Variable Interaction Structure Based Machine Learning Technique for Cancer Tumor Classification

Abstract

Clinical diagnosis of disease samples largely benefits from multivariate statistical analysis, especially the machine learning techniques. One such example of medical application of statistical tools is the cancer tumor classification. In this study, Discriminating Partial Correlation Coefficient Metric (DPCCM) and Variable Predictive Model based Class Discrimination (VPMCD) are proposed as the new machine learning techniques for the same. Publicly available dataset Wisconsin Diagnose Breast Cancer (WDBC) dataset and Wisconsin Breast Cancer (WBC) dataset are analyzed to identify malignant cancer cells from benign cancer cells. Performance of the new classifiers are compared with the existing classifiers. The results based on the overall analysis indicate that DPCCM is potentially a better tool to identify and separate specific tumor samples. It achieves 100% accuracy of classification of test samples for WDBC dataset and 94% on WBC dataset. The performance of the approach is encouraging and its applicability to different cancer datasets (especially multi class problems) is being attempted further. © 2009 International Federation of Medical and Biological Engineering.