A Case Study of Techniques to Select UMVI in Health Care System

Most databases on healthcare have the problem of missing data. The most common method for handling missing data is imputation. The basic idea of imputation is to replace each missing value with a sensible guess, and then to carry on the analysis as if the missing data has never occurred. However, objective predictions with imputed datasets can only be generated if the missing mechanism is fully independent of the present or missing data. This guarantee is often broken in the construction of healthcare datasets due to bias in respondent responses or unintentional errors. The algorithm, therefore, offers an integrated approach to feature selection and missing value imputation to enhance the classification accuracy. The dataset D, of size M, is thus divided into training (D_tr) and test (D_te) sets. The operations are first conducted on the complete data subset D_complete, resulting in a subset reduced to D_complete’ based on the feature selection algorithm. Similarly, incomplete samples of D_tr are reduced to D_incomplete’. Missing values of D incomplete’ are imputed through a learning model trained on D_complete’, after which the same operations are conducted for the test set D_te. Experiments on five UCI datasets show that, even in the presence of reduced-class instances, the feature selection significantly enhances the classification accuracy as compared to the baseline imputation models at different missing rates. Moreover, DT tends to be oversensitive in selecting the features, while GA and IG behave similarly, suggesting that one needs ad-hoc feature selection methodologies to attain improved imputation and classification performances.