Artificial Intelligence Approaches for Gastrointestinal Bleeding Detection in Capsule Endoscopy: A Systematic Review and Critical Analysis

Background/Objectives: Gastrointestinal (GI) bleeding is a common, potentially life-threatening condition, and capsule endoscopy (CE) is the principal investigation for suspected small-bowel bleeding after non-diagnostic upper endoscopy and colonoscopy. Because manual review of thousands of CE frames is labor-intensive, artificial intelligence (AI) has been proposed to automate bleeding detection. Methods: Following PRISMA 2020, this systematic review synthesizes 15 peer-reviewed journal studies (2020–2025) applying traditional machine learning, convolutional neural networks, transfer learning, attention-based or hybrid architectures, segmentation, or object detection to GI bleeding detection in CE. Methodological quality was appraised with an AI-adapted risk-of-bias framework informed by QUADAS-2 and PROBAST. Results: Frame-level classification accuracy ranged from 90.6% to 99.9% (median 98.5%; 13 studies), with the lowest values from evaluations closest to clinical practice; the segmentation and detection studies reported a mean intersection-over-union of 86.9% and an mAP@0.5 of 79.0%. Only two studies (13%) were at low risk of bias for validation strategy, none reported patient-level separation of training and test data, only four (27%) used clinical patient data, and none compared AI with conventional endoscopy or clinical outcomes. Conclusions: Current evidence demonstrates high performance on curated datasets but not yet generalizability or clinical readiness; patient-level, external, and prospective multicentre validation is required.