The predictive performance of Bayesian estimates incorporating pharmacokinetic values for hematology-oncology patients was compared with that of Bayesian estimates incorporating general population values. In study phase 1, medical records were reviewed for 50 adult patients with a hematologic or oncologic diagnosis who had received i.v. gentamicin or tobramycin. Aminoglycoside pharmacokinetic values were calculated for the patients by using a modified two-point Sawchuk-Zaske method, and the subpopulation mean for each variable was determined. In phase 2, data for 10 other hematology-oncology patients receiving aminoglycosides were entered into the Abbottbase Bayesian pharmacokinetics program. Aminoglycoside pharmacokinetic values and serum concentrations for each of these 10 patients were estimated, first using the program's general population values and then repeating the analysis using the subpopulation means for volume of distribution and renal clearance slope obtained in phase 1. The serum aminoglycoside concentrations predicted by each Bayesian method were compared with the actual peaks and troughs. Both the peak and trough predictions of the Abbottbase program using the subpopulation values for volume of distribution and renal clearance slope were significantly less biased than those predicted by the Abbottbase program incorporating the general population values. The methods did not differ significantly in precision. Use of subpopulation pharmacokinetic values in Bayesian predictions of serum aminoglycoside concentrations in hematology-oncology patients reduced bias significantly but had no significant effect on precision.