Single and multiple time-point prediction models in kidney transplant outcome

This study predicted graft and recipient survival in kidney transplantation based on the United States Renal Data System (USRDS) by regression models and artificial neural networks (ANNs). Four categories of models were examined: single time-point models (logistic regression and single-output ANNs) versus multiple time-point models (Cox models and multiple-output ANNs). The results showed that kidney transplant survival could be predicted by regression models and ANNs, with good prediction discrimination and model calibration. In AUC (area under the ROC curve) measurements, logistic regression ranged from 0.71 to 0.81; single-output ANNs from 0.72 to 0.82; Cox models from 0.65 to 0.78; multiple-output ANNs from 0.59 to 0.82. In Hosmer-Lemeshow test, p values for logistic regression and single-output ANNs were above 0.05 indicating good calibration, while p values for Cox models and multiple-output ANNs were in general below 0.05, indicating poor calibration. This study established for the first time that: 1) Single time-point models may be more appropriate than multiple time-point models if the predictors have significantly different effects on short-term versus long-term survival. 2) Logistic regression is able to achieve comparable performance as ANNs if there are no strong interactions or nonlinear relationships among the predictors and the outcomes. 3) Multiple-output ANNs may perform poorly when a large proportion of observations are censored. 4) Appropriate baseline survivor function should be specified for Cox models in order to achieve good model calibration when the clinical decision support is designed to provide exact predicted survival rates instead of a relative percentile in survival.