SeltaSquare

1. 
2. 

• 게시판 리스트 옵션
  • 검색

In the previous content, we explored the foundational principles for adopting AI in pharmacovigilance. In this Part 2, we will examine how those values can be effectively embedded into real-world practice through the remaining guidelines. We have summarized the core principles required for AI to become a trusted partner for field experts, including transparency, data security, fairness, and governance that oversees the entire process—ultimately completing a practical operational framework.

CIOMS WG XIV’s 7 Principles for AI in Pharmacovigilance

[Created by Gemini]

Principle 4. Transparency & Explainability

Transparency and explainability mean that stakeholders must clearly understand what data the AI has been trained on, what algorithms it uses, and what its limitations are in order to establish trust. A key enabling technology for this is Explainable AI (XAI). XAI provides the rationale behind AI outputs by showing which specific words, contexts, or data patterns led to a given result. By making the AI’s decision-making process visible, it creates a foundation of trust that allows users to confidently accept its outputs. This not only helps experts quickly identify errors in AI outputs but also supports faster and more effective decision-making. In regulatory inspections, it also serves as strong evidence to demonstrate the system’s validity by providing a logical basis for decisions. However, the explanations provided by XAI reflect the model’s reasoning process and do not always align with scientific truth; therefore, experts must critically interpret them while being mindful of hallucinations.

In addition to technical transparency, ensuring the real-world usefulness of AI systems requires robust KPI management tailored to pharmacovigilance. Rather than relying solely on overall accuracy, it is essential to carefully manage PV-specific metrics such as precision—the proportion of truly relevant information among selected outputs—and recall—the proportion of relevant information successfully detected out of all available signals. From the stage of metric design, domain experts should be involved to interpret the clinical meaning behind numerical values, thereby establishing a system of continuous quality improvement. This is a key component of successful AI adoption in pharmacovigilance.

Principle 5. Data Privacy

Given the sensitive nature of patient data in pharmacovigilance, strict compliance with privacy regulations must be ensured from the earliest stages of technology adoption. In particular, when using generative AI such as large language models (LLMs), the risk of data leakage or re-identification increases significantly. Therefore, organizations must go beyond basic legal compliance and establish advanced, end-to-end data security frameworks. This does not mean restricting the use of AI, but rather implementing practical safeguards such as privacy-enhancing technologies (PETs) and controlled data access mechanisms. By applying de-identification techniques to prevent AI models from identifying individuals within training data, and by conducting Data Protection Impact Assessments (DPIA) to proactively evaluate privacy risks, organizations can build a safe and trustworthy PV AI environment.

Principle 6. Fairness & Equity

Fairness and equity refer to designing and managing AI systems so that they do not produce biased outcomes against specific groups such as race, gender, or geography. This goes beyond preventing technical errors—it ensures that no population group is misrepresented or unfairly assessed in safety evaluations. To achieve this, potential bias must be identified early using statistical methods during the AI adoption phase, and system performance must be continuously monitored to ensure consistent outcomes even for underrepresented populations. In pharmacovigilance, fairness is directly tied to how accurately AI reflects the actual exposed population for a given medicinal product. Therefore, it is critical to ensure that training and evaluation datasets are sufficiently representative, and where possible, to evaluate performance separately across different subgroups.

A key challenge in ensuring fairness in ICSR (Individual Case Safety Reports) remains the structural limitations of spontaneous reporting systems. Reporting rates vary by country, and access to real-world data (RWD) differs significantly, limiting visibility into underrepresented populations. Ultimately, managing fairness and equity requires confronting these data representativeness issues directly and implementing both statistical and operational measures to ensure that all patients receive equal safety benefits.

Principle 7. Governance & Accountability

The final principle emphasizes that AI must go beyond being a simple technological tool and must be systematically governed within existing pharmacovigilance Quality Management Systems (QMS). Since systems themselves cannot be held accountable, it is essential to clearly define the roles and responsibilities of the entities that manage AI and are accountable for its outputs. This governance framework must cover the entire lifecycle of AI systems—from planning, development, and pre-deployment validation to operation and post-deployment monitoring—and must also include qualification processes for software vendors and service providers. In particular, the “Governance Framework Grid” proposed by CIOMS provides a structured methodology to define control scopes based on risk levels at each stage, ensuring end-to-end transparency across the system lifecycle.

From Detection to Prevention: The Future of Pharmacovigilance Led by Selta Square

The CIOMS 7 Principles discussed above represent not only a minimum set of commitments for successfully integrating AI into pharmacovigilance, but also a practical roadmap for implementation. When these principles are fully embedded into real-world practice, the field can move beyond the long-standing paradigm of “reactive response” and toward a fundamental transformation of pharmacovigilance. We are entering a future where AI does not merely detect adverse events after they occur, but proactively prevents them by predicting risks in advance.

Within this wave of innovation, Selta Square is already delivering tangible outcomes. Most notably, through a joint research collaboration with Samsung Medical Center, we have identified the “Golden Standard Factor” for contrast media adverse reactions, enabling personalized risk prediction guidelines tailored to individual patient characteristics. This demonstrates that prevention—beyond detection—is already an active, real-world capability, and signals that truly personalized safety management integrating genetic and lifestyle data is no longer far away.

AI is now evolving beyond a simple tool into a powerful partner that enhances the efficiency of pharmacovigilance. In this transformation, our role must also evolve—from data analysts to strategic leaders capable of guiding AI systems and making high-value decisions. Selta Square remains committed to standing alongside you as a trusted partner, helping you lead this change and build the safest possible pharmacovigilance systems.