Artificial Intelligence for Intraoperative Bleeding Detection: A Narrative Review

Timely recognition of intraoperative bleeding can support visualization and hemostatic response. This narrative review organizes representative systems by temporal purpose - anticipation, detection, and post-hoc assessment - while treating input sources separately. Video classifiers, object detectors, segmentation-assisted models, and synchronized device - video pipelines demonstrate distinct capabilities, but their reported metrics are not interchangeable. One device-associated detector achieved 81.0% sensitivity with 4.6% precision, corresponding to roughly eight false alerts per operation, whereas other studies evaluated event counts, duration, or warning lead time. We distinguish offline testing, streaming feasibility, and prospective intraoperative evaluation, and relate these evidence levels to event-level accuracy, alert burden, and response time. Current evidence remains limited by small datasets, uncertain transfer across procedures and institutions, workflow integration, and surgeon interaction. Future studies should test transferable contextual cues and determine prospectively whether alerts improve recognition or hemostatic response without excessive interruption.