Introduction:

Heart failure is a common, progressive, and life-threatening condition whose risk is often overestimated. Effective tools are required to discriminate the risk and therefore a system based on the assembly of neural networks was trained for this purpose.

Objective:

To present the results of the training and internal validation of a system based on a set of artificial neural networks for the prognosis of one-month mortality in patients hospitalized for acute heart failure, and to compare the results of each of the individual networks developed and four set systems simple voting and AdaBoost.

Materials and method:

From a cohort of 462 patients diagnosed with decompensated heart failure, 11 networks were trained and then assembled using four systems: simple voting, two systems weighted by operating characteristics (predictive values and likelihood ratios) and Boosting. Operating characteristics for the 30-day prognosis of death were calculated and compared with two clinical rules and logistic regression applied to the same population.

Results:

The various ensemble methods had a better prognostic performance than each of the networks that composed them. Voting weighted by predictive values performed best, with an accuracy of 89.0% (95% CI: 82.6-93.2%) although the results’ confidence intervals overlapped.

Conclusions:

The ensemble of neural networks through voting weighted by predictive values showed an adequate performance for predicting 30-day mortality in acute heart failure.

        · abstract in Spanish     · text in Spanish     · Spanish ( pdf )