Cytokine Signatures Outperform Immune Subsets in Machine Learning Models for Predicting Acute Graft-Versus-Host Disease at Neutrophil Engraftment

Acute graft-versus-host disease (aGvHD) is a major immune complication of allogeneic hematopoietic stem cell transplantation (Allo-HSCT), driven by complex immune-cytokine interactions. This study employed machine learning (ML) algorithms to develop early predictive models for aGvHD using immune and cytokine profiles of Allo-HSCT recipients at the time of engraftment.

Seventy patients with hematological disorders undergoing their first Allo-HSCT were recruited prospectively. Peripheral blood immune subsets and cytokines were analyzed using flow cytometry and ELISA, respectively. ML models, including support vector classifier (SVC), decision tree, and random forest, were trained on 48 features: 34 immune subsets and 14 cytokines.

Patients who developed aGvHD exhibited a reduced CD4⁺/CD8⁺ ratio, lower Tregs, elevated Th1, Th17, cytotoxic natural killer (NK) cells, dendritic cells (DCs), B cell, and increased proinflammatory cytokines (IFN-γ, IL-1β, IP-10, TNF-α, IL-17α, IL-12p70, MIP-1α, MIP-1β, and RANTES). ML models demonstrated excellent predictive performance, with cytokine profiles alone or combined with immune data achieving perfect accuracy (1.00), followed by T-cell (0.96), NK cell (0.93), DC (0.90), and B cell (0.86) models.

Cytokine profiles emerged as superior predictors over immune subsets, supporting their integration into ML-based aGvHD risk prediction. These findings provide a foundation for developing biomarker-guided strategies for early aGvHD detection and management.