Using population-based serum drug concentration cutoff values to predict toxicity: test performance and limitations compared with Bayesian interpretation.

Abstract

The use of population-based serum drug concentration cutoff values for several commonly monitored drugs as tests to distinguish toxicity from nontoxicity was studied. Serum drug concentration and response data published in studies of theophylline, digoxin, aminoglycosides, vancomycin, and procainamide were analyzed to determine the prevalence of toxicity in each group of patients. Serum drug concentration cutoff values were then varied, and the following performance characteristics of the cutoffs as tests for predicting toxicity were calculated for each value: sensitivity, specificity, likelihood ratio, predictive value, and ratio of net consequences. At commonly accepted cutoff values, sensitivity was lower than desirable, positive predictive values were less than negative values, and the ratio of net consequences indicated that false-negative errors are implicitly weighted as more risky than false-positive errors. As the cutoff for each drug was increased, specificity increased but sensitivity decreased, positive predictive values increased but negative predictive values decreased, and the ratio of net consequences increased. Therapeutic drug monitoring would improve if practitioners and laboratorians collaborated to (1) conduct prospective studies of the test performance characteristics of drug concentrations, (2) estimate the pretest probability of toxicity for each patient, (3) combine the pretest estimate with the test characteristics to make a posttest estimate, and (4) develop a more patient-specific, Bayesian approach.