Cancer Screening Using Multi-modal Differential Principal Orthogonal Decomposition

Signatures of a cancer may be encrypted in DNA microarrays, and once found, can be used for diagnoses. The standard Principal Orthogonal Decomposition (POD) method has been used to effectively detect liver and bladder cancers. Supporting work demonstrated feasibility of detecting leukemia and colon cancer via extending the standard POD to use principal features extracted from cancer and healthy sets as input to Support Vector Machine (SVM). In this study, we improved screening performances with inclusion of multiple dominant extracted modes from both cancer and healthy samples. We also investigate the efficacy of combining gene expressions with their derivative information to improve the accuracy of disease detection from previous work. We report sensitivity, specificity, and accuracy from classifications using extended POD with SVM trained with weighted projections onto multiple modes extracted from cancer and normal gene expressions and their derivatives. This is equivalent to mining not only the resembling features, but also the behavioral features. By using multiple modes, classification and prediction can be more distinctively definitive. We found that, in many cases, our new approach using multi-modal POD tends to improve cancer-screening accuracy.