KASABOV, Nikola Kirilov,FUTSCHIK, Matthias Erwin,SULLIVAN, Michael James,REEVE, Anthony Edmund
申请号：
ES03788512
公开号：
ES2590134T3
申请日:
2003.08.15
申请国别(地区):
ES
年份:
2016
代理人:
摘要：
A computer-implemented method to support a medical decision, characterized by comprising: providing a three layer system comprising: a. a first layer comprising: i. a first predictor module operates on gene expression data using microarray evolutionary fuzzy neural network ii. a second predictor module operates on clinical information using a Bayesian classifier b. a second layer comprising the following class: i. A class and ii. Class B and c. a third layer consisting of an output to provide a combined output by combining the output of all class elements of the second layer, wherein: said output of the third layer is fully connected to the elements of class second layer through connection weights α and α 1 all class elements of the second layer are completely connected with the elements of the first layer module through connection weights ß1,1- SS1, SS2 and 1-beta2 and said connection weights are such as to minimize an error of said combined output so that said combined output has a higher precision than the output of Class A or the output of Class B individually, wherein: the parameter values ß1 , beta2, and α are evaluated and quantified for the three layer system using an exhaustive search method according to the steps: i. creating said first predictor module and said second predictor module through training, testing and optimization parameters ii. for each value of beta1 and beta2 for each value and for each value of α, assaying accuracy three-layer system for the entire set of data and iii. choosing the parameter values SS1, SS2, and α providing the strongest precision three-layer system parameter values SS1, SS2, and α are evaluated and quantified for the three layer system according to a method of specialization statistical basis, where each output class of each module is weighted with accuracy class normalized calculated for this module through all modules in the system, where the values of continuous output for output class are multiplied by the we
