Applications Of SVM In The Classification Of Kidney Stones And Calculating Thermodynamics Parameter

Support vector machine (SVM), developed by Vapnik, as a novel type of machine learning method, is gaining popularity and wide applications due to many attractive features and promising generalization performance. Due to its rigorous theory background and remarkable generalization performance, this dissertation introduced SVM to chemoinformatics and medical informatics. We will show the capability of SVM in QSPR analysis and its potential utilities to solve problems in chemistry and medical informatics through several applications in classification and correlation analysis.A brief description of the QSPR principle, realization process and research status was given in Chapter 1 of this dissertation. In this section, we also indicated the shortcoming of the present QSPR method and then it is necessary to introduce new methods. Then, the support vector machine was described in detail. At last, we gave a review of the application of SVM in chemoinformatics and medical informatics.In Chapter 2, the SVM was used to diagnose kidney stones and compared the results to Linear Discriminant Analysis (LDA) . According to results of the two methods, they both show good prediction ability. So it indicates that SVM is an effective tool to the prediction of kidney stones. The formation of kidney stone is connecting with environment, life conditions, body disorder and urinary disease. We discuss the stones of calcium up to 94.4% of all the stones from solubility product and the characters of calcium ion.In Chapter 3, we applied SVM to chemistry. For the first time, the SVM was used to develop a QSPR model that relates the structures of 607 organic compounds to their Gibbs free energy of formation. The descriptors calculated by CODESSA can be represented the molecular structures. Thirteen of these descriptors were selected by forward stepwise regression and were used developing models to predict the formation of Gibbs free energy of organic compounds. MLR and SVM were utilized to construct linear and non-linear models of the organic compounds. The optimal QSPR model based on support vector machine was obtained. The mean-absolute error (MAE) of formation Gibbs free energy was 31.0989 KJ/mol for the whole set, 30.5695 KJ/mol for the training set, and 35.9246 KJ/mol for the test set respectively. The prediction results are more satisfactory than those of MLR.