Risk prediction models for delayed bleeding after colorectal endoscopic submucosal dissection: a systematic review

To systematically evaluate risk prediction models for delayed bleeding after colorectal endoscopic submucosal dissection (ESD) and identify priorities for external validation and further refinement. PubMed, Embase, Web of Science, CINAHL, the Cochrane Library, CNKI, Wanfang, VIP, and SinoMed were searched from inception to July 1, 2026. Study and model data were extracted using the CHARMS checklist. Development quality, risk of bias, and applicability were assessed with PROBAST+AI, and reporting completeness was assessed with TRIPOD+AI. Findings were synthesized descriptively by model, prediction time point, and validation method. Ten studies reporting 10 newly developed models and one updated model were included. Four models used only preprocedural predictors, whereas seven required intraprocedural information. Three models underwent geographical external validation, with reported areas under the receiver operating characteristic curve (AUCs) ranging from 0.550 to 0.780. Concerns about development quality were high in nine of 11 development or updating assessments, and 24 of 31 performance evaluations were at high risk of bias. Full model specifications were reported for 8 of 11 models. Reporting was incomplete for model-building procedures, missing data handling, and analytical code availability. Existing models showed variable discrimination across validation settings. High risk of bias, limited independent external validation, and incomplete calibration assessment restrict confidence in their clinical use. Further studies should validate existing models in independent multicenter cohorts, update them when necessary, and assess their clinical usefulness.