Machine Learning Methods for Predicting Adverse Drug Reactions in Hospitalized Patients

Abstract

Adverse Drug Reaction (ADR) in hospitalized patients have significant impacts on costs as well as on morbidity and mortality. A subset of ADR is preventable at the moment of prescribing. The use of supervised learning tools has been shown to be a good alternative to predict ADR, but more research is needed to make it a standard. A study was conducted to demonstrate how the use of machine learning methods can strengthen the ability to predict ADR in hospitalized patients. Binary classification models were constructed using three class balancing, feature selection and supervised learning approaches known as deep learning, random forest, and gradient boosting trees. Five-fold cross-validation and various evaluation metrics such as AUC, recall and F-measure were used to evaluate the models’ performance. Gradient boosting trees models was the approach that performed best, exhibiting a recall level of 78.3% and an ROC curve AUC of 0.81. Findings on attribute importance were also generated, which should thus prove useful in developing tools for clinical applications. Our predictive models’ results outperform those reported in the recent literature using traditional statistical methods. In this work we have shown the enormous potential of ML tools in this field and their great contribution to the development of information systems that use them.