COLOFIT: Development and internal-external validation of models using age, sex, faecal immunochemical and blood tests to optimise diagnosis of colorectal cancer in symptomatic patients

Abstract

Objective To develop and validate a model using available information at the time of Faecal Immunochemical testing (FIT) in primary care to improve selection of symptomatic patients for colorectal cancer (CRC) investigations. Design Population based cohort study. Setting All adults ≥ 18 years of age referred to Nottingham University Hospitals NHS Trust between 2018 and 2022 with symptoms of suspected CRC who had a FIT. Participants The derivation cohort (Nov/2017-Nov/2021) included 34,435 patients with FIT results who had 533 (1.5%) CRCs at 1-year. The validation analysis included 34,231 patients with first FITs in the derivation cohort with 516 (1.5%) cancers, and 16,735 patients with first FITs in the validation cohort with 206 (1.2%) cancers. Main outcome measures Predicted 1-year CRC diagnosis using Cox proportional hazards modelling with selected multiple fractional polynomial transformations for age, faecal haemoglobin concentration (f-Hb) value, mean corpuscular volume (MCV), platelet count and sex. In the internal-external validation we calculated discrimination and calibration to assess performance and estimated net benefit values across a range of CRC risk thresholds to assess clinical utility. Results In the survival model multiple fractional polynomial transformations were selected for age, f-Hb and platelet count, with MCV included as a linear variable and sex as a binary variable. Haemoglobin was not selected. At a CRC risk threshold of 0.6% (equivalent to f-Hb=10 µgHb/g (µg/g)) overall performance of the validated model across age strata using Harrell ′s C index was ≥ 0.91% (overall C statistic 93%, 95% CI 92%-95%) with acceptable calibration. Using this model would yield similar numbers of detected and missed cancers but require 20% fewer investigations than a f-Hb ≥10 µg/g strategy. For approximately 100,000 people per year with symptoms of suspected CRC, we predict it might save >10,000 colonoscopies with no evidence that more cancers would be missed if we used our model to triage investigations compared to using FIT at the currently recommend level for referral. Conclusions Including age, sex, MCV, platelets and f-Hb in a survival analysis model to predict the risk of CRC yields greater diagnostic utility than a simple binary cut off f-Hb≥10 µg/g. Enacting model-based triage of a symptomatic CRC pathway may decrease the burden on endoscopy whilst maintaining diagnostic accuracy. Further targeted validation of this approach is required in external populations with symptoms of possible CRC.