Description
The information from prostate biopsies is often summarized in heuristic manners which lose information on the spatial distribution of cancer within the prostate. This heuristically determined set of summarizing variables are often used in clinical settings and in prostate cancer predictive models. A data-driven method for summarizing biopsy results may lead to better information and more accurate predictive models. Secondly, the sampling strategy for prostate biopsies generally stays the same as prostate volume fluctuates, so biopsy results may become less representative as prostate volume increases. A popular existing predictive model was tested to assess whether its accuracy suffers when prostate volume increases. 723 prostate biopsy results from Genesis Healthcare Partners were examined and spatial information for Gleason scores, cancer lengths and benign lengths were recorded. Principal component analysis (PCA) was performed with these variables to reveal whether the data could be summarized in a way that might preserve spatial information. In a subset of 204 patients, these components, as well as prostate specific antigen levels, digital rectal exam results and age were used to create several predictive models. Models which used PCA-based or heuristic-based variables were tested in their predictive accuracy of extra-capsular extension. The accuracy of the many models ranged between 69 - 75% with no clear trends as to whether PCA-based models outperformed heuristic based models. Five of the models however did outperform the popular online model published by Memorial Sloan-Kettering Cancer Center on the same dataset. Across the best models, accuracy was consistently worse in high prostate volume cases as expected, but the PCA-based models again did not perform any better.
