Predicting the Need for Hemodialysis in Peritoneal Dialysis Patients with at Least One Episode of Peritonitis Using Artificial Intelligence

Summary Transfer to hemodialysis is a major cause of permanent discontinuation of peritoneal dialysis (PD) treatment. Peritoneal infection is a common cause of discontinuation of peritoneal dialysis and permanent transfer to hemodialysis (HD). Accurately predicting the approximate date of transfer allows for the preparation of a vascular access under optimal conditions. Accurately predicting the date of transfer is therefore of clinical importance. The goal of this study is to develop an artificial intelligence model that predicts, based on a patient’s history of peritoneal infection, whether the patient should be transferred to hemodialysis. Our approach is based on a neural network learning method, specifically a Long Short-Term Memory (LSTM) recurrent neural network (RNN); the training set, representing 80% of the dataset—that is, 9,934 patient sequences—was used to optimize the model’s parameters. The test set (10%, or 1,242 sequences) was kept strictly separate from the other sets and reserved exclusively for the final evaluation of the model’s performance. The monitoring of patients’ conditions over time was based on peritoneal infections. Recurrent neural networks are particularly well-suited to this type of prediction based on time-series data. This study highlighted the predictive performance of such a model on this type of data. Introducing a bias into the training process helped limit false-negative predictions. Our data, drawn from the French-Language Peritoneal Dialysis Registry (RDPLF) database, included 15,800 patients who had experienced at least one episode of peritonitis, with the aim of predicting the likelihood of transfer to hemodialysis following this complication. With our model, the false-negative rate—the proportion of patients who were transferred to HD within 6 months even though the model did not predict it—was 5.3%.