Clinical validation of risk-scoring systems to predict risk of delayed bleeding after endoscopic mucosal resection of large colorectal lesions.

A multicenter cohort study was performed in patients with nonpedunculated lesions ≥20 mm removed by EMR. We assessed the discrimination and calibration of the GSEED-RE and ACER models. Difficulty performing EMR was subjectively categorized as low, medium or high. We created a new model, including factors associated with DB in 3 cohort studies RESULTS: DB occurred in 45 of 1034 (4.5%) EMRs; it was associated with proximal location (odds ratio [OR], 2.84; 95% CI, 1.31-6.16), antiplatelet agents (OR, 2.51; 95% CI, 0.99-6.34) or anticoagulants (OR, 4.54; 95% CI, 2.14-9.63), difficulty of EMR (OR, 3.23; 95% CI, 1.41-7.40), and comorbidity (OR, 2.11; 95% CI, 0.99-4.47). The GSEED-RE and ACER models did not accurately predict DB. Re-estimation and recalibration yielded acceptable results (GSEED-RE area under the curve [AUC], 0.64; 95% CI, 0.54-0.74 and ACER AUC 0.65; 95% CI, 0.57-0.73). We used lesion size, proximal location, comorbidity, and antiplatelet or anticoagulant therapy to generate a new model (GSEED-RE2), which achieved higher AUC values (0.69-0.73; 95% CI, 0.59-0.80) and exhibited lower susceptibility to changes among datasets CONCLUSIONS: The updated GSEED-RE and ACER models achieved acceptable prediction levels of DB. The GSEED-RE2 model may achieve better prediction results and could be used to guide the management of patients after validation by other external groups. Clinicaltrials.gov no: NCT03050333.