MODELLING EARLY DETECTION OF PROSTATE CANCER

Abstract

Prostate cancer is one of the most common cancers among men in the world (excluding non-melanoma skin cancers). According to the statistics from the Canadian Cancer Society, it is the third leading cause of death from cancer in men in Canada. In general, prostate cancer is treatable with 5-year survival rate of 99% for early-stage. However, if the cancer has spread to nearby organs, the 5-year survival rate drops to 28%. Unfortunately, 92% of patients are diagnosed at an advanced stage. Early detection is an ongoing challenge for prostate cancer treatment. Existing statistical models (e.g. principal component analysis) are more likely to inform us the statistical relationship between each metabolite. However, they have a poor performance of predicting the early prostate cancer. In order to improve predictions, in this thesis, we developed models using metabolite profile to identify patients who are likely to having prostate cancer. Several predictive methods such as Support Vector Machine (SVM), K-Nearest-Neighbour (KNN), Random Forest, LASSO, and PLS-DA were used.