Comparison of a pharmacovigilance screening heuristic with expert clinical consensus for adverse drug reaction signal classification at a Pakistani tertiary care hospital

Background and Objective: Adverse drug reactions (ADRs) have a disproportionately higher burden in low-and middleincome countries where pharmacovigilance infrastructure is under-resourced. Electronic medical records (EMRs) can quantify drug-ADR associations but may require interpretive classification. We compared a statistical heuristic against expert clinical consensus for 82 drug-ADR signals from a Pakistani tertiary care hospital. Methodology: This is a retrospective analysis used electronic records from the Department of Internal Medicine at Aga Khan University Hospital, Karachi, Pakistan from 2018-2022. In this study logistic regression with false discovery rate correction was applied to EMR data (2018–2022). A four-category heuristic (True Signal, Likely Confounding, Reverse Causation, Null/No Signal) was compared to independent classifications by expert clinicians. Agreement was assessed using Cohen’s κ and Fleiss’ κ. Results: Heuristic-clinician concordance was 36.6% (κ=0.106, p=0.065). Fleiss’ κ among reviewers was near 0 (p=0.997). Of 52 discordant signals, 69.2% involved experts assigning higher signal-ADR plausibility than the heuristic. Conclusion: Heuristic and clinical judgment might complement each other. Potential hybrid frameworks warrant prospective evaluation for efficient ADR signal refinement in resource-limited settings.