Prediction of Molecular Colorectal Cancer Recurrence Using Machine learning

Abstract

Understanding the attributes that affect the occurrence of colorectal cancer can be very effective to developing methods that help in preventing this cancer disease. In many cases, a patient who receives cancer treatment must be kept under observation for a long period of time as cancer would most probably recur. The proposed approach is a feature-driven classification that predicts related features that greatly influence colorectal cancer recurrence and then uses these features to classify cases using different machine learning approaches. The microarray gene expression is combined with other demographic and clinical data to determine the relation to the recurrence measured using the statistical MRMR method. Then the best features among them that are highly correlated are selected. Different machine learning approaches were used to predict the recurrence, including the Quadratic SVM and the Gaussian Naïve Based approaches with and without the resulting correlated features. Performance improved dramatically when the related features were utilized. Using MRMR, we found that the accuracy of applying Gaussian Naïve Based is calculated as 80.6%, which outperformed the accuracy for Quadratic non-linear SVM by 77%. More data can be used in the future to improve the performance.