The limitations we have lived with

For five decades, post-marketing safety surveillance has depended primarily on spontaneous reporting — the voluntary submission of individual case safety reports by healthcare professionals and patients to national pharmacovigilance databases. The system has genuine strengths: it is global, it captures unexpected events, and it has detected real signals that led to meaningful safety actions.

But its limitations are well-characterized and fundamental. Spontaneous reporting captures an estimated 1-10% of actual adverse events. Reporting rates vary by country, by event seriousness, by time since launch, and by media attention. The denominator — how many patients are actually exposed — is unknown. And the data cannot establish causality, only statistical association.

These limitations have been understood since the 1960s. What has changed is that alternatives are now operationally feasible.

The real-world evidence landscape

Real-world evidence in pharmacovigilance draws from three primary data sources: electronic health records, administrative claims databases, and disease or product registries. Each has distinct strengths and limitations for safety surveillance.

Electronic health records offer clinical granularity — laboratory values, imaging results, clinical notes — that spontaneous reports cannot capture. Claims databases offer scale and longitudinal follow-up that trials cannot match. Registries offer structured, protocol-driven data collection for specific populations or exposures.

The analytical infrastructure to use these data sources for safety surveillance has matured significantly. The FDA's Sentinel System and the EMA's Darwin EU network represent the most advanced regulatory implementations, both using distributed data models that enable population-level analyses without centralizing patient data.

The methodological reality

The promise of real-world evidence in pharmacovigilance is real but bounded by fundamental methodological constraints. Confounding by indication — the fact that patients who receive a drug differ systematically from those who do not — cannot be fully resolved through observational methods alone. Propensity score matching, instrumental variables, and target trial emulation frameworks reduce but do not eliminate confounding.

The path forward is not real-world evidence replacing spontaneous reporting, but a multi-source pharmacovigilance model where spontaneous reports, clinical trial data, and real-world evidence each contribute complementary information to an integrated safety assessment. The regulatory framework to support this model is being built, but the methodological standards that will determine whether real-world evidence is trusted in regulatory decision-making are still being defined.