Explainable Artificial Intelligence in Pharmacovigilance and Drug Safety: A Systematic Review of Enhancing Transparency and Regulatory Acceptance
Virendra S. Gomase, Shenbagavel Vairachilai, Rupali Sharma, Suchita P. Dhamane, Satish Sardana, Bijith Marakarkandy, Sandeep KelkarIntroduction:
Explainable Artificial Intelligence (XAI) is rapidly transforming drug safety assessment by making complex AI models interpretable, transparent, and trustworthy for clinicians, healthcare regulators, and patients. This study emphasizes recent advances in the methodologies of XAI, including rule-based, post hoc, and hybrid models, and assesses the effectiveness of these approaches in pharmacovigilance and drug discovery.
Method:
This paper evaluates the state-of-the-art techniques in XAI methods, such as methodbased or rule-based approaches, post-hoc explanation methods, and hybrid methods in general, used in drug safety and pharmaceutical-related research and applications. The role and treatment of explainability in AI in the EU AI Act and EU MDR regulations will also be discussed in this paper.
Results:
Integration of XAI into clinical prediction systems (CPS) evaluation will increase the level and quality of transparency and error identification, and clinician trust, although with a trade-off between interpretation and accuracy within prediction tasks. There is a shift towards greater focus on transparency and human evaluation mentioned within regulatory frameworks, although there is no standardization with respect to clinical prediction systems (CPS) evaluation and validation of XAI
Discussion:
Emerging technologies emphasize the importance of clinician-centric design, personalized and causal inference-based explanations, and the integration of technical aspects and clinician feedback. Large research gaps are present in the standardization of evaluation criteria, ethical integration, and real-world clinically valid XAI systems.
Conclusion:
In clinical applications associated with drug safety, emerging areas of research should be targeted towards collaborations, evaluations, and XAI design adaptable to meeting growing demands for operation, acceptability, and patient safety.