Reasearch For The Multivariate Time Series Classification Modeling Based On The Tumor Clinical Index

Lung cancer is one of the common malignant tumors, in recent decades theincidence and mortality of lung cancer are significantly increased. The treatment oflung cancer had no obvious increase in nearly a decade, the main reason is lungcancer biology is complex, high malignant degree, and80%of with lung cancerpatients are diagnosed at advanced stage. Researches on modeling of tumorprogression can help doctors understand the disease and choose appropriate treatment.So the modeling of tumor progression issue has been considered as one of the mostimportant area for researching.The advanced lung cancer patients mostly receive traditional medical care inChina and it’s difficult to make the treatment of the patients coordinated. So, theresearch plan is proposed by our lab is modeling for lung cancer tumor progressionprediction while don’t limit treatment, directly studying the soft index that revealtreatment effects, and combining with the development trend of the disease. Thispaper introduces the characteristics of the detection index from the perspective of thetime series data. Then, research the feature selection scheme multidimensional timeseries in view of the target model. On this basis, specific research and solutions toestablish the classification model of the patient population are proposed. Then add thisdocument refers to the classification module and auxiliary functions on the basis ofthe existing DTAMS. Finally, the research work is summarized and prospected. Themain work of research is summarized as follows:(1) For the nonlinear, relevance, multidimensional and timing of soft targets andsoftening indicators(introduced in section1.3), this paper proposes mutualinformation theory and the between-class class separability criterion theory-basedfeature selection algorithm. To extract the nonlinear relationship between theindicators, mutual information theory is used to convert the multi-dimensional timeseries sample. Then, sort the indicators by the between-class class separabilitycriterion. Next, improve the insufficient of criterion, consider the redundancy betweenthe indicators, and reduce the index while maintaining the characteristics of the data.Finally, convert the feature matrix into a feature vector as input to the model.(2) Based on the feature selection, SVM classification model based on PSOalgorithm is proposed. In order to obtain the optimal model parameters for theclassification model, for the situation of standard PSO algorithm will fall into localoptimal solution, improve PSO model structure, with the maximum possible obtainingglobal optimal solution. In addition, in view of the SVM model supports binaryclassification, combining the theory of binary tree of the knowledge, extends the SVM model consists of two classified to four categories. Compared with the commonlyused classification model, this model has a higher classification accuracy.(3) Participate in DTAMS system development, complete the patientclassification application modules involved in this paper independently. The moduleconsists mainly of medical records analysis, feature selection and classificationmodeling and evaluation functions. DTAMS system experiments provide aconvenient platform for the group members, this all integrated into the platform tomake my own contribution to improve the entire prototype system.In addition, we completed the data management of patients with retrospectivedata, design databases and electronic data. The paper is an important part of theresearch group as a whole study, to some extent, promote the in-depth study of tumorprogression modeling.