Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.631134
Title: Machine learning based data pre-processing for the purpose of medical data mining and decision support
Author: Rahman, M. Mostafizur
Awarding Body: University of Hull
Current Institution: University of Hull
Date of Award: 2014
Availability of Full Text:
Access through EThOS:
Access through Institution:
Abstract:
Building an accurate and reliable model for prediction for different application domains, is one of the most significant challenges in knowledge discovery and data mining. Sometimes, improved data quality is itself the goal of the analysis, usually to improve processes in a production database and the designing of decision support. As medicine moves forward there is a need for sophisticated decision support systems that make use of data mining to support more orthodox knowledge engineering and Health Informatics practice. However, the real-life medical data rarely complies with the requirements of various data mining tools. It is often inconsistent, noisy, containing redundant attributes, in an unsuitable format, containing missing values and imbalanced with regards to the outcome class label. Many real-life data sets are incomplete, with missing values. In medical data mining the problem with missing values has become a challenging issue. In many clinical trials, the medical report pro-forma allow some attributes to be left blank, because they are inappropriate for some class of illness or the person providing the information feels that it is not appropriate to record the values for some attributes. The research reported in this thesis has explored the use of machine learning techniques as missing value imputation methods. The thesis also proposed a new way of imputing missing value by supervised learning. A classifier was used to learn the data patterns from a complete data sub-set and the model was later used to predict the missing values for the full dataset. The proposed machine learning based missing value imputation was applied on the thesis data and the results are compared with traditional Mean/Mode imputation. Experimental results show that all the machine learning methods which we explored outperformed the statistical method (Mean/Mode). The class imbalance problem has been found to hinder the performance of learning systems. In fact, most of the medical datasets are found to be highly imbalance in their class label. The solution to this problem is to reduce the gap between the minority class samples and the majority class samples. Over-sampling can be applied to increase the number of minority class sample to balance the data. The alternative to over-sampling is under-sampling where the size of majority class sample is reduced. The thesis proposed one cluster based under-sampling technique to reduce the gap between the majority and minority samples. Different under-sampling and over-sampling techniques were explored as ways to balance the data. The experimental results show that for the thesis data the new proposed modified cluster based under-sampling technique performed better than other class balancing techniques. In further research it is found that the class imbalance problem not only affects the classification performance but also has an adverse effect on feature selection. The thesis proposed a new framework for feature selection for class imbalanced datasets. The research found that, using the proposed framework the classifier needs less attributes to show high accuracy, and more attributes are needed if the data is highly imbalanced. The research described in the thesis contains the flowing four novel main contributions. a) Improved data mining methodology for mining medical data b) Machine learning based missing value imputation method c) Cluster Based semi-supervised class balancing method d) Feature selection framework for class imbalance datasets The performance analysis and comparative study show that the use of proposed method of missing value imputation, class balancing and feature selection framework can provide an effective approach to data preparation for building medical decision support.
Supervisor: Davis, Darryl N. Sponsor: University of Hull
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.631134  DOI: Not available
Keywords: Computer science
Share: