Post-marketing safety of tarlatamab in small cell lung cancer based on FAERS and WHO-VigiAccess with SHAP-based interpretable machine learning analysis of immune-related adverse events.

Background: Tarlatamab is a DLL3-targeted bispecific T-cell engager approved for previously treated extensive-stage small cell lung cancer (ES-SCLC). However, its post-marketing safety profile in routine practice remains incompletely characterized. Methods: We conducted a retrospective pharmacovigilance study using the FDA Adverse Event Reporting System from Q2 2024 to Q3 2025, with a descriptive cross-database comparison using WHO-VigiAccess. Disproportionality analyses were performed using reporting odds ratio, proportional reporting ratio, Bayesian confidence propagation neural network, and multi-item gamma Poisson shrinker algorithms. Clinical priority scoring, subgroup analysis, time-to-onset analysis, multivariable logistic regression, and interpretable machine learning with SHAP (SHapley Additive exPlanations) were further applied. Results: A total of 942 reports with tarlatamab as the primary suspect drug were identified, comprising 1,346 adverse events. At the preferred-term level, 30 signals met all four disproportionality criteria. The most frequent and strongest signals were cytokine release syndrome (CRS; n = 201; ROR 223.84, 95% CI 192.15-260.77) and immune effector cell-associated neurotoxicity syndrome (ICANS; n = 106; ROR 312.43, 95% CI 254.68-383.26). Common additional signals included pyrexia, dysgeusia, hypotension, and ageusia. Potentially under-recognized events, such as intestinal perforation, dyspnoea at rest, and incontinence, were also detected. Most adverse events occurred early after treatment initiation, with a median time to onset of 3 days; CRS and ICANS both showed a median onset of 1 day. In multivariable analysis, concomitant medication use was associated with higher reported odds of CRS (OR 2.551, 95% CI 1.353-4.811), whereas reports from Japan and the year 2025 were associated with lower reported odds. The CRS report-level classification model showed acceptable discrimination (AUC 0.733), whereas the best machine-learning model for adverse events of immune disorders classification within FAERS reports showed only modest performance (validation AUC 0.639). SHAP analysis indicated that country and therapy type contributed more to model output than age and sex, suggesting that the model primarily captured reporting context and treatment complexity rather than intrinsic biological susceptibility. Cross-database comparison with WHO-VigiAccess showed a broadly concordant reporting pattern, with CRS and ICANS remaining the most commonly reported toxicities. Conclusion: Post-marketing data indicate that tarlatamab has a distinct safety profile dominated by early-onset immune-mediated and neurologic toxicities, particularly CRS and ICANS. Early monitoring and prompt supportive management during initial treatment cycles are essential. These findings broaden current knowledge of tarlatamab safety in real-world practice and support prospective studies to pharmacovigilance-based signal stratification and monitoring strategies.