The detection of Adverse Drug Reactions (ADRs) in clinical records plays a pivotal role in pharmacovigilance (PhV). Achieving near-ideal practice relies on well-trained health professionals, who are trained to identify, assess, and report to health authorities ADRs occurring after drug marketing approval, including those that are infrequent. However, the number of experts trained in this practice is low and despite reporting ADRs being mandatory for healthcare professionals, pharmacovigilance still suffers from a significant under-reporting, accounting for only 5-10% of all ADRs. Yet, drug safety is crucial for assessing the benefit/risk ratio of a given drug. It is therefore important to circumvent under-reporting and to be able to collect ADRs automatically from medical reports. The most natural approach would be to train a model in a supervised manner, which requires annotation of a large volume of data, but this is unfortunately not possible. We therefore propose here an unsupervised approach to distinguish between ADRs-related and non-related reports. From a more formal point of view, we address this problem as a clustering task aiming at distinguishing medical reports containing the description of an ADR from those without.