AI & ML in Pharmacovigilance

Software and digital systems used for case intake, processing, signal detection, or reporting; includes e-reporting, mobile apps, and cloud-based safety databases.

Documented evidence that computerized PV systems consistently perform as intended and meet GxP requirements (GVP Module I, Annex 11).

Risk that AI/ML models perpetuate or amplify biases (e.g., under-reporting in certain populations, coding bias) affecting signal detection or case handling.

Principle that qualified personnel must review and take responsibility for PV decisions supported or suggested by AI; automation does not replace accountability.

Monitoring of social media and digital channels for potential adverse event reports or product complaints; EMA and FDA have issued guidance on handling.

In digital/social media reports, the reporter may be identifiable by username, profile, or direct contact. Same validity criteria apply as for traditional reports.

System (typically validated) for storing and managing ICSRs and supporting regulatory reporting and analysis.

Ethical use of AI in PV: fairness, transparency, accountability, and human oversight to avoid harm and maintain trust.

Documented validation of AI/ML models used in PV (accuracy, reproducibility, bias check) per regulatory expectations.

Ability to explain how an AI system reached its output; important for PV decisions and inspections.

Policies for selection, quality, and representativeness of data used to train PV AI models.

Tracking of algorithm versions used in PV (e.g., signal detection) for reproducibility and audits.

Regular backup of safety database for business continuity and compliance.

Use of large datasets (EHR, claims, social) in PV; EMA Big Data Steering Group and guidance.

Use of predictive models (e.g., for risk of AE, case priority) in PV.

Use of rules or AI to triage incoming cases (e.g., serious, expedited).

Algorithm (rule-based or ML) to identify potential duplicate reports.