Related Experiment Video
Updated: May 28, 2026
A Data-Driven Approach to Quantifying Immune States in Sepsis
Published on: February 7, 2025
Machine learning-driven analysis of serum GDF15 trajectories identifies novel sepsis sub-phenotypes and predicts
Qinxue Wang1, Jiawei Wang2, Yuhan Zhao3
1Department of Geriatric Intensive Care Unit, the First Affiliated Hospital with Nanjing Medical University, Nanjing, Jiangsu, China; Department of Critical Care Center, the First Affiliated Hospital with Nanjing Medical University, Nanjing, Jiangsu, China.
Background:
Sepsis heterogeneity complicates management and prognosis. Growth differentiation factor 15 (GDF15) may offer novel insights into sepsis sub-phenotyping. This study explored serum GDF15 trajectories for sub-phenotyping and prognostic stratification.
Methods:
A multicenter prospective cohort (March-October 2023) enrolled sepsis patients from four Chinese ICUs. Serum GDF15 was measured on days 1, 3, and 7 post-admission. Group-based trajectory modeling (GBTM) was employed for sub-phenotyping. Prognostic differences were analyzed using logistic regression and survival analysis; XGBoost machine learning determined key discriminators.
Results:
This study included 284 sepsis patients, with median age 65 (interquartile range [IQR] 52-75), 66.9% male, and 26.1% 28-day mortality. GBTM identified three distinct GDF15 trajectory subgroups according to levels measured on days 1, 3, and 7: low-level (n = 160, 56.3%), intermediate-level (n = 87, 30.6%), and high-level (n = 37, 13.0%). The 28-day mortality was significantly different across the groups (16.9%, 32.2%, and 51.4%, respectively; p < 0.001). After multivariable adjustment for demographics, disease severity scores, and other key clinical parameters in a series of sensitivity analyses, GDF15 trajectory subgrouping remained a robust, independent predictor of mortality. In the most comprehensively adjusted model, the high-level trajectory was associated with a 7.8-fold increased risk of 28-day mortality compared to the low-level group (95% CI: 3.2-19.1, p < 0.01). XGBoost analysis revealed Sequential Organ Failure Assessment (SOFA) score and lactate as critical discriminators of trajectory sub-phenotypes.
Conclusion:
Machine learning-driven GDF15 trajectory analysis delineates three sepsis sub-phenotypes with divergent prognoses. The high-level trajectory independently predicts 28-day mortality, highlighting GDF15 as a pathophysiologically anchored biomarker for precision management. These findings warrant validation in larger cohorts to optimize sepsis stratification strategies.