45 Predicting long length of stay in a paediatric intensive care unit using machine learning. (30th November 2020)

• Record Type: Journal Article
• Authors: East, Abigail; Ray, Samiran; Pope, Rebecca; Cortina-Borja, Mario; Sebire, Neil J
• Is Part Of: Archives of disease in childhood. Volume 105(2020)Supplement 2
• Page Start: A15
• Page End: A16
• Publication Date: 2020-11-30
• Subjects: Children -- Diseases -- Periodicals; Infants -- Diseases -- Periodicals; 618.920005
• DOI: 10.1136/archdischild-2020-gosh.45
• Languages: English
• ISSNs: 0003-9888

Abstract

Abstract : Introduction: Length of stay (LOS) prediction modelling in intensive care units is a valuable capacity planning tool as hospitals attempt to clear the backlog of surgical patients resulting from the COVID-19 pandemic. Recent work in adults has demonstrated the benefits of using machine learning over statistical methods for LOS prediction, however machine learning approaches have not been applied to paediatric populations. Objectives: The study set out to develop machine learning models to predict long LOS in the paediatric intensive care unit at Great Ormond Street Hospital using electronic patient records. Methods: Paediatric intensive care patients between 1st May 2019 and 30th April 2020 were extracted from electronic patient records. Random forest, XGBoost, and multilayer perceptron models were built to predict LOS greater than three or seven days. The dataset contained demographics, ventilation data, and summary statistics of physiological time-series data, taken from the first twelve hours of admission. The performance of the machine learning classifiers was compared to a baseline logistic regression model. Results: There were 564 patients in the study population, of whom 307 had a LOS greater than three days and 105 had a LOS greater than seven days. Using the seven-day threshold, the optimal model was the random forest, which achieved an AUC of 0.785 and correctly classified 42.9% of long LOS patients. Using the three-day threshold, the optimal model was the multilayer perceptron, which achieved an AUC of 0.737 and correctly classified 85.7% of long LOS patients. The performance of the machine learning models was variable, and they did not unanimously outperform the baseline models. Conclusions: The machine learning models performed poorly in predicting long LOS. Further work is required to assess the clinical utility and value of deep learning methods in an operational setting.