AUTHOR=Makrinioti Heidi , Maggina Paraskevi , Lakoumentas John , Xepapadaki Paraskevi , Taka Stella , Megremis Spyridon , Manioudaki Maria , Johnston Sebastian L. , Tsolia Maria , Papaevangelou Vassiliki , Papadopoulos Nikolaos G.
TITLE=Recurrent Wheeze Exacerbations Following Acute Bronchiolitis—A Machine Learning Approach
JOURNAL=Frontiers in Allergy
VOLUME=2
YEAR=2021
URL=https://www.frontiersin.org/journals/allergy/articles/10.3389/falgy.2021.728389
DOI=10.3389/falgy.2021.728389
ISSN=2673-6101
ABSTRACT=Introduction: Acute bronchiolitis is one of the most common respiratory infections in infancy. Although most infants with bronchiolitis do not get hospitalized, infants with hospitalized bronchiolitis are more likely to develop wheeze exacerbations during the first years of life. The objective of this prospective cohort study was to develop machine learning models to predict incidence and persistence of wheeze exacerbations following the first hospitalized episode of acute bronchiolitis.
Methods: One hundred thirty-one otherwise healthy term infants hospitalized with the first episode of bronchiolitis at a tertiary pediatric hospital in Athens, Greece, and 73 age-matched controls were recruited. All patients/controls were followed up for 3 years with 6-monthly telephone reviews. Through principal component analysis (PCA), a cluster model was used to describe main outcomes. Associations between virus type and the clusters and between virus type and other clinical characteristics and demographic data were identified. Through random forest classification, a prediction model with smallest classification error was identified. Primary outcomes included the incidence and the number of caregiver-reported wheeze exacerbations.
Results: PCA identified 2 clusters of the outcome measures (Cluster 1 and Cluster 2) that were significantly associated with the number of recurrent wheeze episodes over 3-years of follow-up (Chi-Squared, p < 0.001). Cluster 1 included infants who presented higher number of wheeze exacerbations over follow-up time. Rhinovirus (RV) detection was more common in Cluster 1 and was more strongly associated with clinical severity on admission (p < 0.01). A prediction model based on virus type and clinical severity could predict Cluster 1 with an overall error 0.1145 (sensitivity 75.56% and specificity 91.86%).
Conclusion: A prediction model based on virus type and clinical severity of first hospitalized episode of bronchiolitis could predict sensitively the incidence and persistence of wheeze exacerbations during a 3-year follow-up. Virus type (RV) was the strongest predictor.