Linearisation of feed-forward artificial networks to study input importance

Sammanfattning: In the use of articial neural networks (ANN) for real-life applications there is a need for determining the importance of each input variable for the decisions of the articial neural network. By comparing more complex ANN:s with the simple perceptron, a basic method to determine some form of importance is found. This is tested on three dierent datasets of which one is based on a real life dataset of acute coronary syndrome (ACS). The results from these tests indicate that the method works.