Safety signal dilution in cancer trials: Estimation of toxicities through term fragmentation.

Abstract

e24134 Background: To harmonize global clinical reporting, standardized frameworks are used to classify the severity of adverse events (AEs). Despite standardization, a single AE can be represented by multiple distinct codes. When a clinical event is distributed across synonymous terms, the resulting 'signal dilution' may lead to the underreporting of toxicities and a distorted safety profile. We conducted a proof-of-concept analysis to evaluate a potential signal dilution in AE reporting, focusing on renal AEs in oncology trials. Methods: We systematically reviewed phase 3 oncology randomized controlled trials (RCTs) published in NEJM, The Lancet, and JAMA between 2016 and 2021. We assessed AE coding variability using the Common Terminology Criteria for Adverse Events (CTCAE) and Medical Dictionary for Regulatory Activities (MedDRA). Signal dilution was quantified by analyzing: 1) Minimum AE frequency reporting thresholds and 2) Term fragmentation, defined as the number of distinct terms used for clinically related events. Renal toxicity was used as a case study, aggregating related terms to compare pooled vs. reported incidence rates. Results: A total of 94 phase 3 RCTs yielding 107 unique treatment comparisons met the inclusion criteria. Renal AEs were reported in 46 (43%) comparisons. Median reporting thresholds for all-grade AEs were 10% (IQR 5-10) in primary manuscripts and 5% (IQR 1.5-10) in supplements. The most frequent related terms were blood creatinine increased (37.5%), acute kidney injury (37.5%), nephritis (25%), renal failure (16.7%) and renal impairment (14.6%). An average of 1.5 (range 1-4) related terms were used per trial; 29% (14/46) of trials utilized ≥2 related terms. Notably, there was a significant 40% increase in term fragmentation within experimental vs. control arms (mean 1.4 vs. 1.0 terms, respectively; p = 0.002). Pooling related terms increased cumulative renal toxicity incidence by an average of 6.3% (range 0-33.5%) relative to the single highest reported term. Furthermore, 22% (6/27) of trials did not report renal AEs because individual terms fell below the mandatory reporting threshold in the main text, despite the aggregate pooled signal exceeding those limits. Modeling a scenario with one additional related term under trial-specific thresholds indicates that the rate of unreported aggregate signals would increase to 74% (20/27) of trials and mask a cumulative AE incidence of 13.5% (range 0-43.5%). Conclusions: Current oncology trial reporting conventions lead to safety signal dilution. By splitting clinically related toxicities into distinct terms, the true incidence of AEs, particularly in experimental arms, is systematically underestimated. To ensure evidence-based risk-benefit trade-offs, regulatory and editorial standards should mandate the reporting of pooled, clinically related terms and lower the reporting thresholds for toxicities.