission. For comparison, the bronchiolitis risk of admission score (BRAS) and the respiratory severity score-heart rate (RSS-HR) were also calculated at admission. A total of 112 infants were included, and three severity levels were identified. Mild bronchiolitis corresponded to infants not requiring NICU admission or supplementary ventilation. Moderate cases involved infants needing supplementary ventilation without NICU admission. Severe bronchiolitis was associated with both NICU admission and supplementary ventilation. BRAS showed only a weak association with the newly defined clinical severity score (τ = 0.354, p < 0.001), while RSS-HR showed no significant correlation. In contrast, the ML model demonstrated strong predictive performance (τ = 0.731, p < 0.001 in the test set). A simplified decision tree based on three admission variables-age, retractions, and oxygen saturation-also showed good predictive ability (τ = 0.654, p < 0.001).

conclusionsA new dynamic approach for assessing bronchiolitis severity at hospital admission, capable of robust and adaptable prediction across clinical settings, has been elaborated. WHAT IS KNOWN: • Several bronchiolitis severity scores exist, but many lack validation, are subjective, and have limited applicability across settings, making early and accurate severity prediction difficult. WHAT IS NEW: • Our study introduces a three-level, outcome-based severity score and a dynamic ML tool, plus a simplified ARS decision tree, all outperforming conventional scores in predicting severity at admission.