Managing Clinical Use of High-Alert Drugs: A Supervised Learning Approach to Pharmacokinetic Data Analysis

Drug-related problems, particularly those that result from sub- or overtherapeutic doses of high-alert medications, have become a growing concern in clinical medicine. In this paper, we use a model-tree-based regression technique (namely, M5) and support vector machine (SVM) for regression to develop learning-based systems for predicting the adequacy of a vancomycin regimen. We empirically evaluate each system´s accuracy in predicting patients´ peak and trough concentrations in different clinical scenarios characterized by renal functions and regimen types. Our data consist of 1099 clinical cases that were collected from a major tertiary medical center in southern Taiwan. We also examine the use of bagging for enhancing the prediction power of the respective systems and include in our evaluation a salient one-compartment model for performance benchmark purposes. Overall, our evaluation results suggest that both M5 and SVM are significantly more accurate than the benchmark one-compartment model in predicting patients´ peak and trough concentrations across all investigated clinical scenarios. M5 appears to benefit considerably from bagging, which has a positive but seemingly smaller effect on SVM. Taken together, our findings indicate supervised learning techniques that are capable of effectively supporting clinicians´ use of vancomycin or similar high-alert drugs in their patient care and management.