Multilayer perceptron neural network to estimate left ventricular systolic dysfunction in hospitalized heart failure patients using accessible clinical and biochemical variables

Abstract

Introduction: Assessment of left ventricular ejection fraction (LVEF) is essential in heart failure, but echocardiographic measurement is not always available in resource-limited clinical settings.

Objective: To develop a neural network model to predict reduced LVEF using common clinical and biochemical variables in hospitalized patients with heart failure.

Materials and Methods: A cross-sectional study was conducted using a secondary database of 584 patients. The dependent variable was LVEF. A multilayer perceptron neural network model was developed using the following predictors: BNP, troponin, hemoglobin, BMI, age, eGFR, and creatinine. Its performance was compared to a binary logistic regression model, and both were evaluated using area under the curve (AUC), sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and overall accuracy.

Results: The neural network achieved an accuracy of 75.3% in the test set and 73.5% in the training set, with a sensitivity of 72%, specificity of 78.2%, AUC of 0.811, and an odds ratio of 8.07. BNP was the most influential predictor, followed by hemoglobin, troponin, eGFR, age, BMI, and creatinine. In comparison, the logistic regression model performed worse (accuracy: 54.8%; sensitivity: 44.9%; specificity: 62.5%; AUC: 0.65), particularly in negative predictive value (82.9% in the neural model vs. 56%).

Conclusions: The neural network model effectively predicted reduced LVEF using routine clinical variables, outperforming logistic regression and offering a practical alternative in settings lacking immediate access to echocardiography.