Confounding-aware disproportionality analysis reveals disease-inherent versus drug-attributable endocrine safety signals of immune checkpoint inhibitors.

Background: Conventional disproportionality analysis assumes comparable background event rates between case and control cohorts. In patients with primary endocrine malignancies treated with immune checkpoint inhibitors (ICIs), this assumption fails: disease-inherent hormonal dysregulation and therapy-related sequelae inflate background event rates and may produce signals misattributed to ICI exposure. No prior study has used tumor-type-specific reference populations to separate drug-attributable from disease-inherent signals in this setting. Methods: Forty quarterly FAERS cycles (Q1 2015 to Q4 2024) were analyzed after deduplication of 2,264,070 reports. In a pooled analysis, 329 ICI-treated endocrine cancer cases were compared with 80,191 ICI-treated non-endocrine cancer controls across six pre-specified endocrine immune-related adverse event categories, with false discovery rate (FDR) correction. The central step was a confounding-aware reanalysis comparing ICI-exposed patients of a given tumor type against a reference cohort with the same malignancy but no ICI exposure, using logistic regression adjusted for age and sex. The thyroid carcinoma association was tested using reporter type, Firth penalized regression, multiple imputation, and a tipping-point analysis. Results: Pooled adrenal insufficiency did not reach the signal threshold after FDR correction (ROR = 1.922; 95% CI 1.051-3.513; q = 0.143). This non-significance concealed two opposing patterns. In ACC, adrenal insufficiency was the strongest uncorrected subgroup signal yet was not associated with ICI exposure after adjustment (OR = 0.58; 95% CI 0.09-1.95); with four exposed events and 12.6% power, this subgroup cannot confirm or exclude an association. In thyroid carcinoma, ICI exposure showed markedly higher adrenal insufficiency reporting odds (OR = 13.33; 95% CI 4.81-31.85), remaining positive under reporter adjustment, Firth regression, and multiple imputation, with a tipping-point analysis indicating only implausibly extreme confounding could nullify it. Conclusion: Pooled pharmacovigilance can misclassify drug-attributable and disease-inherent events. Tumor-type-specific reference modeling separates them: ACC data are most consistent with disease-inherent pathophysiology, whereas thyroid carcinoma shows a robust adrenal insufficiency signal. We interpret the thyroid signal as hypothesis-generating, since differential endocrine surveillance may contribute to its magnitude, and prospective validation is warranted.