Data Mining In Clinical Medicine: Research And Application

With the development of hospital informatization and the application of hospitalinformation systems and electronic medical records, the data capacity of the hospitaldatabase has been expanding continually. These precious medical data are very valuablefor disease diagnosis, treatment and medical research. Currently, however, mostdatabase operation in hospitals was limited to some ordinary process, such as data entry,modify, query, and delete, which were low-end operation of the medical database, lackof data integration and analysis, much less for automatic acquisition of knowledge andmedical decision support. On the other hand, traditional data analysis tools can only dosome light processing in the face of mass data, rather than obtaining the inherentrelevance and hidden information among the data, which resulted in the so calledpredicament of “data rich but information poor”. To get out of this predicament, atechnology known as data mining later was developed as the time require, which canintelligently and automatically convert the data into useful information and knowledge.Data mining is the nontrivial process of identifying valid, novel, potentially useful, andultimately understandable patterns in data.First, based on the study of data mining theory and investigation of clinicalapplication needs, this paper proposed some improvements for two data miningtechniques: na ve Bayesian method and BP ANN (artificial neural network), whichwould be used in later chapter. The improvements for traditional Naive Bayesianmethod include attribute weights based on the Wald value of logistic regression andclass conditional probabilities calculation based on kernel density estimation. And thosefor BP ANN include adding momentum term and adjusting learning rate adaptively.Then, ICU data set and transfusion data set were used to test and validate theirapplications in clinical medicine respectively. With the ICU data, a naive Bayesianmodel was constructed to predict the PHM (probability of hospital mortality) of ICUpatients. Compared with the logistic regression model, it suggested that the naive Bayes model showed an obvious advantage in resolution. As to transfusion data, we used theimproved BP ANN to predict values of vital signs and Hb indicator of the patients afterred blood cell transfusion, and established a multi-indicators scoring model to evaluateand pre-judge the effect of blood transfusion. By the validation on the test sample set, itsuggested that the multi-indicators scoring model, which was based on Numericalprediction with BP ANN, showed a better ability to distinguish bad transfusion withprediction accuracy of89.7%, which turned out that it can be used for accuratepre-judgment of the effect of the transfusion.