| アブストラクト | BACKGROUND: Thrombocytopenia during immune checkpoint inhibitor treatment in lung cancer may be coded as immune, marrow suppression, coagulation, or non-specific. We tested whether coding-defined report phenotypes differed across co-reported treatment contexts. METHODS: We analyzed FAERS (2015Q1-2026Q2) and JADER through Japanese fiscal 2025Q4 (31 March 2026). Strict primary lung cancer reports were grouped as ICI without chemotherapy (C1), ICI plus chemotherapy (C2), or chemotherapy without ICI (C3). Firth models compared mutually exclusive immune-like and marrow-like anchors. Complementary analyses evaluated the broad cohort, treatment date compatibility, C3 small-cell influence, phenotype variants, onset windows, treatment-variable ablation in elastic-net report phenotype models, and a 120-report random stratified blinded review with independent ratings, third-person adjudication, inverse-selection weighting, and bootstrap uncertainty. RESULTS: FAERS included 1,362 immune-like or marrow-like anchors. Relative to C1, adjusted odds ratios (ORs) for immune-like versus marrow-like phenotype were 0.111 (95% CI, 0.069-0.173) in C2 and 0.014 (0.006-0.029) in C3; C3 was exploratory because only eight reports were immune-like. In 2,835 broad cohort reports, adjusted immune-like composition was 13.0% (95% bootstrap interval, 11.0%-15.1%) in C1 and 4.7% (3.4%-6.2%) in C2. Treatment date and stricter phenotype analyses were directionally consistent. JADER reproduced the C2 versus C1 direction but showed calibration shift; treatment-variable ablation reduced cross-database discrimination. Reviewers agreed on 111/120 reports (92.5%; 95% CI, 87.5%-96.7%; Cohen's kappa, 0.897; Gwet's AC1, 0.901). In the locked 927-report frame, design-weighted final categories were 23.4% immune probable, 37.3% immune possible, 13.6% indeterminate, and 25.8% non-immune probable. Among 100 determinate reports, weighted AUROC was 0.892 (0.827-0.948), PR-AUC 0.956 (0.922-0.982), and calibration-in-the-large intercept 2.054 (1.707-2.432), indicating underprediction. CONCLUSION: Coding-defined thrombocytopenia report composition differed across treatment contexts. Blinded review supported reproducible structured category assignment and score-based report ordering, while calibration shift and treatment-variable ablation defined important limits. Findings support pharmacovigilance case prioritization and hypothesis generation, not clinical diagnosis, incidence estimation, patient-level etiologic attribution, or causal inference. |
|---|
