All Times ET
Keywords: Supervised Learning, Precision Medicine, Clinical Decision Support, Physiologic Time Series, Predictive Modelling
Adverse drug reactions (ADR) are among the top five leading causes of death in the US. Concerns about inducing an ADR may deter clinicians from properly delivering beta blocker therapy to cardiac patients in the Intensive Care Unit (ICU). Our team deployed supervised and unsupervised machine learning algorithms on physiological signals data combined with clinical and demographic markers to classify and predict the patient-level risk of Carvedilol-related ADRs in ICU patients. By examining data prior to drug administration, ADR cases could be discriminated from controls with medically meaningful accuracy in a matched case-control study, per the Area Under the Receiver Operating Characteristics Curve (ROC) metric (GLM: ROC=0.84 (0.80-0.87); Random Forest: ROC=0.84 (0.81-0.87); XGBoost: ROC=0.83 (0.79-0.86)). Further, adding heart rate data to a traditional model based on clinical and demographic data alone (GLM: ROC=0.58) resulted in large increases in predictive value. Similar results were observed in a class-imbalanced setting designed to mimic ADR incidence in the ICU (GLM: ROC=0.83 (0.83-0.84); Random Forest: ROC=0.84 (0.84-0.85); XGBoost: ROC=0.83 (0.83-0.84)). The precision and recall metrics computed from the Random Forest model (41% precision, 50% recall) in the class-imbalanced design showcases the efficacy of such a model as a predictive tool if introduced into a clinical setting.
Successful integration of physiological signals data with clinical data could alert clinicians to impending ADRs before drug administration, thereby reducing patient harm. Broader use of this novel approach utilizing other physiologic signals has potential to classify other adverse events in clinical settings.