CD3+ T-cell count prediction for anti-thymocyte globulin treatment monitorization in kidney transplant recipients: a...

Antithymocyte globulin (ATG) therapy is conventionally monitored by measuring peripheral blood CD3+ T-cell counts. However, CD3+ T-cell quantification requires flow cytometry and may not be routinely available in centers. This study aimed to identify clinical and laboratory predictors of CD3+ T-cell depletion and to develop a machine learning model for predicting attainment of predefined therapeutic CD3+ T-cell thresholds following ATG induction therapy in kidney transplant recipients.

Adult transplant patients who underwent kidney transplantation were retrospectively evaluated. ATG doses were administered as induction therapy, and the subsequent peripheral blood CD3+ T-cell counts were obtained. Demographic, anthropometric features, and laboratory results were recorded. Statistical and machine learning methods were applied to identify parameters predictive of CD3+ T-cell counts. Prediction models were developed and compared to the logistic regression model to estimate the likelihood of reaching certain CD3+ T-cell count cut-offs after transplantation to evaluate and monitor ATG effect.

In the analysis of 397 transplant patients of 99.2% grafts from living donors, 57.2% of patients achieved the predefined day-1 CD3+ T-cell < 50 cell/μl threshold, while 57.5% of patients reached the predefined day-2 CD3+ T-cell < 30 cell/μl threshold. The machine learning model's performance in predicting target threshold attainment resulted in ROC-AUC values of 0.75 and 0.80 for Day 1 (test and validation sets), whereas Day 2 predictions yielded ROC-AUC scores of 0.70 and 0.66, respectively. The predictive performance of the machine learning model was superior to logistic regression prediction, and decision curve analysis showed that the model provided clinically meaningful net benefit for decision-making for Day 1 and Day 2 predictions.

The clinical response to ATG treatment for induction immunosuppression in kidney transplant patients may be adequately predicted using low-cost and accessible laboratory tests alongside patient-specific characteristics without the need for CD3+ T-cell quantification, through machine learning prediction models. This study provides an initial framework for further development of more accurate machine learning models with higher prediction power for clinical use in ATG effect prediction.