Discovery of Risk Factor Control Targets Using Machine Learning and Network Analysis for One-Year Renal Allograft Dysfunction

초록

Purpose: Early identification of graft loss risk and timely therapeutic intervention are crucial for preventing late renal allograft failure and improving long-term graft function. This study was conducted to create a sparse model capable of predicting the one-year renal allograft dysfunction and to build a factor network suggesting risk control targets. Methods: development data were constructed using the Korean Organ Transplant Registry (KOTRY). The XGBoost algorithm was trained to predict the model outcome with 112 features, and the relevant factors were selected. The statistical significance of factors was calculated using multiple logistic regression for the development data. The model outcome was one-year eGFR < 45 mL/min/1.73 m2, and model performance was measured using AUC, sensitivity, and specificity. The clinical importance of the model outcome was assessed using long-term graft survival and rejection-free survival. The factor network was built using inter-factor partial correlations and the statistical significance of each factor. Results: The model achieved an AUC of 0.82, a sensitivity of 0.8, and a specificity of 0.8 using seven pre- or peri-transplantation factors. Model prediction was significantly associated with a five-year survival of graft and rejection-free survival. Post-transplantation stay and discharge eGFR ≥ 88.0 were the most prominent risk and preventive nodes on the network, respectively. Donor age and discharge eGFR < 59.8 had a high impact on model prediction and could be effective risk control targets for their multiple connections to other risk nodes. Conclusions: One-year renal allograft dysfunction could be predicted early after transplantation. The long-term outcomes of kidney transplantation might be improved by preemptive measures on donor age, kidney function at discharge, and post-transplantation stay.