Machine learning in kidney transplantation: a new possibility for the graft survival prediction
Giuseppe Ietto1, Linda Liepa1, Mirco Gallazzi2, Marika Morabito1, Marta Ripamonti1, Federica Masci1, Elia Zani1, Domenico Iovino1, Cristiano Parise1, Matteo Zanchetta1, Dorotea Confalonieri1, Giulio Carcano1.
1General, Emergency and Transplant Surgery Department, ASST-Settelaghi and University of Insubria, Varese, Italy; 2Informatic and communication Department, University of Insubria, Varese, Italy
Background: One of the major open questions in the field of kidney transplantation is the graft survival, and therefore life expectancy of the receiver. In the last decades the information available about the recipient, the donor and the organ transplanted have increased considerably. In addition, the knowledge about transplantation has grown, so it could be possible to make a more accurate prediction about the outcome. In the current state of the art, these analyzes are conducted by applying traditional machine learning techniques. However, these algorithms require a large and precise amount of data. Our aim is to create a new machine learning technique to predict our patients’ life expectancy and organ durability without requiring complete information and huge amounts of data to be able to produce reliable results.
Methods: To create the first algorithm, we selected only the data that by their nature could allow to compare the patients who received the kidney transplant, such as the pre and post operative creatinine value, the years of dialysis, the Charlson score etc., relating them to the age and Charlson score of the donors.
Results: Our study cohort consists of 362 patients, who underwent kidney transplantation in Ospedale di Circolo, Varese, from 2013 to 2021. Due to the incomplete nature of the data, in order to trace a patient's life expectancy, the distance between two patients at a time is calculated considering only those features where in both patients the value is not null, otherwise the feature is eliminated for that comparison. Based on this value the distance is then calculated to find the 10 most similar patients. Finally, a post-transplant kidney longeviness is calculated. A keynote from this work concerns the predictive limit of years of post-transplant organ survival. Specifically, the maximum number of years we can currently predict is 9 years given that the data we are working on have a limited range. Data processing is still ongoing, due to the limited number of the statistical sample.
Conclusions: From the preliminary results obtained, the method proposed in our study shows the potential of a system capable of obtaining an approximate but adequate estimate of the longevity of the transplanted organ. It can also be considered a starting point for the application of advanced deep learning techniques for the prediction of patients life expectancy and organ durability, which could integrate other scores already validated. Further studies are needed to improve the algorithm.
right-click to download