Systems Bioengineering of Septic Shock Metabolism: Citrulline, β-Hydroxybutyrate and Plasma Biomarker-Based Phenotyping

Background: Although advances in critical care have improved short-term outcomes, sepsis survivors continue to face substantial chronic morbidity and impaired long-term survival. Conventional threshold-based tools such as the Sequential Organ Failure Assessment (SOFA) and Modified Early Warning Score (MEWS) show moderate and variable discrimination across cohorts. Reported areas under the receiver operating characteristic curve (AUROCs) must therefore be interpreted in relation to the population, prediction horizon, and outcome used in each study rather than as direct head-to-head comparisons. Objectives: This review evaluates how artificial intelligence (AI) could be linked with dynamic plasma metabolites, particularly citrulline and β-hydroxybutyrate (3-HB), to support biologically informed sepsis phenotyping, while critically examining mechanistic evidence, clinical limitations, and translational readiness. Data Synthesis: Machine-learning and natural language processing architectures have shown promising discrimination in many early-detection studies, with pooled AUROCs near 0.87 and reported prediction windows extending to 48 h. However, performance estimates vary with cohort composition, outcome definition, and validation design, and they should not be ranked against unrelated biomarker studies. Human sepsis studies generally associate low or persistently low citrulline with impaired intestinal function and organ injury, but no sepsis-specific decision cutoff has been externally validated. For 3-HB, an AUROC of 0.8429 for septic liver injury was derived from a cohort of 57 patients and has not been shown to add value beyond routine liver tests or illness-severity measures. Murine experiments provide mechanistic hypotheses for ketone-mediated organ protection, but model-specific and sometimes opposing nutritional effects limit direct translation. These metabolites are therefore best considered candidate longitudinal features for multimodal phenotyping rather than stand-alone clinical triggers. Conclusions: Biologically informed algorithmic surveillance is a promising direction, but clinical implementation requires prospective serial sampling, explicit adjustment for renal, hepatic and nutritional confounders, head-to-head comparison with routine markers, and external validation of calibration and clinical utility. Until these requirements are met, citrulline and 3-HB should support research phenotyping rather than direct treatment selection.