Introduction: Acute Graft-versus-Host-Disease (aGVHD) is a major immune complication following allogeneic hematopoietic stem cell transplantation (Allo-HSCT), initiated by conditioning regimen-associated tissue damage. It involves the complex interplay of immune cells and cytokines. Our study aims to leverage machine learning (ML) algorithms on the immune and cytokine profile of Allo-HSCT recipients to develop biomarker-based classification models to predict the onset of aGVHD at the time of engraftment.

Methods: Seventy patients diagnosed with hematological disorders who had undergone Ist Allo-HSCT were recruited from All India Institute of Medical Sciences, New Delhi, India. Peripheral blood was collected from the patients at the time of engraftment, and the immune cell subtypes and cytokine profiles were analyzed using flow cytometry and ELISA respectively. The individual cell counts were then processed using basic ML models, including support vector classifier with RBF kernel, Decision Tree, and Random Forest, chosen for their mathematical simplicity and feature importance advantage of Decision Trees and Random Forests. Various data settings were utilized in the study: combined immune and cytokine counts, only immune cell counts, only cytokine counts, only T-cell counts, both T- and NK-cell counts, only dendritic cell counts, and only B-cell counts. These configurations were selected to investigate how different data sets impact the prediction of aGVHD before its onset.

Results:

At the engraftment flow cytometric analysis of reconstituted lymphocytes in patients who developed acute GVHD revealed that there was a remarkable increase in the decrease in the ratio of CD4+/CD8+ T-cell (p: ≤0.0001), Tregs (p: ≤0.0001) with an increase in the cytotoxic regulatory NK-cell (p: ≤0.0001), dendritic cells (p: ≤0.0001) and B-cell (p: ≤0.0001). The levels of pro-inflammatory cytokines (IFN-γ, TNF-α, IL-1β, MIP-1α, IL-17α), and Th17- and Th1-cells were elevated with consequent decline of the levels of anti-inflammatory cytokine IL-10, and Th2-, Th9-, and Th22-cells.

Machine learning based on 40 parameters [all immune cell subsets n=34 and all cytokines (n=6)]. The correlation heat map shows a higher correlation of aGVHD with the cytokine profile with or without immune cells (accuracy: 1), T-cell with or without NK-cell (accuracy: 1) than for any other individual cell [NK cell (accuracy 0.93), dendritic cell (accuracy: 0.86), and B cell (accuracy: 0.86)]

Conclusion:

The current models classify perfectly, indicating the potential for a machine learning (ML) algorithm in predicting the onset of aGVHD. However, a study with a larger sample size is required to validate these classification models and mitigate the risk of overfitting observed due to the consistently high performance.

The study also highlights the potential of cytokine profiles as a viable alternative to T-cell counts, as evidenced by the correlation heat map and classifier models. These findings provide valuable insights into dataset requirements and future directions for integrating ML models into aGVHD prediction.

Disclosures

No relevant conflicts of interest to declare.

This content is only available as a PDF.
Sign in via your Institution