Experience of using machine learning to optimize renal replacement therapy: a review

This review examined studies on the use of machine learning tools to optimize renal replacement therapy. Publications were searched using the databases eLibrary, PubMed, and Scopus. The TITLE and ABSTRACT fields were used to select the metadata for the following queries: artificial intelligence AND (dialys* OR hemodialys*) and machine learning AND (dialys* OR hemodialys*). The use of different word forms for key concepts was considered, and the period of publications was not limited. The primary search yielded 669 full-text articles. Three independent experts (i.e., two nephrologists and one machine learning specialist) manually selected articles. Article selection focused on the use of machine learning algorithms, solution of a specific issue of optimization of dialysis therapy, and availability of a model specifically developed for patients on dialysis and not intended for other clinical tasks. Review of articles revealed that predicting dialysis adequacy is the most common dialysis-related task wherein machine learning tools are used. Dialysis adequacy refers to the effective removal of nitrogenous waste products. Post-procedural complications are a potential indirect sign of dialysis inadequacy, the risk of which is what artificial intelligence models are trained to identify. They are also used as basis for selecting therapy for renal anemia and phosphorus–calcium metabolism disorders, as various factors should be considered in these cases, including changes over time. Several studies of machine learning are aimed at predicting survival of patients on dialysis and determining the optimal time to initiate renal replacement therapy in patients with chronic kidney disease. A comprehensive review of the articles showed several limitations, including small sample sizes, insufficient testing during the preliminary preparation and analysis stages, and low reproducibility of many studies. The most relevant studies described the clinical implementation of systems based on machine learning algorithms after testing them on different patient populations.