B1725
Title: Machine learning methodologies for the prediction of rapid lung function decline
Authors: Judith Dexheimer - Cincinnati Children's Hospital (United States) [presenting]
Abstract: Cystic Fibrosis (CF) affects more than 70,000 individuals worldwide. Patients have an average of 9 visits per year. We collected data from the US Cystic Fibrosis Foundation Patient Registry which contains clinical encounter-level data obtained from patients at accredited care centers. We developed machine learning methodologies to predict the decrease in lung function, FEV1-Indicated Exacerbation Signal (FIES). We developed five machine learning models: Logistic Regression (LR), Random Forest (RF), K-Nearest Neighbors (KNN), Artificial Neural Network (ANN), and a Recurrent Neural Network (RNN). Outcomes included area under the receiver operating characteristic curve (AUROC), sensitivity, specificity, recall, and precision. From the 20,153 patients with data in the registry, 13,285 were included in the analysis. Patients were split into a training and validation cohort. Using 10-fold cross validation, the RNN and RF performed best with AUROC = 0.900 and AUROC=0.885 respectively. Findings across all methods will be compared, and implications of recall and precision will be discussed including accounting for longitudinal dependencies in electronic health record data.
