A novel Bayesian drug dosing program (Abbott Pharmacokinetic Systems, Theophylline Program) was evaluated. The predictive accuracy of this method was assessed in 10 healthy volunteers receiving single intravenous test doses. Estimates for clearance and distribution volume were compared with those obtained from the area under the curve. The observed prediction error depended largely on sampling time. The deviations were lowest for the distribution volume during the first 60 min and for clearance at 12 hours after theophylline administration. Furthermore the Bayesian technique was prospectively evaluated in 10 hospitalized and 22 outpatients treated with sustained-release theophylline preparations (Uniphyllin, Bronchoretard, PulmiDur). Predictive precision and accuracy were adequate, if theophylline was given twice daily. The highest predictive accuracy was achieved in outpatients, if predictions were based on trough concentrations. In 19/22 outpatients prediction errors were within a clinically acceptable range (mean prediction error +/- standard deviation; -0.6 +/- 2.1 mg/l). Moreover in hospitalized patients (n = 5) with twice-daily maintenance regimens, concentration-time curves could mainly be predicted with sufficient accuracy. Hospitalized patients (n = 5) with once-daily dosing showed large fluctuations between peak and trough theophylline concentrations in serum. In these patients a reliable prediction of the concentration-time curves was not possible apparently due to non-linearity of theophylline kinetics. Relatively large prediction errors were found in one patient with acute viral respiratory illness and 3 patients with altered absorption. Despite certain limitations the clinical application of the Bayesian forecasting method tested appears to be promising.