Dynamic prediction of renal survival among deeply phenotyped kidney transplant recipients using artificial intelligence: an observational, international, multicohort study

For individuals with end-stage renal disease, kidney transplantation is the best treatment, in terms of quality and quantity of life and costs for society. Considering the organ shortage worldwide, the allograft failure is a real burden in chronic kidney disease, contributing to the growing number of the seven million individuals with end-stage renal disease. For these reasons, health agencies such as the Food and Drug Administation and the European Medicines Agency, have emphasized the need for developing a prediction model that would accurately predict the risk of allograft failure, which would allow clinicians to continuously enhance their prognostic judgements, and adapt the patient monitoring and management.

However, developing such model is a difficult task because of the high heterogeneity of kidney recipients and number of potential risk factors. This requires the construction of datasets with thousands of patients, long-term follow-up, with the collection of histological, immunological, clinical and functional data. The goal of the study was thus to create deeply phenotyped cohorts of kidney recipients, and to apply robust prediction model to the data.

This ambitious project was carried out by a research team from the Institut National de la Santé et de la Recherche Médicale (Inserm), led by the Professor Alexandre Loupy. The need to deal with diverse data and complex statistical methods called for the creation of international consortium with experts, including nephrologists, pathologists, epidemiologists, immunologists, statisticians and mathematicians.

« The expertise of our consortium depends on the various and complementary skills of researchers, physicians and statisticians that constitute it » states the Professor Alexandre Loupy, senior author of the study, « This multidisciplinarity is unusual in research, but represents a major advantage, because we can easily interact with the adequate people. »

To adopt the best strategy in the development of dynamic prediction model, the consortium conducted a literature review in kidney transplantation and noticed that to date, no prediction models were taking into account the patient trajectories, that is, the parameters of patient assessed over time, which may hold value for attaining high prediction performances. As a consequence, the consortium have opted for an integrative and dynamic approach, called joint modelling, and collected many parameters repeated over time that may help predicting the allograft failure.

After a long-standing work of several years, more than 13,000 patients from Europe, the US, and South-America were included in the study. More than 400,000 repeated measurements of estimated glomerular filtration rate and proteinuria were recorded. These cohorts of patients, that are unique in the world, permitted to capture the differences in clinical pratice and populations between hospitals and countries, which is a major asset.

The consortium tested numerous models to obtain the best prediction performances. After months of research, their verdict is quite clear : the performances of their final model, called DISPO (dynamic, integrative system for predicting outcomes in kidney transplantation) are the highest ever reached in kidney transplantation.

« The computational methods we used were very time-consuming, which obliged us to acquire powerful computers. » underlines the Doctor Marc Raynaud, epidemiologist and first author of the study, « The next step is to integrate new parameters to the model, in particular immunological data, to try improving the prediction performances. »

For this study, the consortium received the prestigious Leonardo Da Vinci prize at the ESOT congress.

Published on Wednesday 27 October in the Lancet Digital Health, this prediction model is promising because it has been validated in a high number of clinical scenarios and subpopulations, and in several countries and continents, which supports its generalizability. Overall, it may improve the clinicians’ interpretations of the patients, thus enhancing the decision-making and patient management. ​

Read the article

Dynamic prediction of renal survival among deeply phenotyped kidney transplant recipients using artificial intelligence: an observational, international, multicohort study
Marc Raynaud, PhD,  Olivier Aubert, MD, Gillian Divard, MD, Prof Peter P Reese, MD, Prof Nassim Kamar, MD, Daniel Yoo, MPH, Chen-Shan Chin, PhD, Élodie Bailly, MD, Prof Matthias Buchler, MD, Prof Marc Ladrière, MD, Prof Moglie Le Quintrec, MD, Prof Michel Delahousse, MD, Ivana Juric, MD, Prof Nikolina Basic-Jukic, MD, Prof Marta Crespo, MD, Prof Helio Tedesco Silva Jr, MD, Kamilla Linhares, MD, Prof Maria Cristina Ribeiro de Castro, MD, Prof Gervasio Soler Pujol, MD, Prof Jean-Philippe Empana, MD, Prof Camilo Ulloa, MD, Prof Enver Akalin, MD, Prof Georg Böhmig, MD, Edmund Huang, MD, Prof Mark D Stegall, MD, Prof Andrew J Bentall, MD, Prof Robert A Montgomery, MD, Prof Stanley C Jordan, MD, Prof Rainer Oberbauer, MD, Prof Dorry L Segev, MD, Prof John J Friedewald, MD, Prof Xavier Jouven, MD, Prof Christophe Legendre, MD, Prof Carmen Lefaucheur, MD, Prof Alexandre Loupy, MD