The Authors’ Reply: Correspondence: The First Asian Kidney Transplantation Prediction Models for Long-Term Patient and Allograft Survival

We appreciate the insightful comments from Dr Takamasa Miyauchi and colleagues. We completely agree our model could be improved in many ways. The calibration power of our model was adequate, with C-statistics for predicting death-censored graft survival and patient survival of 0.69 and 0.64, respectively.¹ C-statistics of 0.60–0.75 reflect possibly helpful, but not the best, discrimination power.²

Half of our study data was based on living donor kidney transplantation (LKT), in which there is low variation among input variables such as serum creatinine, panel reactive antibody, and cold ischemic time. To improve the discrimination power, more information regarding LKT would be needed. Data regarding preexisting donor-specific antibody, desensitization protocols, ABO incompatibility, donor total kidney volume from pre-transplantation evaluation, and perioperative allograft biopsy, should be collected and included in the future models.^3-6 Multinational Asian collaborative studies with countries such as Japan and South Korea where LKT predominate, together with countries in the region that continue to have high rates of deceased donor kidney transplants, may help to increase the clinical relevance and generalizability of future prediction models.

Post-transplantation details are also important for predicting long-term outcomes, as noted by Dr Miyauchi and colleagues. However, one of the main objectives in developing our model was to answer to the question most commonly asked by patients before hospital discharge: “How long will my new kidney last?”⁷ Our model provides a mechanism to answer this question in an evidence-based and timely way. Nevertheless, a clinician should emphasize that the survival probabilities from our model are “average” values based on pre-transplantation data, which are subject to change, and also depend on post-transplantation complications. The ideal prediction model should dynamically incorporate post-transplantation clinical course outcome data using artificial intelligence and machine learning.

As mentioned in the Discussion section of our article, each country has different practice patterns in the care of kidney transplantation patients. In addition, different cultures and healthcare systems likely also contribute to allograft and patient outcomes.⁸ Our models had superior predictive ability in our study population than those derived from kidney transplant recipients in the USA. For this reason, each country or region should ideally develop their own models and test against other available models to determine the best prediction strategy for their patients.