Genetics, primary care records and lifestyle factors for short-term dynamic risk prediction of colorectal cancer: prospective study of asymptomatic and symptomatic UK Biobank participants

Abstract

Objectives: To quantify the contributions of polygenic scores, primary care records (presenting symptoms, medical history and common blood tests) and lifestyle factors, for short-term risk prediction of colorectal cancer (CRC) in general and symptomatic individuals.

Methods and analysis: This prospective cohort study used data from the UK Biobank with follow-up until 2018. It included 160 507 participants with linked primary care records and a subcohort of 42 782 participants with recent CRC-related symptoms. The outcome was the first-recorded CRC diagnosis within 2 years. Dynamic risk models with time-varying predictors were derived using a super-landmark framework. Model discrimination was assessed through Harrel's C-index, and predictor contributions to model discrimination were evaluated using inclusion-order-agnostic Shapley values.

Results: C-indices (95% CIs) were 0.73 (0.72 to 0.73) and 0.69 (0.68 to 0.70) for the general and symptomatic participants, respectively. Shapley contributions to model discrimination (95% CIs) were core predictors (eg, age, sex) 33% (25% to 42%) (symptomatic: 34% (9% to 75%)), polygenic scores 16% (8% to 26%) (8% (-21% to 35%)), primary care blood tests 32% (19% to 43%) (41% (16% to 73%)), medical history 11% (4% to 17%) (9% (-25% to 37%)), lifestyle factors 6% (0% to 11%) (-5% (-32% to 13.4%)) and symptoms 3% (-2% to 7%) (13% (-19% to 41%)).

Conclusions: Polygenic scores contribute substantially to short-term risk prediction for CRC in both general and symptomatic populations; however, the contribution of information in primary care records (including presenting symptoms, medical history and common blood tests) is greater. Lifestyle factors not routinely collected in primary care contribute minimally.

Keywords: Biostatistics; Bowel cancer; Cancer screening; Colorectal cancer; Genetic markers.

Figure 1. (A) HRs from bidirectional stepwise Cox regression for the study cohort; (B): discriminative contribution of predictors using Shapley values (C-index>0.5); (C) C-indices from 200 bootstrap samples for each combination of predictor sets. Colour-coding indicates the predictor set in all figures. Overall mode, C-index: 0.728 (0.726– to 0.731). BMI, body mass index; NSAIDs, non-steroidal anti-inflammatory drugs.

Figure 2. (A) HRs from bidirectional stepwise Cox regression for the ‘symptomatic’ cohort, characterised by haemorrhoids, constipation, rectal bleeding and diverticular disease, as shown in online supplemental figure 3; (B) discriminative contribution of predictors using Shapley values (C-index>0.5); (C) C-indices from 200 bootstrap samples for each combination of predictor sets. Overall model C-index: 0.689 (0.682– to 0.695). BMI, body mass index; NSAIDs, non-steroidal anti-inflammatory drugs; UK education system qualifications (or equivalent): Advanced Level (A-level), Advanced Subsidary Level (AS-level), Ordinary Level (O-level), General Certificate of Secondary Education (GCSE), Certificate of Secondary Education (CSE), National Vocational Qualification (NVQ), Higher National Diploma (HND), Higher National Certificate (HNC).