PNS143 APPLICATION OF MACHINE LEARNING IN PREDICTING HOSPITAL READMISSION: A SYSTEMATIC REVIEW OF LITERATURE

Abstract

Reducing hospital readmission helps to improve quality of inpatient care. Advances in machine learning provide great opportunities in understanding readmission risk. This study synthesizes the literature on application of machine learning-based models in predicting hospital readmission and further evaluates their performance. Electronic databases PUBMED, MEDLINE and EMBASE were systematically searched from January 1, 2015 through December 10, 2019 by two independent reviewers. Only observational studies published in English language, using machine learning technique-based predictive models for hospital readmissions in US patients were included. Data were extracted on study characteristics (data source, population and setting, outcome, sample size); and model performance (algorithm, model derivation/validation, discrimination, key predictive variables) using Covidence platform. Of 522 citations reviewed, 46 studies met the inclusion criteria. The most common outcome was 30-day readmission. The study population primarily comprised of heart failure patients (12 studies), all inpatients (8 studies), and postoperative patients (7 studies); the remaining studies focused on mental health, ICU, pediatric, and rehabilitation inpatients. Twenty-nine studies used electronic health records, three studies used administrative claims data, while the remaining fourteen studies employed population-based data sources. A number of thirty-six studies used classical machine learning algorithms (c statistic range: 0.46-0.85), including decision tree(17 studies) , lasso regression(6 studies), neural networks(4 studies), support vector machine(4 studies), and K-nearest neighbor (1 study); and nineteen of these studies included both machine learning algorithms and traditional regression methods, of which thirteen found ML outperformed traditional regression methods. In the remaining ten studies, deep learning models were explored (c statistic range: 0.60-0.76), of which eight studies found that deep learning models improved predictive performance of readmissions compared to traditional classification models. Machine learning algorithms are increasingly used to predict hospital readmission in US patients. These machine learning algorithms tend to have better predictive performance for hospital readmission compared to traditional classification models.