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

Signal detection is a core task in pharmacovigilance, and disproportionality analysis (DA) remains the dominant approach for spontaneous reporting systems (SRS). However, SRS data are sparse and heavily zero-inflated, reflecting rare adverse drug reactions (ADRs) and substantial under-reporting, which can distort standard count-model assumptions. Using Medicines and Healthcare products Regulatory Agency (MHRA) interactive Drug Analysis Profiles (iDAPs) for selective serotonin reuptake inhibitors (SSRIs), we derived system organ class (SOC)-level adverse event counts. Goodness-of-fit testing indicated that SOC counts were variably consistent with Poisson or negative binomial (NB) distributions, motivating the use of zero-inflated Poisson and zero-inflated NB models. Model adequacy was assessed using Pearson-residual dispersion (absolute fit) and Akaikes Information Criterion (relative fit). Zero-inflated models provided improved fit for 10 out of all 27 SOCs (Endocrine disorders, Immune system disorders, Metabolism and nutrition disorders, Infections and infestations, Renal and urinary disorders, Hepatobiliary disorders, Ear and labyrinth disorders, Pregnancy, puerperium and perinatal conditions, Congenital, familial and genetic disorders, Neoplasms benign, malignant and unspecified (including cysts and polyps)) and yielded signals broadly concordant with DA, including for metabolism-related events. These results support zero-inflated modelling as a complementary signal detection method for iDAPs-derived count data, particularly where internal model-based validation and explicit handling of excess zeros are desired. Key PointsO_LIStandard disproportionality methods can struggle with sparse pharmacovigilance data because many drug-event combinations have lots of zeros. C_LIO_LIZero-inflated models better reflected the structure of the data for many system organ classes, improving model fit without changing the overall safety message. C_LIO_LIThe approach produced signals broadly consistent with traditional analysis, supporting its use as a complementary method rather than a replacement. C_LIO_LIHandling excess zeros directly may improve detection of rare or under-reported adverse drug reactions in spontaneous reporting systems. C_LIO_LIThese findings could strengthen post-marketing safety monitoring for SSRIs and other medicines with sparse reporting patterns. C_LI Plain Language SummaryWe studied whether a different way of analysing reports of side effects could help spot safety concerns for antidepressants called selective serotonin reuptake inhibitors (SSRIs). Using UK medicines safety reports, we compared a standard counting method with a method designed for data that contain many categories with no reports at all. The new method fitted the data better for many body-system groups and gave signals that were broadly similar to the standard approach, while also handling the large number of empty categories more realistically. This matters because side effects in safety databases are often rare and under-reported, which can make them hard to detect. Our findings suggest that these newer statistical methods could improve how researchers and regulators monitor medicine safety, especially when the data are sparse and unevenly reported.