A dynamic prediction model for early detection of colorectal cancer using routine blood test results from primary care


Session type:

Pradeep S. Virdee1, Jacqueline Birks2, Tim Holt2
1Centre for Statistics in Medicine, University of Oxford, 2University of Oxford



Colorectal cancer is the fourth most common type of cancer in the UK. Around 55% of patients are diagnosed at a late stage (Stage 3 and 4), where likelihood of survival is reduced: five-year survival is 93% at Stage 1 versus 10% at stage 4. This highlights the importance of early detection. The full blood count (FBC) is a common blood test in primary care. Some FBC indices, including haemoglobin, mean cell volume, and platelet count, change over time as the cancer develops. We built a dynamic prediction model that utilises changes in repeated FBC measurements to identify risk of colorectal cancer diagnosis two years in the future.


We performed a cohort study using data from the Clinical Practice Research Datalink and National Cancer Registration and Analysis Service. We developed a multivariate joint model of longitudinal and time-to-event data for males and females separately. Using historical repeated FBCs over five years prior to baseline (last included FBC), age-adjusted trajectories in haemoglobin, mean cell volume, and platelet measurements informed two-year risk of colorectal cancer diagnosis, using a Cox model. Model performance was assessed using Harrell’s c-statistic for discrimination.


Developing joint models is computationally challenging so we used a random sample of 226,130 males and 226,914 females, of whom 0.4% (n=870) and 0.3% (n=677) were diagnosed two years after their baseline FBC, respectively. Simultaneous age-adjusted decreases in haemoglobin and mean cell volume and increase in platelets from the population trajectory (patients with no diagnosis recorded) increased the risk of diagnosis for both males and females (each p<0.05). The c-statistic was 0.749 (95% CI: 0.729, 0.768) for males and 0.756 (95% CI: 0.738, 0.775) for females.


Our dynamic prediction model has performs adequately in identifying patients at risk of future diagnosis. Further model performance statistics will be presented, including plots of observed versus predicted probabilities for calibration.

Impact statement

Our prediction model has the potential to utilise small changes in FBC indices occurring simultaneously over time to identify patients who need further investigation for colorectal cancer. Such changes can appear before overt symptoms occur, so the prediction model could facilitate earlier detection.