Hybrid prediction model with missing value imputation for medical data

• Proposed novel hybrid prediction model with missing value imputation.
• HPM-MI has improved accuracy, sensitivity, specificity, kappa and ROC on 3 datasets.
• The best accuracy is achieved for diabetes, hepatitis, and breast cancer datasets.
• MVI is one of the important step of proposed model.

Accurate prediction in the presence of large number of missing values in the data set has always been a challenging problem. Most of hybrid models to address this challenge have either deleted the missing instances from the data set (popularly known as case deletion) or have used some default way to fill the missing values. This paper, presents a novel hybrid prediction model with missing value imputation (HPM-MI) that analyze various imputation techniques using simple K-means clustering and apply the best one to a data set. The proposed hybrid model is the first one to use combination of K-means clustering with Multilayer Perceptron. K-means clustering is also used to validate class labels of given data (incorrectly classified instances are deleted i.e. pattern extracted from original data) before applying classifier. The proposed system has significantly improved data quality by use of best imputation technique after quantitative analysis of eleven imputation approaches. The efficiency of proposed model as predictive classification system is investigated on three benchmark medical data sets namely Pima Indians Diabetes, Wisconsin Breast Cancer, and Hepatitis from the UCI Repository of Machine Learning. In addition to accuracy, sensitivity, specificity; kappa statistics and the area under ROC are also computed. The experimental results show HPM-MI has produced accuracy, sensitivity, specificity, kappa and ROC as 99.82%, 100%, 99.74%, 0.996 and 1.0 respectively for Pima Indian Diabetes data set, 99.39%, 99.31%, 99.54%, 0.986, and 1.0 respectively for breast cancer data set and 99.08%, 100%, 96.55%, 0.978 and 0.99 respectively for Hepatitis data set. Results are best in comparison with existing methods. Further, the performance of our model is measured and analyzed as function of missing rate and train-test ratio using 2D synthetic data set and Wisconsin Diagnostics Breast Cancer Data Sets. Results are promising and therefore the proposed model will be very useful in prediction for medical domain especially when numbers of missing value are large in the data set.