Prediction Models for Kidney Graft Outcomes Integrating Pre-Transplant Non-HLA Antibody Panels

초록

Non-HLA antibodies may influence graft outcomes beyond conventional HLA mismatch and donor-specific antibodies. We aimed to develop pre-transplant prediction models for rejection and long-term graft dysfunction by integrating non-HLA antibody panels with clinical variables. We analysed 384 kidney transplant recipients using demographics, clinical data, and immunologic variables including HLA mismatch, HLA-DSA status, and > 40 non-HLA antibodies. Primary outcome was rejection (ABMR, IFTA or TCMR). Graft dysfunction was defined as serum creatinine ≧ 1.8 mg/dL at 5 years. After PMM-based imputation, data were split 7:3 into training and test sets. Elastic Net logistic regression (α = 0.25) with cross-validated λ selection was applied. Model discrimination (ROC, AUC, AUPRC) and Firth-penalized calibration were performed. The rejection model achieved a test AUC of 0.825 (95% CI 0.742–0.907) and AUPRC 0.880 (sensitivity 0.870, specificity 0.700). In Firth regression, the composite risk score was strongly associated with rejection (β = 4.67, p < 3.8 × 10−9). Donor type, HLA mismatch, HLA class II DSA, younger age, and antibodies including LMNA and PLA2R increased risk, whereas Fibronectin and LMNB1 were protective. The graft dysfunction model showed a test AUC of 0.82 (95% CI 0.74–0.91) and AUPRC of 0.443, exceeding the random expectation (0.19). The risk score remained significant in Firth regression (β = 2.99, p < 0.0001). PTPRN and PECR antibodies were major risk contributors, while LMNB1 and CXCL11 were protective. Models incorporating pre-transplant non-HLA antibody panels with clinical variables demonstrated strong performance in estimating rejection and graft dysfunction risk, and key non-HLA antibodies contributing to model prediction were identified. With further validation, these models may support risk stratification and personalized post-transplant surveillance strategies in kidney recipients.