Transcriptomic stratification value of a dual-axis myeloid imbalance framework in sepsis: an integrative study based on a discovery cohort and external validation cohorts

Background Sepsis is a life-threatening syndrome characterized by infection-induced dysregulation of the host response and subsequent organ dysfunction, with marked clinical heterogeneity. Conventional stratification approaches based on single inflammatory markers or clinical severity scores often fail to robustly capture host-response states. Recent studies using public transcriptomic cohorts have suggested that reproducible immune phenotypes and myeloid functional remodeling exist in sepsis; however, substantial variability remains across studies in phenotype definitions, gene-set composition, and clinical interpretability. Identifying a stratification framework from public transcriptomic data that is both biologically interpretable and reproducible across cohorts remains a key challenge in precision stratification of sepsis. In this study, we constructed a dual-axis myeloid imbalance framework composed of an antigen-presentation-associated myeloid program and an interferon/stress-responsive myeloid program using public whole-blood and peripheral blood mononuclear cell transcriptomic cohorts, and evaluated its value for outcome stratification and disease-state discrimination through module-level, gene-level, and cross-cohort validation. Methods Three public transcriptomic cohorts were included. GSE65682 served as the discovery cohort for comparing host transcriptomic differences between 28-day ICU survivors and non-survivors. GSE57065 served as the external validation cohort for comparing septic patients with healthy controls and for assessing early temporal dynamics at 0, 24, and 48 hours. GSE48080 was analyzed as a PBMC-based supplementary cohort providing orthogonal supportive evidence. Two core modules were constructed using fixed representative gene sets: the antigen-presentation-associated myeloid program (MS2_APC) and the IFN/stress-responsive myeloid program (MS3_IFN_stress). A dual-axis index was then derived to quantify the degree of imbalance between the two axes. Based on this framework, we performed phase-space distribution analysis, module distribution comparisons, cross-cohort effect-size integration, representative gene-expression heatmaps, focused volcano plots, dual-axis gene bubble heatmaps, and ROC analyses. Results In GSE65682, 28-day ICU non-survivors showed significantly lower MS2_APC and significantly higher dual-axis index than survivors, whereas MS3_IFN_stress alone did not reach statistical significance, indicating that outcome-related signals were primarily characterized by suppression of the APC-related axis and greater dual-axis imbalance. The phase-space plot further showed that non-survivors clustered more toward a relatively low-APC, highly imbalanced region defined by MS2_APC and MS3_IFN_stress. External validation in GSE57065 demonstrated that septic patients, compared with healthy controls, had significantly reduced MS2_APC and markedly increased MS3_IFN_stress and dual-axis index, supporting stable disease-state relevance of this framework. Cross-cohort effect integration further indicated that reduction of MS2_APC and elevation of the dual-axis index were more consistent across cohorts and comparison settings, whereas GSE48080 provided only trend-level PBMC support with limited statistical strength. Gene-level analyses showed that APC-related representative genes exhibited more coherent negative shifts across both discovery and validation cohorts, whereas IFN/stress-related genes displayed more dispersed patterns. ROC analysis suggested that the dual-axis index had limited discriminatory performance for outcome stratification in GSE65682 but strong discrimination between septic and healthy states in GSE57065. Conclusions We propose and validate a dual-axis myeloid imbalance framework in sepsis. This framework was associated with adverse outcomes in the discovery cohort, showed clear disease-state relevance in an external whole-blood cohort, and further suggested at the gene level that APC attenuation is more reproducible across cohorts than isolated IFN/stress-related variation. This framework provides a transcriptomic perspective for understanding host-response heterogeneity in sepsis, although its clinical predictive performance and bedside applicability still require validation in independent clinical cohorts.