Using machine learning and clinical registry data to uncover variation in clinical decision making

Abstract

Clinical registry data contains a wealth of information on patients, clinical practice, outcomes and interventions. Machine learning algorithms are able to learn complex patterns from data. We present methods for using machine learning with clinical registry data to carry out retrospective audit of clinical practice. Using a registry of stroke patients, we demonstrate how machine learning can be used to: investigate whether patients would have been treated differently had they attended a different hospital; group hospitals according to clinical decision making practice; identify where there is variation in decision making between hospitals; characterise patients that hospitals find it hard to agree on how to treat. Our methods should be applicable to any clinical registry and any machine learning algorithm to investigate the extent to which clinical practice is standardized and identify areas for improvement at a hospital level.