Machine learning model combining features from algorithms with different analytical methodologies to detect laboratory-event-related adverse drug reaction signals

摘要: Background The importance of identifying and evaluating adverse drug reactions (ADRs) has been widely recognized. Many studies have developed algorithms for ADR signal detection using electronic health record (EHR) data. In this study, we propose a machine learning (ML) model that enables accurate by integrating features from existing based on inpatient EHR laboratory results. Materials methods To construct an reference dataset, extracted known drug–laboratory event pairs represented test the EU-SPC SIDER databases. All possible pairs, except ones, are considered unknown. To detect pair, three algorithms—CERT, CLEAR, PACE—were applied to 21-year We also constructed ML models (based random forest, L1 regularized logistic regression, support vector machine, neural network) use intermediate products CERT, PACE as inputs determine whether pair is associated. For performance comparison, evaluated sensitivity, specificity, positive predictive value (PPV), negative (NPV), F1-measure, area under receiver operating characteristic (AUROC). Results All measures outperformed those with sensitivity 0.593–0.793, specificity 0.619–0.796, NPV 0.645–0.727, PPV 0.680–0.777, F1-measure 0.629–0.709, AUROC 0.737–0.816. Features related change or distribution shape were important detecting signals. Conclusions Improved indicated applying our data feasible promising more comprehensive signals.