Development and external validation of the HCH and HPMS prognostic indices for sepsis: a retrospective model development study using a Multi-Objective Non-Newtonian Fluid optimization algorithm.

BACKGROUND: The pathological heterogeneity of sepsis makes it challenging for traditional scoring systems to balance early-warning sensitivity, dynamic progression characterization, and mechanistic interpretability. Developing a multi-objective optimization algorithm incorporating rheological properties to screen novel composite parameters from routine indicators, thereby constructing a precise sepsis assessment tool with both clinical feasibility and pathological interpretability, offers a new approach to inform clinical decisions and prognosis prediction for sepsis patients. This retrospective prognostic model development and validation study aimed to develop and externally validate novel composite indices for sepsis mortality prediction. METHODS: Using data from the eICU (derivation cohort, n = 3,965) and MIMIC-IV (temporal validation cohort, n = 2,917) databases, we developed the Multi-Objective Non-Newtonian Fluid Optimization (MONNF) algorithm through physically inspired dynamic viscoelastic regulation, multi-target synergistic constraints, and medical knowledge-embedding strategies integrated with key pathological dimensions of sepsis. This algorithm screened novel indicators capturing the pathophysiological interplay among circulatory compromise, coagulopathy, and metabolic derangement. Subsequently, an Under-sampling Synchronous Evolutionary Ensemble (USEE) prediction model was established to validate its cross-center generalizability and clinical utility. RESULTS: The initial phase of MONNF identified the Hypoxia-Coagulation-Hemoglobin Index (HCH; components: lactate, INR, hemoglobin) with a 28-day mortality ROC-AUC of 0.67, superior to SOFA (ROC-AUC = 0.63). The later phase constructed the Hepatic-Pulmonary-Metabolic Synergistic Index (HPMS; components: total bilirubin, PaO₂/FiO₂, bicarbonate, albumin) with a ROC-AUC of 0.64. Risk inflection points of the new indicators revealed critical intervention windows (pnonlinear < 0.001), while Boruta feature importance ranking confirmed their higher decision weight than traditional variables. The USEE model demonstrated optimal predictive performance in independent validation (ROC-AUC = 0.84, PR-AUC = 0.71). CONCLUSION: In this retrospective development and validation study, a novel optimization algorithm based on dynamic non-Newtonian fluid properties successfully uncovered composite indicators that profoundly characterize core pathological features of sepsis. This work provides an innovative solution to simplify clinical stratification and address the tripartite paradox in prognostic assessment: efficiency versus sensitivity versus interpretability.