Cytokine Signatures Outperform Immune Subsets in Machine Learning Models for Predicting Acute Graft-Versus-Host...

Area of Science:

• Immunology
• Transplantation Medicine
• Computational Biology

Background:

• Acute graft-versus-host disease (aGvHD) is a significant immune complication following allogeneic hematopoietic stem cell transplantation (Allo-HSCT).
• Complex immune-cytokine interactions drive aGvHD pathogenesis.
• Early prediction of aGvHD is crucial for patient management.

Purpose of the Study:

• To develop early predictive models for aGvHD using machine learning (ML).
• To identify key immune and cytokine biomarkers for aGvHD prediction at engraftment.
• To compare the predictive performance of immune subsets versus cytokine profiles.

Main Methods:

• Prospective recruitment of 70 patients undergoing Allo-HSCT.
• Analysis of peripheral blood immune subsets via flow cytometry and cytokines via ELISA.
• Training of ML models (SVC, decision tree, random forest) on 48 features (34 immune, 14 cytokine).

Main Results:

• aGvHD patients showed altered immune profiles (reduced CD4+/CD8+ ratio, lower Tregs) and elevated pro-inflammatory cytokines.
• ML models achieved excellent predictive performance.
• Cytokine profiles alone or combined with immune data yielded perfect accuracy (1.00), outperforming immune subset-specific models.

Conclusions:

• Cytokine profiles are superior predictors of aGvHD compared to immune subsets.
• Integration of cytokine profiles into ML models enhances aGvHD risk prediction.
• Findings support biomarker-guided strategies for early aGvHD detection and management.