Abstract / Summary
Background/Objectives: Infection is a major complication of acute myeloid leukaemia (AML) induction, and analyses relating treatment-emergent infection to survival are prone to immortal-time bias. We characterised baseline and treatment-emergent infection during induction and used bias-aware methods to distinguish a pre-treatment risk signal from a survival-time artefact. Methods: Our study analysed a retrospective single-centre cohort of 91 adults with newly diagnosed AML receiving induction under universal posaconazole prophylaxis (2019–2025) with Cox models, 14- and 30-day landmark analyses and Firth penalised logistic regression. Results: Active infection at diagnosis occurred in 23/91 (25.3%) and treatment-emergent infection in 54/91 (59.3%). A naive Cox model showed an implausibly protective association for treatment-emergent infection (hazard ratio [HR] 0.46, 95% CI 0.21–0.99). In 14- and 30-day landmark analyses (HR 0.59 and 0.60; 95% CIs included 1) the estimate moved toward the null and lost statistical significance; because infection-onset dates were unavailable, landmark analysis only partially corrects this bias, and the independent association between treatment-emergent infection and survival could not be reliably estimated. Active infection at diagnosis showed a fragile, exploratory association with later mortality that was nominally significant in only one of several pre-specified specifications (30-day landmark HR 2.56, 95% CI 1.03–6.37). Conclusions: The apparently protective association of treatment-emergent infection with survival is consistent with immortal-time bias; it is a complication and marker of course, not a survival determinant. Baseline infection is a hypothesis-generating candidate pre-treatment risk marker that requires prospective validation.