Explainable Survival Models for Post-Marketing Safety Signal Prioritization Using Reporting and Literature Evidence Download PDF

Post-marketing safety surveillance must sift through large volumes of potential drug–adverse event associations, where signal timing is central to patient protection. Current prioritization tools often emphasize present evidence strength while giving less attention to when a signal may mature. Disproportionality analysis and manual review can help identify candidate associations, but they do not forecast the expected trajectory of signal maturation. They also provide limited transparency about which evidence streams most influence the urgency assigned to a signal. This article develops a conceptual explainable survival modeling framework for predicting time-to-signal confirmation or escalation for drug–adverse event pairs. The framework uses reporting trends, literature evidence, and product-specific covariates while decomposing each prediction into interpretable drivers. A deep survival model, such as DeepSurv, is proposed for longitudinal drug safety data with time-varying reporting and literature features updated over repeated surveillance intervals. SHAP values are used post hoc to attribute predicted hazard to specific features and evidence streams. Conceptually, the model could identify drug–event pairs that warrant earlier review than static prioritization approaches. For each prioritized signal, it would provide a transparent rationale, such as a rising reporting trajectory combined with emerging literature evidence accelerating the expected signal timeline. Explainable survival models could bring a more dynamic and auditable form of analytic support to post-marketing signal prioritization. Their greatest value lies in aligning time-to-event prediction with the evidentiary reasoning already used by pharmacovigilance experts.