In pharmacovigilance, linking the adverse reactions by patients to drugs they took is a key activity typically based on the analysis of patient reports. Yet generating potentially interesting pairs (drug, reaction) from a record database is a complex task, especially when many drugs are involved. To limit the generation effort, we exploit the frequently occurring patterns in the database and form \textit{association rules} on top of them. Moreover, only rules of minimal premise are considered as output by concept analysis tools, which are then filtered through standard measures for statistical significance. We illustrate the process on a small database of anti-HIV drugs involved in the HAART therapy while larger-scope validation within the database of the French Medicines Agency is also reported.