Finding Safety Signals in Sparse Data: Zero-Inflated Models for Pharmacovigilance studies

Signal detection is a core task in pharmacovigilance, with disproportionality analysis remaining the dominant approach for spontaneous reporting systems. However, data from these systems are sparse and contain many zeros, reflecting the rarity of adverse drug reactions and significant under-reporting. Researchers analyzed adverse event data from the British regulatory agency MHRA for SSRIs at the level of system organ classes. Goodness-of-fit testing showed that data were partially consistent with Poisson or negative binomial distributions. Zero-inflated models provided improved fit for 10 out of 27 examined adverse event categories, including endocrine disorders, immune system disorders, and metabolism-related disorders. The results support the use of zero-inflated models as a complementary method for safety signal detection, particularly when internal model validation and explicit handling of excess zeros are desired.