Validating predictive models for colorectal cancer

Abstract

InterviewsQ What is the current status of research regarding the development of models for predicting colorectal cancer?In the primary care setting there are very few well-designed and accurate p rediction models for identifying individuals with undiagnosed cancer that have also undergone evaluation by independent investigators. It is worth stressing how important it is that any prediction model should undergo evaluation by independent investigators. Prediction models always appear to work well on the data used to develop the model, and authors of such models will naturally tend to downplay such issues and be rather optimistic about their model. It is thus crucial to try and quantify such o ptimism and provide a more honest estimate of the performance and critical e valuation of the prediction model. Q Your recent UK-based study evaluated the performance of QCancer® (Colorectal) prediction model [1]. What were the main findings?QCancer (Colorectal) is a new prediction model for identifying individuals with undiagnosed colorectal cancer. The results from our study provided an independent evaluation of the performance of QCancer (Colorectal) on a large cohort of 2.1 million general practice patients. In particular, our results confirmed that QCancer (Colorectal) offered an improvement over using individual signs and symptoms in predicting whether a patient has u ndiagnosed colorectal cancer.Q The analysis on this study included 2.1 million patients. What are the challenges in working with such large amounts of data?Data sets of this magnitude are relatively uncommon and without a doubt the main challenge is that of the computational burden this places on the actual computing and calculating of the performance measures that describe the predictive a bility of the prediction model. ASK THE EXPERTSGary Collins*: Gary Collins is a senior medical statistician at the Centre for Statistics in Medicine, University of Oxford, UK. His main research interests are in the validation of clinical prediction models, and the methodological and design issues in conducting validation studies funded by the UK Medical Research Council. He is also interested in reporting of these studies and is leading a group of prediction modeling experts, clinicians and journal editors to produce a consensus-based guideline to improve the quality of studies reporting the development and validation of prediction models.