Mapper-based topological subtyping of early diuretic response in ICU patients with acute heart failure: a retrospective cohort study with external validation in eICU-CRD

Purpose: Fixed-threshold definitions of diuretic resistance may miss heterogeneous and time-varying response patterns in ICU acute heart failure (AHF). We aimed to identify stable and interpretable subtypes of early loop-diuretic response using a Mapper-based topological framework with external validation. Methods: In MIMIC-IV v3.1, we included adult ICU stays with an AHF diagnosis and first intravenous loop diuretic within 24 h of ICU admission (t0). t0-anchored features captured furosemideequivalent exposure, cumulative urine output and diuretic efficiency index (DEI) over 2 h/6 h/12 h, and hemodynamic/renal context. Mapper graphs were built with two-stage parameter search and evaluated by bootstrap co-membership Jaccard similarity (200 iterations; 80% resampling). Branch labels were assessed for incremental prognostic value for in-hospital mortality using Cox and logistic regression, and for AKI within 48 h (KDIGO creatinine proxy) as a mechanistic endpoint. We performed external validation in eICU-CRD v2.0 using a harmonized pipeline. Results: The cohort comprised 2,373 ICU stays. The optimal Mapper yielded 58 nodes and 393 edges (coverage 1.000) with high stability (mean Jaccard 0.924±0.038; stability score 0.944). Four major branches (A–D) showed marked mortality differences (log-rank p = 1.659 × 10−19). After adjustment, branches C and D had higher mortality risk than A and improved discrimination (AUC 0.712 to 0.737). AKI48h also differed across branches (p = 2.137 × 10−7). In eICU-CRD v2.0, the external validation recapitulated a comparable topology and major-branch structure with a consistent mortality-risk gradient (log-rank p = 5.441 × 10−4). Conclusion: t0-anchored Mapper subphenotyping provides a stable and interpretable representation of diuretic-response heterogeneity in ICU AHF with independent prognostic and mechanistic relevance.