Application Of Regularized SOM Clustering Algorithm In Disease Diagnosis

The intellectual disease diagnosis studies the characteristics of various diseases automatically learned from clinical medical big data,which investigates and identifies data from patients automatically to the assistance for the judgment of doctors to reduce misdiagnosis.The main work is as follows.First of all regularized penalty items are added to the competition rules of winning elements of SOM(Self Organizing Maps)to solve the over-fitting problem in the SOM prediction.Over-fitting problem is avoided also by the regularized operators of tikhonov algorithm as well as the use of penalty items to restrict the weight of output neurons and the acquisition of balance between target function and penalty items.Next,information entropy is introduced to ITR_SOM algorithm and.the EITR_SOM algorithm in proposed,to solve the problem of lacking of interpretability in the SOM fitting process.It is proved by comparative tests that the thought of algorithm improvement in this thesis is feasible in improving cluster accuracy of SOM and in the prevention of over-fitting problem.The algorithm mentioned has an high increase in accuracy in the tests towards standard&irregular data cluster,especially the standard of entropy has a notable decrease like 2/3 in the ITR-SOM algorithm as well as more than 3/4 in the EITR-SOM algorithm.Finally,the proposed algorithm is applied to the diagnosis of intelligent diseases through the diagnose prediction of heart disease and mammary cancer which are compared with BP algorithm and DBN algorithm.The algorithm improves the training algorithm by nearly 10 percentage points when the training data is less,as well as that the entropy is kept in a same level,which speaks for that the improvement effect is very significant.