Use of prediction models to investigate respiratory support therapy of infants with acute viral Bronchiolitis: retrospective observational study using machine-learning techniques at a large tertiary center

Acute viral bronchiolitis (AVB) is the most common lower respiratory tract infection during the first year of life and the most frequent reason for hospitalization during infancy, generating extensive cost for healthcare systems. The overall aim of this thesis was to understand and optimise current respiratory therapies for AVB patients who presented to a large tertiary hospital in the USA. The study design was retrospective and observational. It used machine-learning techniques and causal inference algorithms to inform clinical decision making. Specifically, it compared the effectiveness of high-flow nasal cannula (HFNC) with that of standard treatment. The primary outcome was length of hospital stay. The dataset contained all AVB patients under the age of one year who presented to the Children's Hospital of Los Angeles between 01/2008 and 03/2017. In total, 891 patients were admitted and treated with either standard nasal cannula therapy or HFNC. The dataset was reduced to 599 cases after excluding significant co-morbidities and outliers to ensure the study was truly focussing on AVB patients. The analysis was performed in four steps: descriptive statistics, feature selection, data visualisation, propensity score matching, and predictive analytics. Propensity score matching was used to match patients in the standard group with those in the high flow group. Subsequent regression analysis estimated the average treatment effect of HFNC on the primary outcome. Due to decision bias, propensity score matching could not demonstrate a treatment effect of high flow therapy on hospital length of stay. This finding was in accordance with the latest literature. The list of the examined confounding variables included patient demographics, common co-morbidities, viral cause, vital parameters, and clinical descriptors of the respiratory state of the patient. In total, the combined influence of 22 covariates on the treatment choice and outcome was investigated. A newly created data-driven respiratory severity score incorporated those 22 covariates and converted them into individual scores. The sum of the individual scores generated a respiratory severity score. Respiratory severity scores, obtained at different times of the hospital stay, and other covariates (risk factors) were used to fit machine learning models that predicted hospital length of stay, prolonged length of stay (>5 days), the need for high flow therapy, and failure of standard therapy or high flow therapy. The results were highly significant. In addition to face and content validity, construct and prediction validity were successfully evaluated by applying statistical and machine learning tools. The respiratory severity score demonstrated promising characteristics when used in a fully computerised healthcare setting. As soon as full validation is achieved, it has the potential to become a useful instrument for clinical decision making, randomized controlled trials and comparative effectiveness research.