#2277 Artificial Intelligence for predicting heart failure in chronic kidney disease: analysis of a 5000-patient cohort

Le résumé fourni par la source

Abstract Background and Aims Chronic kidney disease (CKD) and heart failure share a complex, bidirectional interaction, and are therefore an area of crucial interest in nephrology due to their prevalence and effect on mortality. It is basic to identify the risk of heart failure in patients with CKD as soon as possible to improve treatment outcomes and complications. It is important to note that while there is wide experience in prediction of heart failure, the use of models based on artificial intelligence is practically non-existent with this objective in mind. Based on the experience of our group in the prediction of mortality in patients with CKD and in CKD progression with machine learning algorithms, this study aims to predict the onset of heart failure in CKD patients. This study addresses this issue using machine learning techniques to offer new opportunities for accurate prediction of cardiac disease. This approach facilitates the individualization of risk assessment and allows the implementation of early and personalized interventions that can significantly influence the management of both diseases. Method Design: Retrospective observational study of a historical cohort from the Registry of Renal Patients of Catalonia (RMRC) and the Catalan Agency for Health Quality and Evaluation. Sample: 5000 patients with CKD, selected for having the minimum number of missing data. Follow-up: 10 years, from January 2010 to December 2020. Inclusion criteria: patients older than 18 years with CKD. Variables: 333 variables: a) Age, gender, weight, height (4); b) Status on the transplant waiting list (2); c) Renal replacement therapy (3); d) Diagnoses (ICD-10) excluding group I50 diagnoses, used as the outcome (146); e) Laboratory variables (78); f) Pharmacological treatments (100). Label and models: The algorithm used was the Light Gradient Boosting Machine (LGBM), with which four models were trained to perform a binary classification to predict whether a patient would develop heart failure in a period of 3, 4, 5 or 6 years. Label 0 was used to represent the absence of heart failure, while label 1 was used to represent the presence of heart failure. The methodology used to train the models was as follows: 1. Pre-processing of the data to manage missing data and data errors. 2. Creating different datasets, one for each prediction horizon. 3. Training and evaluating the LGBM models (with 5-fold CV). Results Age: 63 ± 13 years, Gender: 66% male and 34% were female. Different prediction horizons were tested, and the best results were obtained for 6 years where was achieved an area under the curve of 0.86 and an accuracy of 0.76. The 5 variables with major relevance according to SHAP values (SHapley Additive exPlanations) and in this order are: age, CKD diagnosis, hypertension, haemoglobin, and lymphocytes. The results presented in Fig. 1 and Table 1 correspond to the mean obtained for the 5-folds of the Group Cross-Validation. Conclusion –Advanced artificial intelligence algorithms present promising results for the prediction of heart failure in CKD patients. –This enables early, personalized, and more effective treatments for CKD patients.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.