An explanation facility for a neural network trained to predict atrial fibrillation directly after cardiac surgery

An explanation facility is presented that makes it possible to elucidate why a neural network assigns a particular class label to a case. The explanation facility is developed as a remedy to the black-box problem of neural networks which impedes their use for (clinical) decision support. The method is based on ideas from feature selection. For a classified case, variables are identified that can possibly change the classification of the case. These variables are subsequently ranked according to their importance for the classification of the case. The method is evaluated on a classification problem in cardiology: the prediction of atrial fibrillation directly after cardiac surgery.