الفهرس 
Research OverviewOnly 14 pages are availabe for public view
 |  | from | 32 |  |  |
| | | from | 32 | | |
Abstract
The research aims to use of a proposed model for the prediction that involves large number of multiple simple two-layered neural networks connected in a complete parallel structure called a macro-neural model. Macro-neural networks are a type of feed-forward neural networks in which the network architecture is composed of many interconnected simple networks. The type of the devised architecture can affect both generalization and convergence as well as provide a solution to several known problems in conventional designs such as over-fit problem and the unknown size of the hidden layer. Convergence to a global minimum and reliability are important properties of feed-forward neural networks. These properties are studied by examining different training examples for survival rate prediction in lung cancer. A modification to the survival perdition is presented to over-come problems inherent in the proposed model. The performance of the proposed model is examined with respect to the conventional methods. Detail error analysis is presented and several data sets are analyzed and compared. In most of the comparisons carried out the proposed model proved to be superior to conventional survival rate prediction methods. However, the generalization of the obtained results cannot be claimed for any problem as the effectiveness of the proposed model has to be examined for more diverse problems before making the conclusion that the proposed model is indeed superior for any problem. The main improvements that has been achieved include the increase of the prediction accuracy to 77.8%, with area under the ROC close to 0.829. Such accuracy is clearly better than the one recorded for the typical statistical regression analysis which was only 71% with area of 0.723. The prediction even appears to be more accurate than previous studies employing similar data with standard three-layer neural network which deteriorates fast as the period of prediction in the future increases.