Automated Medical Diagnosis Based On Supervised Manifold Dimension Reduction

With the development of computer technology, human society has been in the information age. In the field of medical diagnosis, a large number of high dimensional data will be inevitably encountered. Traditional medical diagnosis technology is mainly influenced by human subjective factors, the diagnosis accuracy rate is lower, the time people spent in diagnosis is longer. Research showed that the diagnostic accuracy rate of automated medical diagnosis technology is higher, and the misdiagnosis rate is lower compared with the traditional method.Currently, automated medical diagnosis technology has not been widely used. Traditional expert system relies on the database for medical diagnosis, which can be understood by medical workers. However, the data collected in the database of expert system is more complex, the redundancy is higher and the accuracy of medical diagnosis is lower. Support Vector Machine Classification method can classify the collected medical information, to some degree, it relieves the limitation of traditional expert system database and improves the accuracy of diagnosis. But Support Vector Machine has a black box effect, which is difficult to interpret the reasoning process and conclude the feature of black box. It can not be intuitive to see the processing process, and intelligibility is not well. Manifold dimensionality reduction algorithm based on machine learning can reduce the dimension of high dimensional data to the low dimensional visual space, and the visualization of the intermediate process is easy to be understood and analyzed by the medical workers. It has guiding significance for medical diagnosis. Many dimensionality reduction algorithms are applied to the field of medical automation. However, dimension reduction algorithm can only reduce dimension of medical information but unable to classify information. This paper will propose the idea which combines dimension reduction and classification to deal with high dimensional medical data. It is shown that the low dimensional mapping plus a linear classification decision making surface is conductive to improve the intelligibility. A large number of medical data is processed by dimension manifold algorithm, which reduces the redundancy of data and improves the accuracy of calculation analysis.This thesis mainly focuses on the manifold dimension reduction and classification techniques. The research work and main research results include:1.The thesis proposes a classification algorithm based on equal metic mapping(it's called SIMBA algorithm for short). SIMBA algorithm is integrated into the supervision information based on ISOMAP algorithm, and the high dimensional medical data is extracted, decision tree algorithm is used to classify the dimension data. Test data can also be implemented by SIMBA algorithm. The visualization of the intermediate process can enhance the intelligibility, which can be understood easily by medical workers. After comparison with other traditional classification algorithms, the proposed algorithm has higher classification accuracy than other traditional classification algorithms.2.This thesis proposes a new algorithm based on LLE algorithm(It's called DLLEA algorithm). The main idea of DLLEA algorithm is:DLLEA algorithm integrates monitoring information based on LLE algorithm and use linear Support Vector Machine algorithm to classify the result of dimension reduction. Test data can also be implemented by DLLEA algorithm. After comparison with other traditional classification algorithms, the proposed algorithm has higher classification accuracy than other traditional classification algorithms.3.This thesis proposes a supervised classification algorithm based on LSE algorithm(It's called SLSE algorithm for short). The basic idea of SLSE algorithm is: SLSE algorithm integrates monitoring information based on local spline algorithm and use KNN classification algorithm to classify the new test sample points. SLSE algorithm which combines LSE algorithm and LDA algorithm, integrates monitoring information based on LSE algorithm. SLSE algorithm generates a clear mapping faction for data points and can easily get the projection of dataset.