Late Breaking Abstract - Machine learning algorithms to predict bronchiolitis hospitalisations in the first year of life

<bold>Background:</bold> Reduced lung function is associated with lower respiratory tract infections in early life and may predict bronchiolitis risk. <bold>Aim:</bold> To utilize tidal breathing flow-volume loops (TBFVL) from healthy infants and train-test machine learning algorithms to generate predictive models for bronchiolitis hospitalisations. <bold>Methods:</bold> TBFVL were measured at 6 weeks of corrected age in infants born to mothers with asthma using an ultrasonic flow meter (Spiroson; Eco Medics, Duernten, Switzerland). TBFVL raw data were segmented into ‘windows’ of 10 breaths to extract signal descriptive features encompassing statistical, temporal and spectral domains. Fifteen prediction models were trained using a 70:30 training-testing data split and evaluated for Area under the ROC curve (AUROC), along with accuracy, recall, precision, kappa, F1 score and Matthews correlation coefficient. <bold>Results:</bold> Three hundred and sixty-seven infants were included in the analysis, of whom 22 (6.0%) were admitted to hospital, and 47 (13%) were either admitted or presented to the hospital emergency department, for bronchiolitis during the first year of life. Forty-four signal descriptive features were selected from 2645 TBFVL segments and included along with 8 demographic features as predictors for bronchiolitis hospitalisations. The best machine learning algorithm (Extra Tree Classifier) provided an AUROC of 0.95 for hospital admissions and AUROC of 0.97 for hospital admissions and/or presentations. <bold>Conclusion:</bold> The ability to predict bronchiolitis risk in healthy infants could alleviate challenges faced by health care providers to ensure both equality and equity in accessing RSV prophylaxis.