Research On Decision Model For Incomplete Mixed Data Of Electronic Health Records

The electronic health records are electronic historical records which are formed when people are engaged in healthcare-related activities and saved for future inquiries.After more than ten years of development,China has accumulated a large amount of data in this field.Using machine learning algorithm to automatically discover latent medical rules from abundant electronic health record data plays an important role in disease prevention,disease control and disease treatment.However,the features of electronic health records,namely incompleteness and fuzziness,restrict the application of traditional machine learning methods.Therefore,it is very necessary to establish a set of machine learning methods specifically for these features,and apply them to the decision model of the electronic health records.Firstly,a new method of measuring the distance between two fuzzy variables based on the area of the membership function is proposed.This method is consistent with the distance based on max-min close degree,and has more simple and convenient calculation.Based on this measurement method,the application range of k-NN imputation algorithm is extended from crisp data to fuzzy data and even to crisp-fuzzy mixed data.Specifically,the new distance measurement method of two fuzzy variables is used for searching similar samples of fuzzy samples and crisp-fuzzy mixed samples which containing missing values.Then the concept of fuzzy variable reduction value is introduced into the fuzzy data decision model,and an extreme learning machine based decision model for fuzzy data is established.Fuzzy variables are replaced by their several reduction values,which are trained by classical extreme learning machin algorithm.The actual fuzzy outputs are calculated by using inverse matix of reduction value equations.Another decision model for crisp-fuzzy mixed data is also eatablished by transforming fuzzy data to crisp data with the use of fuzzy variable reduction values.Finally,the proposed imputation algorithm and decision model are applied to electronic health records of a practical maternal and child health care management system,and an integral set of decision model for incomplete mixed data is given.The major novel work of this thesis includes the following four aspects:(i)A new distance measurement between two fuzzy variables is proposed,and its features are discussed;(ii)The k-NN imputation algorithm is extended to fuzzy data environment and crisp-fuzzy mixed data environment to imputate incomplete fuzzy data and incomplete crisp-fuzzy mixed data;(iii)Fuzzy variable reduction values are introduced into decision model for fuzzy data and crisp-fuzzy mixed data,and two decision models are correspondingly established;(iv)Featues of the electronic health records are analysized,which named incompleteness and fuzziness,and an integral set of decision model for these features is established.