Bowel Cancer Risk Prediction Using NHS 111 Call and Primary Care Prescription Data
I built and tested a way to use NHS 111 calls and GP prescribing records to flag people at higher risk of bowel cancer and to rank how urgently they should be seen, and found that adding prescription data roughly doubles the cancers found for the same number of patients reviewed.
This project was completed by Piyali Dutta, Lead Data Engineer, in the Data Engineering Team, as part of the Data Science MRes at the University of Leeds.
This project builds directly on an earlier NHS England study that used 111 calls, hospital data and demographics to build risk cohorts for nine cancers. That study deliberately left out bowel cancer because it already has a national screening programme. That left an obvious gap. Bowel cancer is still the UK’s fourth most common cancer and second biggest cancer killer, survival falls from about 90% at stage 1 to about 10% at stage 4, and the existing tools (NICE NG12 red-flag referral, FIT screening, QCancer and ColonFlag) mostly act once a patient has already presented. GP prescribing records were an untapped, population-wide signal. Iron and laxative prescriptions, for example, plausibly reflect anaemia and altered bowel habit before diagnosis. So I chose to ask whether prescription data adds value beyond demographics and 111 calls, and to extend the earlier study’s framework to bowel cancer, a screened cancer, with a new urgency-scheduling stage.
I linked four NHS sources (the cancer registry, patient demographics, NHS 111 calls and GP prescribing) for about 35.5 million adults aged 40 and over, and built the work on the Federated Data Platform using Palantir Foundry. I trained the same XGBoost model twice, once on demographics and 111 calls (the baseline) and once with prescription features added (the full model). I tested both on a held-out population of about 7 million patients, where bowel cancer prevalence is 0.217%. I evaluated them in three ways. Method A ranks every patient by risk and flags the top-N. Method B distils the model’s most important features into a simple two-condition rule that needs no model to run. Method C takes a flagged cohort and predicts time to diagnosis, then splits patients into four urgency bands.
The headline result is that standard accuracy barely moves (AUC 0.768 to 0.772), but the operational picture is very different. Reviewing the top 10,000 highest-risk patients, the baseline finds 100 cancers and the full model finds 210, a 9.7× lift over background, or 1 in 48 flagged patients with cancer versus 1 in 100. The best simple rule (a new iron prescription in the last six months plus any laxative prescription) finds 83 cancers at 3.67× lift, so there is a real trade-off between yield and ease of deployment. For urgency, the full survival model gives a clean gradient, with 3.35% of the critical band having cancer against 0.6% in the lowest band, while the baseline gives almost no timing signal. Prescriptions also change who is flagged: women rise from 8.3% to 41.0% of the top 50,000, and people aged 70 to 79 from 2.8% to 12.1%.
The limitations are real. This is retrospective work on a single health system. Some subgroup results rest on small numbers. Prescriptions may partly reflect a work-up already under way, which is the lead-time problem. Some subgroup gains, such as those for unrecorded ethnicity, may be confounded by age. The next steps I propose are a prospective pilot as an adjunct to NICE NG12 and QCancer, linkage to consultation and screening data, and external validation.
| Output | Link |
|---|---|
| Project Link | Link |
| Project Link | Link |
| Project Link | Link |