Radiomics based risk stratification for recurrence free survival in oropharyngeal cancer patients

Oropharyngeal Cancer, along with other cancers, requires the use of certain techniques to help plan for treatment. The use of image capturing technology to gather information about a particular cancer is routinely done in clinical practice. Using this information, we seek to predict how likely a particular cancer could return after initial treatment. Being able to predict overall survival and the likelihood of a cancer returning can help compare and improve treatments. Machine learning can help ingest these data and build statistical and mathematical models to help with these predictions. Our approach uses these large-scale image data to identify groups of similar patients and define a risk stratification to improve the prediction of patient’s survival. We apply feature selection to identify a set of important and meaningful data from the whole set. Once we have a dataset that has the relevant information needed for predictions, we build an ensemble survival model, and assign the patients into groups based on this data. We then use the clinical only data to perform risk prediction using a combination of different models that build on each other. Next, we combine the assignments of the patients along with other clinical data and see a noticeable improvement for Overall Survival and Recurrence Free Survival.