Abstract

Background: Several models have recently been developed to predict mismatch repair (MMR) gene mutations. Their comparative performance with clinical criteria or universal molecular screening in a population based colorectal cancer (CRC) cohort has not been assessed.

Methods: All 1222 CRC from the EPICOLON cohort underwent tumour MMR testing with immunohistochemistry and microsatellite instability, and those with MMR deficiency (n = 91) underwent MLH1/MSH2 germline testing. Sensitivity, specificity and positive predictive value (PPV) of the PREMM_1,2 and the Barnetson models for identification of MLH1/MSH2 mutation carriers were evaluated and compared with the revised Bethesda guidelines (RBG), Amsterdam II criteria, and tumour analysis for MMR deficiency. Overall discriminative ability was quantified by the area under the ROC curve (AUC), and calibration was assessed by comparing the average predictions versus the observed prevalence.

Results: Both models had similar AUC (0.93 and 0.92, respectively). Sensitivity of the RBG and a PREMM_1,2 score ⩾5% was 100% (95% CI 71% to 100%); a Barnetson score >0.5% missed one mutation carrier (sensitivity 87%, 95% CI 51% to 99%). PPVs of all three strategies were 2–3%. Presence of MMR deficiency increased specificity and PPV of predictive scores (97% and 21% for PREMM_1,2 score ⩾5%, and 98% and 21% for Barnetson ⩾0.5%, respectively).

Conclusions: The PREMM_1,2 and the Barnetson models offer a quantitative systematic approach to select CRC patients for identification of MLH1/MSH2 mutation carriers with a similar performance to the RBG.