External validation of aminoglycoside models used in web calculators and clinical decision support systems after laboratory conversion to serum creatinine isotope dilution mass spectrometry assay.

ABSTRACT

BACKGROUND: Models to predict gentamicin t(½) from serum creatinine (SCr) estimated creatinine clearance (CrCl) are currently being incorporated into smart-device applications and clinical decision support modules without external validation. OBJECTIVE: The aim of this study was to determine whether such models remain viable after conversion to isotope dilution mass spectrometry (IDMS) SCr assay. METHODS: This study analyzed data from retrospective reviews of the medical records of nonobese adults receiving the aminoglycoside gentamicin and having ≥2 evaluable serum gentamicin concentrations after laboratory IDMS SCr conversion, from January 2008 to August 2009, at a tertiary care hospital in Florida. A literature search found a number of cited aminoglycoside models. This group of models was classified as group 1. The World Wide Web was also searched for the term aminoglycoside dosing calculators, with 6 models found and referred to as group 2. Predictive performance measures were used to compare the model results with the t(½) calculated from gentamicin concentrations using the Nelder-Mead algorithm. RESULTS: The records of 39 patients met the inclusion criteria (23 men, 16 women; age range, 18-86 years; range of estimated CrCl, 55-115 mL/min) and provided the "gold standard" aminoglycoside t(½). A gentamicin t(½) was predicted from several published models (group 1) and from other models used in online smart-device applications (group 2) and clinical decision modules. The median (interquartile range) root mean square errors were 0.48 (0.44 to 0.65) and 0.48 (0.45 to 0.70) hours from group-1 and -2 models, respectively. The median mean relative prediction errors were 9% (-14% to +13%) and 11% (+1% to +21%) from groups 1 and 2. The median mean absolute prediction errors were 21% (19% to 28%) and 21% (20% to 30%) from groups 1 and 2. Adjusting SCr by +20% improved the predictive ability in 3 of 12 cited models and in 5 of 6 models used in applications. CONCLUSIONS: Models to predict gentamicin t(½) should be externally validated at one's institution before use. The findings from the present study provide a framework for conducting external validation.