Contrastive learning of adverse events to provide effective and interpretable vector representations for machine-assisted pharmacovigilance

Description

Olivér Balogh and co-authors teach in 2026 a machine to map drug side effects as points in a space, so similar harms sit near each other. Drug safety watching still runs on text reports and counting rules, which blunt modern learning tools. Their method learns vector representations of adverse events from real spontaneous report data, using contrastive learning, pulling true pairs together and pushing strangers apart.

The resulting map was probed from several angles: clustering, meaning checks, and spread comparisons. Its neighborhoods reflected both what an event does and what causes it, and it captured safety knowledge better than standard medical vocabularies and general text encoders. Fed into a classifier, the vectors beat the counting method regulators commonly use at predicting drug event links. Interpretable structure is the selling point: a safety officer can see why two events sit together, not just that a model said so.

Catalog record

Source: https://europepmc.org/articles/PMC13541282