An artificial neural network for the prediction of mortality in patients with chronic kidney disease

Abstract

Introduction: Early mortality in patients with chronic kidney disease represents a major health problem, which is why the design of novel prognostic models is a priority. Objective: Design an artificial neural network for the prediction of mortality in patients with chronic kidney disease on hemodialysis. Methods: A prospective cohort analytical study was conducted in patients with chronic kidney disease on hemodialysis during the period from January 1, 2013 to December 31, 2017. A total of 36 attributes were analyzed in 392 patients. The multilayer perceptron was used to design an artificial neural network composed of 12 variables. Finally, the classification table, the discriminatory capacity of the algorithm and the normalized importance of the prognostic variables were evaluated. Results: The artificial neural network presented overall correct classification percentages of a 96.3% in the training sample and a 96.7% in the validation sample. The discriminatory capacity was very good, area ROC of 0.989. The most important normalized predictors of mortality were cardiovascular disease, albumin, and sepsis. Conclusions: The artificial neural network contributes to the stratification of mortality risk of patients with chronic kidney disease on hemodialysis. The model has good discriminatory capacity and indicators of statistical effectiveness. The prognostic variables identified are easy to determine and interpret, which is why it is considered a predictive tool with useful implementation in medical decision making in the clinical setting.

Keywords

• aprendizaje automático
• enfermedad renal crónica
• hemodiálisis
• inteligencia artificial
• mortalidad
• redes neuronales de la computación.