Neural network model to predict 14-day dialysis in patients with acute renal failure.

OBJECTIVES: To develop and evaluate a multilayer perceptron neural network model to predict the need for 14-day dialysis in hospitalized patients with acute renal failure. MATERIALS AND METHODS: Analytical, longitudinal study using an international secondary database available in the Dryad scientific and medical data repository (https://datadryad.org) of patients with acute kidney failure; 42 clinical and laboratory variables were included. The database was divided into training (69.20%) and test (30.8%). An artificial neuronal network was trained with a hidden layer of 9 neurons (hyperbolic tangent) and output with Softmax function, using cross entropy as loss function. Model performance was evaluated by accuracy, area under the curve (AUC) and classification metrics. RESULTS: There were included 4985 patients. The model had a cross-entropy error of 64,582 (training) and 60,260 (testing), with error rates of 0.7% and 1.0%. The key neurons were H1:2, H1:3, H1:6, and H1:7, with H1:3 (previous dialysis, SOFA, creatinine) standing out. The AUC was 0.995, with accuracy of 99.30% (training) and 99% (testing). The pseudo-probability distribution showed high confidence in classification, and the gain curve identified almost 100% of positive cases in the 10% of highest risk. CONCLUSIONS: The multilayer perceptron network showed high accuracy in dialysis prediction in acute renal failure patients, supporting clinical decision making. Validation in external cohorts is recommended.