Application Of Date Mining Technology In Identifying Drugger’s Human Pulses

In recent decades, the rapid development of database technology and mass storagehardware makes it easier for people to collect data. In response to the challengesproposed by the new data set, traditional data analysis technology possesses variouslimitations while data mining technique makes a breakthrough in the limitations. Withthe development of medical informationization and the explosion of diagnostic data, itis necessary to extract and analyze the potential significant knowledge attained by datamining technology.Addicted to drugs cause serious damages to human health, such as damaging thecentral nervous system, causing other organs （heart, liver, and lungs） dysfunction andpathological changes. According to the theory of Chinese Diagnostics, human pulsesignal contains rich physiology and pathology information such as the cardiovascularsystem information which owns an important clinical value in diagnosing the lesion ofthe human body. We studied on30pulse signal samples （15healthy peoples,15heroinaddicts） collected in Chongqing rehab, and established two classifiers （C4.5DecisionTree and LMBP Neural Network） to differentiate them via the data mining softwareSAS/EM.To efficiently dig out useful information from large miscellaneous data, datapreprocessing is in need. Wavelet transform has good localization properties inprocessing non-stationary signal. Except for adopting time-domain signal as inputfeatures to divide the pulses signal preliminary, we also extract detail coefficientsthrough wavelet transform as the second characteristic parameters. The pulse signal isdecomposed into three levels through db4wavelet Mallat multiresolution algorithm inmatlab wavelet toolbox, so as to clean data, reduce dimension and prepare for thefollowing classification.As one method of data mining based on machine learning, C4.5decision treepossesses simple structure, fast classification, strict theoretical basis, which is useful tosolve small sample learning problems. We proposed a C4.5decision tree classifier toidentify the drug addicts, which takes the gain rate of information to select attribute. Theexperimental results show that the classifier selects T34（the threshold value-0.1736） asthe root node when the time domain parameters are inputted to the C4.5decision treemodel. Its recognition rate is up to93.3%with B06and B01are mistakenly identified while the system robustness is poor. When each level of the detail coefficients are usedto training C4.5classifier, the system selected cd1₃7, cd2₅and cd3₉as the bestdecision points, in which the best accuracy is93.3%. Although its preparation rate failsto increase, it has better robustness and stability.With high fault tolerance and reliability, BP artificial neural network isextraordinary suitable for research of biomedical signals. It doesn’t need prior statisticalhypothesis of the data and noise, summarizing expert knowledge and experience into astrict provision, as well as, its self-organization and adaptive learning ability greatlyrelaxed the constraint conditions of the traditional identification methods. In order toimprove the classification performance, we also introduced the Levenberg-MarquardtBP algorithm, and established a single hidden layer （H1） and a double hidden layer BPnetwork （H2） classifier. Using the detail coefficients at the third level for training andtesting the H1and H2, the results show that both of their overall classification accuracyis96.67%. The experimental results show that their recognition rate are the same whilethe H1’average error rate is lower.Finally, we compared the four methods’ performance in terms of accuracy,computation speed, robustness and interpretability. We concluded that single hiddenlayer of BP network model considering detail coefficients as the feature vector has thebest and desired classification results.