Identification and monitoring of immune-related adverse reactions of PD-1/PD-L1 inhibitors based on symptom network analysis

Tiivistelmä

Background PD-1/PD-L1 inhibitors have become first-line cancer therapies by disrupting immune checkpoint downregulation, thereby potentiating antitumor immunity. However, this enhanced immune activation may also target normal tissues, triggering multisystemic immune-related adverse events (irAE) that complicate clinical management. Conventional pharmacovigilance analyses typically assess individual adverse events in isolation, limiting the ability to capture inter-symptom relationships. Symptom network analysis, an emerging computational approach, constructs co-occurrence networks to identify core symptoms, key associations, and bridge nodes across organ systems, thus enabling a systematic understanding of complex suspected irAE profiles. Methods Adverse event records for six PD-1/PD-L1 inhibitors were extracted from two spontaneous reporting databases: FAERS (Q1 2005-Q3 2024) and JADER (Q1 2004-Q1 2025). Potential signals were screened using reporting odds ratio(ROR), proportional reporting ratio(PRR), and multivariate gamma-Poisson shrinkage (MGPS) methods. Symptom co-occurrence networks were subsequently constructed using EBIC-GLASSO based on Tetrachoric correlation matrices, which are specifically designed for binary adverse event data and correct for attenuation due to dichotomization. Four centrality metrics—Strength, Closeness, Betweenness, and Expected Influence—were calculated via the qgraph package. The bootnet package performed 1,000 bootstrap iterations to evaluate 95% confidence intervals for edge weights and centrality stability coefficients. Results In FAERS, nivolumab (48,330) and pembrolizumab (37,224) generated the highest report volumes. PD-1 inhibitor networks consistently centered on a systemic-gastrointestinal symptom cluster comprising decreased appetite, fatigue, diarrhea, and nausea, whereas PD-L1 inhibitor networks additionally featured prominent hematologic abnormalities. In JADER, interstitial lung disease and adrenal insufficiency emerged as central nodes, reflecting cross-database differences in reporting patterns and physician-initiated reporting practices. Strength and Expected Influence centrality metrics showed good to excellent stability (CS ≥ 0.672 in most networks), supporting the robustness of the core symptom clusters. Conclusion Symptom network analysis effectively delineates core suspected irAE symptoms and inter-system associations for PD-1/PD-L1 inhibitors. FAERS-derived networks emphasize systemic-gastrointestinal syndromes, while JADER reveals pulmonary-endocrine predominance in Japanese populations. These findings demonstrate the utility of symptom network analysis as a complementary approach to traditional pharmacovigilance for characterizing complex suspected irAE patterns across different data sources and populations.