AI System Validation in Pharmacovigilance
The integration of artificial intelligence technologies into pharmacovigilance processes introduces a set of engineering and regulatory challenges that traditional Computer System Validation frameworks do not fully address.
The Regulatory Context: GVP Module VI and the Entry of Artificial Intelligence
GVP Module VI governs the processes of collection, management, and reporting of safety reports. The introduction of AI systems into the pharmacovigilance cycle does not exempt the organisation — and in particular the QPPV — from the responsibility of human oversight on every clinical-regulatory decision.
Human-in-the-Loop: Architecture and Validation Requirements
The concept of human-in-the-loop (HITL) in pharmacovigilance is not a generic good practice principle: it is an architectural requirement that must be designed, documented, and validated as an integral part of the system.
A compliant AI-supported PV system requires that:
- The AI system produces outputs labelled with confidence levels and justification (explainability)
- Every output classified as a potential ICSR is mandatorily subject to review by qualified personnel
- The final decision remains the exclusive prerogative of a qualified human operator
- The HITL workflow is fully traced in the system, with an audit trail compliant with Annex 11, Section 9
AI PV System Validation: Engineering Approach
Our approach includes: documented Intended Use; supplier assessment per Annex 11; risk-based validation scope per ICH Q9 R1 and GAMP 5 Second Edition; statistical test design for probabilistic systems (accuracy, precision/recall, F1-score on segregated reference datasets); and audit trail verification per Annex 11 and ALCOA+ principles.
GMP Annex 22 (Consultation Guideline) is the first European GMP framework specifically dedicated to AI/ML systems in regulated pharmaceutical environments. Core principles include model lifecycle management, formal change control for model updates, explainability by design, and mandatory human oversight.
EU AI Act Annex III: High-Risk AI Systems in Healthcare
The EU AI Regulation (EU AI Act, Reg. EU 2024/1689) classifies in Annex III, point 5, AI systems used in healthcare as high-risk systems. Practical implications for pharmaceutical organisations include: registration in the EU database prior to deployment; mandatory conformity assessment; human oversight by design; automatic logging requirements; and a dedicated AI quality management system.
Explainability and Traceability of AI Decisions in GVP
Explainability is not merely an ethical requirement: in a regulated GVP context, the ability to justify why the AI system classified an event as a potential ICSR is necessary for: operator review (HITL); responses to regulatory inspections; and the defence of pharmacovigilance decisions in liability contexts.
Our Operational Approach
Dalia IA intervenes in AI pharmacovigilance system validation through a structured, phased approach: regulatory gap analysis; Intended Use definition and Risk Assessment; validation plan drafting (GAMP 5 Second Edition, adapted for probabilistic AI/ML systems); test execution and documentation; inspection-ready documentation package; and post-validation model monitoring.
Frequently Asked Questions
Is validation of an AI system used in pharmacovigilance mandatory?
Yes. Any computerised system used in regulated GxP processes — including AI-supported pharmacovigilance systems — must be validated according to applicable requirements (Annex 11, GVP Module VI). The use of AI or ML does not reduce this obligation.
What is the difference between an AI support system and an AI decision system in PV?
An AI support system produces suggestions, classifications, and signals that are always subject to human review before any regulatory action. Human-in-the-loop design is not an organisational choice; it is a regulatory and architectural requirement.
How does the EU AI Act apply to AI systems in pharmacovigilance?
The EU AI Act classifies as high-risk AI systems used in healthcare (Annex III, point 5). Requirements include a conformity assessment, a dedicated AI quality management system, registration in EU databases, and human oversight by design.
How is change control managed for an AI model that is periodically retrained?
Model retraining constitutes a system modification and must follow a formal change control process: impact assessment on the existing validation, regression testing on the new model, documentation update, and formal approval before return to production.
What is meant by explainability of an AI system in pharmacovigilance?
Explainability is the ability of an AI system to provide, for each output, a human-readable justification that explains the factors contributing to that classification or decision.