Short-term mortality after intensive care admission in acute lymphoblastic versus predominantly acute myeloid leukemia: a retrospective cohort study using MIMIC-IV with eICU external replication.

Background: Adults with acute leukemia admitted to intensive care have high early mortality. Whether acute lymphoblastic leukemia (ALL) adds prognostic information beyond baseline severity remains clinically important. Methods: We performed a retrospective MIMIC-IV cohort study of 486 first ICU stays from 372 adults with acute leukemia. Primary outcomes were hospital and ICU mortality. ALL was compared with a predominantly acute myeloid leukemia comparator using complementary confounder-only and severity-adjusted models and ICD-derived remission stratification, with directional replication in an independent eICU cohort. Results: ALL admissions were younger and less severely ill at ICU entry. Crude hospital mortality was lower in ALL than in the predominantly AML comparator (23.8% vs. 38.6%; P = 0.010). After adjustment for age and comorbidity alone, ALL showed a non-significant reduction in the odds of hospital mortality (adjusted OR, 0.688; 95% CI, 0.385-1.228). Adding first-day Sequential Organ Failure Assessment (SOFA), which partly overlaps with leukemia-related complications, moved this estimate toward the null (adjusted OR, 0.905; 95% CI, 0.485-1.689). ICU-mortality estimates were also non-significant, although the severity-adjusted ICU estimate crossed the null (confounder-only adjusted OR, 0.690; severity-adjusted OR, 1.076). Because first-day SOFA may lie on the subtype-to-outcome pathway, the confounder-only estimate is reported as the total-association estimate (not conditional on the possible mediator), and all confidence intervals remained wide. Model calibration and discrimination were acceptable. In an independent eICU cohort (n=82), the crude and severity-adjusted ALL estimates for hospital mortality (odds ratios 0.667 and 0.445) were likewise non-significant. Conclusion: In MIMIC-IV, the crude ALL survival advantage narrowed after adjustment and was not statistically significant. In eICU, crude and severity-adjusted estimates were likewise non-significant but moved further from, rather than toward, the null. Wide confidence intervals in both cohorts, the small ALL sample, overlap between first-day SOFA and leukemia-related complications, and unmeasured leukemia-specific factors leave a modest residual subtype effect possible. These are within-ICU associations affected by selection into intensive care; incomplete post-discharge death ascertainment mainly limits longer-horizon interpretation.