An important role for the WHO Programme for International Drug Monitoring is to identify signals of international drug safety problems as early as possible. Since 1998, Bayesian Confidence Propagation Neural Network (BCPNN) data mining has been in routine use for screening of the WHO adverse reaction database, Vigibase. The identification of drug/adverse drug reaction combinations that have disproportionately high reporting relative to the background of all reports constitutes the first, quantitative step in the Uppsala Monitoring Centre (UMC) signal-detection process. In order to improve the signal-to-noise ratio and to focus on possible signals that are less likely to be detected by individual national pharmacovigilance centres, an expert group considered a number of possible subsidiary selection algorithms to be added as a second filtering step before potential signals were sent to the UMC expert panel for clinical review. As a result of these deliberations, three selection algorithms were implemented for routine use in 2001: 'serious reaction and new drug', 'rapid reporting increase' and 'special interest terms'. The effect of applying these algorithms has been critically evaluated on the basis of the ratio of associations selected to signals found and some modifications decided. Bearing in mind that any filtering strategy is likely to exclude some potential true signals from consideration, we think that triage strategies based on a combination of pragmatic thinking and experience are effective, provided that the results are reviewed at regular intervals and the algorithms adjusted on the basis of performance.