Research On Statistical Pattern Recognition Model For Chmp-markers Using Exausitive Method

Treating Chinese herbal medicine property (CHMP) as the access point, the ’973program’-’Research of basic theories of CHMP’ set up a hypothesis called ’tri-element of property-effect-material’:Chinese herbal medicine (CHM) is an entity of existence, with property being subjective and efficacy being objective reflection of CHM essential nature, besides CHMP functions through its efficacy. Efficacy reflects the core of CHMP, while material composition within CHM is the fundamentality to present its property. However, any single composition or efficacy cannot express the CHMP integrally. The interrelationship among property, composition and efficacy is the key in the research of CHMP theory.Therefore, based on CHMP theory, firstly we combined the chromatographic fingerprint of CHM (High-performance liquid chromatography, HPLC based) and statistical pattern recognition to build the CHMP recognized model and identify CHMP-markers. Then, according to "treating the hot syndrome with cold property medicine and treating cold syndrome with hot property medicine", we designed animal experiments to verify the CHMP-markers.Following the above research idea, we selected61classical CHMs in common use (their properties were determined in the light of ’Shen Nong’s Herbal Classic’,30’cold’ and31’hot’ respectively). Here we utilized HPLC to obtain the data set and exhaustive method was used to build model based on Waikato Environment for Knowledge Analysis(WEKA),which integrated many machine learning methods.80methods from7types of classifier (Bayesian classifier、tree、rule、function lazy classifier、meta algorithms Miscellaneous Classifiers) were built.First, using several evaluation metrics (accuracy rate, true positive rate, false-positive rate, precision rate, recall rate, F-measure, ROC area and the composite score) to select candidate statistical pattern recognition models. Second, the good ability of CHMP-markers identification was used to select appropriate statistical pattern recognition models. Abstracting ’cold’ material composition and ’hot’ material composition with the guidance of appropriate statistical pattern recognition models, and then animal experiments can be designed to verify the CHMP-markers.Analysis results:(1) Based on HPLC data, comparing the effect of80models from7types of classifier using exhaustive method25models can be selected as candidate statistical pattern recognition models. This result showed that more than a few models can be used to classify CHMP correctly.(2)5models (Bayesian Logistic Regression Ridge Logistic SMO、SPegasos、PLS) were selected as candidate appropriate statistical pattern recognition models considering the visualization of CHMP-markers. This result showed that models appropriated for recognizing CHMP-markers are not so many.(3) Statistical simulation showed that4models (Bayesian Logistic Regression Ridge Logistic、SMO、PLS) had higher sensitivity for CHMP-markers recognition than the other models. After comparing the CHMP-marker feature region predicted by the4methods with the real difference between hot and cold of61herbs. PLS model was determined as the most appropriate pattern recognition model for CHMP-markers from so many models(4) With the guidance of PLS model ’cold’ material composition and ’hot’ material composition were abstracted, the CHMP-markers were verified, by appropriate animal experiments. Major innovation:(1) Proposed exhaustive strategy to select appropriate pattern recognition model for CHMP-markers and constructed4key steps for selection. This is one innovation point in theory.(2) By way of comparing the effect of classifying,the ability of recognizing CHMP-markers,4models(Bayesian Logistic Regression、Ridge Logistic、SMO、 PLS) were considered as appropriate pattern recognition model for CHMP-markers. Providing alternative methods for recognizing CHMP-markers.(3) By comparing the CHMP-marker feature region predicted by the4methods with the real difference between hot and cold of61herbs and appropriate animal experiments, PLS model was find the most appropriate pattern recognition model for CHMP-markers. This furnish a standard method for recognizing CHMP-markers...