Artificial Intelligence for Intraoperative Bleeding Detection: A Narrative Review

1. Introduction

Intraoperative bleeding is a time-critical surgical problem: recognition and control must occur while blood can obscure the operative field and interrupt ongoing dissection. Perioperative mortality studies underscore the clinical importance of hemorrhage and major bleeding complications [1,2]. For surgical assistance, the relevant question is not only how much blood has been lost, but whether an evolving event can be recognized early enough to support the surgical team’s response.

Existing reviews examine blood-loss estimation and hemorrhage prediction [3,4]. Eppler and colleagues synthesize automated detection across several intraoperative adverse events and identify heterogeneous definitions and reporting [5]. For bleeding assistance, a further distinction is essential: recognizing visible blood, predicting an event before onset, and delivering a useful alert are different achievements. Detection scores alone do not establish warning lead time or interruption burden. The anticipatory evidence included here comprises two small studies, underscoring the gap between interest in early warning and its validation [6,7]. This review connects temporal tasks and input sources to validation settings and alert-level endpoints, clarifying what existing results establish for intraoperative use.

We organize recognition by temporal purpose: anticipation before bleeding onset, detection of an ongoing event, and post-hoc assessment of its course or severity. Video-derived context and synchronized device signals are input characteristics within these tasks. Bounding boxes and segmentation are included when they support recognition or alerting; preoperative risk prediction, quantitative blood-loss estimation, detailed localization, and autonomous hemostasis are addressed in companion reviews. Conventional image-processing and severity-grading studies provide methodological and clinical comparators. One simulated suction-planning study is included only to illustrate externally supplied context at the recognition–response interface.

We combined a structured keyword search with citation searching. Three term sets—bleeding or hemorrhage prediction; robotic, open, or intraoperative setting; and deep learning, context-aware, generative, or agentic AI—yielded 24 query combinations, run on 28 April 2026 in Google Scholar and supplemented by PubMed/MEDLINE and Scopus. English-language reports published from 2020 onward were considered. Ninety-eight full-text reports were screened; ten met the core criteria for evaluated systems recognizing imminent or active bleeding from operative video or synchronized signals, including two reports from the same group, and three additional reports served as simulation or adjacent-task comparators. Most exclusions involved nonsurgical hemorrhage, preoperative risk models without operative video, blood-loss quantification, or non-bleeding intraoperative outcomes. The resulting evidence base therefore contains fewer than fifteen primary studies of AI recognition of intraoperative bleeding from operative video or synchronized signals. Preprints were considered when no archival version existed, and reports sharing development data were not treated as independent validation.

The review is intended for surgeons and clinical teams seeking to assess the usefulness and limitations of bleeding alerts, and for AI and surgical-engineering researchers seeking clear recognition targets and appropriate evaluation methods. By connecting task definitions, computational approaches, and deployment barriers, it also supports clinical–technical teams planning validation studies and integration into operative workflows.

2. Clinical and Operational Definitions of Bleeding

2.1. From Blood-Loss Thresholds to Visual Bleeding Events

Clinical grading based on accumulated blood loss addresses a different question from recognition of an event in surgical video [8]. A detector requires an operational definition of what is visible and when it becomes positive. Active flow, blood accumulation, and newly appearing blood after device activation are related but non-equivalent targets. For example, active-bleeding detection can explicitly distinguish flowing blood from pooling, whereas a blood-accumulation classifier may label the resulting collection of blood without establishing that flow continues [9,10]. Table 1 summarizes these distinctions and their annotation consequences.

Intervention-based definitions add clinical meaning but also introduce dependence on surgical practice. Electrosurgical bleeding labels can incorporate whether a hemostatic response is required, while severity grading can relate visible blood to the method used to control it [12,13]. These labels differ from active-flow annotations in both target and dependence on intervention decisions.

2.2. Procedural Context and Clinical Actionability

Bleeding recognition depends on the surrounding action as well as blood appearance. Abrupt instrument movement and electrosurgical activation provide distinct temporal cues, while the current dissection scene helps determine the significance of new blood [6,12]. These cues motivate context-sensitive recognition rather than a general account of surgical risk factors. A system must distinguish a change suggesting new bleeding from persistent staining or a previously recognized pool.

Clinical actionability also depends on whether bleeding compromises visualization or requires a change in operative behavior. Recognition of blood, recognition of active flow, and assessment of the need for intervention therefore represent distinct outputs. Their relationship depends on procedure-specific clinical labels.

2.3. How Often Recognition Fails, and How Long a Response Takes

Reliable estimates of missed or delayed intraoperative bleeding recognition remain scarce. Intraoperative adverse events are underreported when ascertainment relies on retrospective records rather than direct observation [5], although this evidence does not quantify real-time recognition failure by the operating team. Human assessment can also be imprecise: in video review of a cadaveric carotid-injury exercise, fellowship-trained experts underestimated blood loss by a mean of 131 mL and identified fewer than half of episodes exceeding 500 mL [14]. These findings establish a clinically relevant recognition gap without defining its frequency during live surgery.

The available response interval also depends on event severity. In the same simulation, successful management required hemostasis within five minutes [14], whereas Horita and colleagues annotated 254 bleeding events across 27 colectomies, or roughly nine events per operation [9]. Together, these studies illustrate substantial variation in bleeding context and frequency but do not establish a general onset-to-intervention threshold. The clinical value of the 0.662-s advance warning reported by Rahbar and colleagues therefore cannot yet be determined from available human-response data. Evaluations of early-warning systems should align bleeding onset, human recognition, alert delivery, and hemostatic intervention on a common timeline and report event severity alongside alert burden.

3. Bleeding Recognition: Tasks and Methods

Temporal purpose and input source provide separate ways to describe a recognition system. Table 2 places each representative study once, according to its primary task and actual inference inputs. Context extracted from video remains a video input; labels used during training do not constitute an additional runtime modality. Post-hoc assessment connects detected events to operative interpretation. No included recognition study used contextual information supplied outside the video or device stream at inference time; the column is retained to mark that absence.

Figure 1 connects these task distinctions to event matching and alert delivery; the evaluation settings are developed in Section 4.

We use three descriptive evidence levels in Table 3: L1, offline image or video evaluation; L2, an integrated streaming or interactive interface demonstrated on replayed or experimental inputs; and L3, prospective operation during surgery with evaluated outputs. The highest demonstrated setting is listed. These levels describe evaluation setting, not study quality or proven clinical benefit; prospective enrollment alone does not establish L3, and an exploratory L3 study is not a comparative outcome trial.

3.1. Anticipation Before Bleeding Onset

Anticipation predicts bleeding before its independently annotated onset. Rahbar and colleagues used local-entropy changes to identify abrupt tool motion, reporting a mean advance warning of 0.662 s in a small video series [6]. Checcucci and colleagues ran a neural-network alert application during 10 robotic prostatectomies and reviewed its outputs against recorded videos; they reported a next-second event after confidence exceeded 93–95% [7]. The latter provides exploratory intraoperative evidence (L3), although its timing framework was still under development. Together, these studies motivate short-horizon warning without resolving whether the signal reliably precedes visible onset across procedures.

3.2. Detection of Ongoing Bleeding

Video-based methods span classification, spatial detection, segmentation, and temporal event analysis. Horita and colleagues distinguished active flow from pooling with YOLOv7 (AP₅₀ 0.574; 48.5 FPS), while Sunakawa and colleagues reported a Dice coefficient of 0.77 for bleeding-region segmentation during hepatectomy [9,11]. Marullo and colleagues combined blood-accumulation classification with tool segmentation, and BLAIR extended video recognition to a replay interface evaluated against a urologist’s recognition times [10,16]. Earlier recognition relative to an observer remains distinct from prediction before independently annotated onset. Rus and colleagues additionally demonstrated an augmented-reality route for displaying detections [17].

Temporal models can summarize event course as well as presence. Hong and colleagues evaluated entire-video bleeding counts and duration, reporting three-fold mean absolute errors of 12.3–13.2 events and 5.09–5.83 min across two AMAGI variants [15]. These summaries do not identify which events were missed or when an alert would arrive.

Synchronized device signals can narrow the detection interval. Ribbens and colleagues combined video with generator activation times and reported fragment-level sensitivity of 81.0% with 4.6% precision [12]. Section 5 translates this precision into alert burden. Temporal association with device activation provides context for detection but does not itself establish causation.

3.3. Post-Hoc Assessment and Clinical Context

Post-hoc assessment relates recognized bleeding to its severity or management. Sawa and colleagues linked a red-pixel index to hemostatic management during robot-assisted pancreatoduodenectomy, analyzing 885 bleeding scenes and four conversion scenes [13]. This exploratory grading scheme is an operational comparator, rather than a learned detector. Hemostatic actions supply reference labels; they are not an additional model input. Such work helps define clinically meaningful targets, while transfer to other procedures remains an open question.

4. Evaluation and Reporting Requirements

4.1. Data, Annotation, and Validation

Dataset size requires counts of independent operations as well as images, clips, and events: adjacent frames are not independent clinical observations. Simulated resources such as SOCAL support method development, but validation of instrument detection or outcome prediction on these data does not establish clinical bleeding-alert performance [19]. Patient- or operation-level partitions reduce leakage; external testing examines transfer beyond the development setting. Confidence intervals need to account for clustering within operations.

Reproducible evaluation requires an explicit target, onset and offset rules, annotator expertise, disagreement resolution, and treatment of obscured intervals. Complete non-bleeding periods matter because selected bleeding clips can conceal false-alert burden. The analysis unit and test prevalence accompany each result, especially when comparing frames, activation fragments, and complete events.

4.2. From Image Scores to Event-Level Performance

Sensitivity and precision describe threshold-specific classification, AP₅₀ bounding-box detection, and Dice or intersection over union (IoU) spatial overlap. AUROC and area under the precision–recall curve (AUPRC) summarize threshold ranges but do not define an acceptable alert threshold. Count and duration errors provide operative summaries, while low precision in device-associated detection shows why sensitivity alone can misrepresent alert utility [12,15]. Appendix A distinguishes reported measures from the review’s proposed priorities.

Event-level evaluation requires explicit onset, matching, and deduplication rules. Figure 1A distinguishes pre-onset warning lead time from post-onset detection delay. Event sensitivity should be paired with false alerts per operative hour, because repeat notifications may count as one matched event yet still interrupt the team. Comparisons therefore require common event rules, analysis units, and test prevalence.

4.3. Timing and the Delivered Alert

Figure 1B separates model throughput (FPS) from end-to-end alert latency. System latency spans capture to display, including buffering, processing, any future-frame waiting, and alert-handling delays; surgeon interpretation is separate. Thresholds, persistence, and suppression determine which outputs reach the display and when. Historical context does not necessarily add steady-state waiting, whereas future-frame dependence does. Horita and colleagues identified display delay as unevaluated despite fast inference [9]. Hardware, resolution, temporal dependencies, and alert-handling rules therefore accompany timing results.

5. Clinical Translation and Barriers to Deployment

5.1. What the Current Evidence Supports

The evidence levels in Table 3 separate recognition tests from interface demonstrations and intraoperative use. BLAIR provides replay-based evaluation, while the earlier Checcucci alert study reports prospective operation during surgery [7,16]. These settings support different claims. Neither an L3 designation nor favorable surgeon ratings establish reduced blood loss, faster hemostasis, or fewer complications without an appropriate clinical comparison.

5.2. Data and Generalization Barriers

Changes in camera appearance, instruments, bleeding prevalence, and hemostatic practice can alter both model inputs and the meaning of a positive label. Horita and colleagues identified procedure-dependent backgrounds and limited representation of severe bleeding as constraints on transfer [9]. External validation therefore needs routine and difficult cases from independent settings, with explicit separation of visual domain shift from differences in what clinicians label actionable. Local threshold adjustment cannot establish reliability for bleeding patterns absent from development data.

5.3. Technical and Workflow Integration Barriers

Deployment requires stable video access, synchronized auxiliary signals when used, and interpretable output at the point of care. In activation-associated detection, generator availability and timestamp alignment determine which video interval is examined [12]. Systems should communicate unavailable or unassessable states when inputs fail or image quality deteriorates because of smoke, contamination, occlusion, camera movement, or interrupted capture. Interface placement must also balance visibility with distraction and anatomical obstruction. These requirements make acquisition reliability, failure communication, and interface design part of clinical integration.

5.4. Human Factors and False-Alert Burden

Checcucci’s study reported a median of 85 anticipated bleeding episodes per patient (IQR 70–90) and a 3% false-positive percentage over a median operative time of 94 min, approximately one warning per minute [7]. The report does not provide the denominator or deduplication rule needed to interpret these figures. If 3% refers to distinct warnings, the implied number of genuine precursors is unusually high; if repeated warnings mark the same event, the reported percentage does not describe alert reliability. The study therefore establishes intraoperative feasibility more clearly than an interpretable alert rate.

Ribbens and colleagues analyzed 2,623 activation fragments from 45 operations, including 21 device-induced bleeding events [12]. At 81.0% sensitivity and 4.6% precision, approximately 17 true events would be detected alongside roughly 350 false positive classifications, or about eight per operation. Because analysis already excluded most non-activation periods, continuous deployment could face a larger alert denominator. This burden is more clinically informative than sensitivity alone.

Surgeon-rated overdetection provides a complementary measure of how outputs are perceived [9]. Alert persistence, dismissal, and meaning should fit the operative task and distinguish visible blood from a recommendation to intervene. Prospective interface studies should therefore evaluate interruption, learning, overreliance, and responses to failure in addition to model discrimination.

5.5. Regulatory Pathway

An intraoperative bleeding alert is software as a medical device, subject to US Food and Drug Administration oversight. The evidence levels used here map onto that framework only partially. L1 offline performance supports analytical validation but no clinical claim; L2 streaming demonstrations address integration rather than clinical validity; L3 prospective operation approaches submission-grade evidence only when intended use, alert threshold, and reference standard are fixed in advance rather than interpreted afterward. The claim itself sets the burden: a system that marks visible blood is a narrower claim than one indicating a need for hemostasis, and a narrower claim is easier to substantiate and less useful. Because retraining and site adaptation change performance, a predetermined change-control plan carries as much weight as the initial evaluation. Teams should therefore settle the intended-use statement before designing validation, since an exploratory study reinterpreted after the fact rarely satisfies reviewers or regulators.

6. Future Directions

6.1. Transferable Context and Precursors

The key question is which contextual inputs improve recognition beyond blood appearance. A direct test would hold the detector, dataset, and event definition fixed, add instrument activity, temporal history, operative phase, or synchronized signals one at a time, and compare event sensitivity with false alerts per operative hour. Inputs that do not improve this trade-off add complexity without demonstrated value. Externally supplied context remains largely untested for recognition; in simulated response planning, Zargarzadeh and colleagues combined a scene image with human-provided pool conditions and textual priority rules [20]. Cross-procedure testing can further distinguish transferable bleeding precursors from procedure-specific motion patterns, while surgeon disagreement about actionability can be modeled as uncertainty rather than forced into a binary label.

6.2. Shared Resources and Adaptation

A useful first benchmark is narrower than a general one. A workable specification is a single procedure, complete unedited operations rather than selected bleeding clips, onset and offset annotated by two surgeons under a declared disagreement rule, and a held-out institution fixed in advance. Results reported as event sensitivity and false alerts per operative hour on that set would be comparable across groups, which current reports are not. Controlled-access hosting, in which models are sent to the data and only metrics are returned, permits such comparison where video cannot be released. Open problems include adapting to new cameras or devices with few labels, detecting unfamiliar appearances, and deciding when a system must withhold an unreliable output. These questions extend beyond improving a score on the original dataset.

6.3. Prospective and Surgeon-Centered Studies

Prospective silent-mode studies can characterize live performance before alerts influence behavior, followed by surgeon-visible comparisons that test recognition, response, and interruption burden. Multicenter studies can then assess transfer across operators, procedures, and rare severe events.

Practical design constraints include data-use agreements for operative video, annotation cost for complete non-bleeding intervals, and the ethics of withholding a potentially useful but unvalidated alert. Controlled-access data environments can reduce sharing barriers, while prespecified interim analyses and stopping rules can govern progression from silent to visible deployment. These design choices should be planned with institutional review boards before prospective testing.

6.4. Limitations of This Review

This targeted narrative synthesis includes neither systematic risk-of-bias assessment nor quantitative pooling. Its literature base is small because the field is small, as the screening counts in Section 1 indicate, but selection among eligible reports was purposive and a systematic search might recover additional borderline work. The included studies span heterogeneous targets, datasets, and evaluation units, and related reports may share development data. The evidence levels describe testing settings rather than methodological quality. The review therefore provides a framework for interpreting representative evidence, not a ranking of systems or an estimate of the field’s overall maturity.

7. Conclusion

AI-based video analysis and synchronized signals support bleeding anticipation, detection, and post-hoc assessment at different stages of validation. Separating temporal purpose from input source clarifies what each system recognizes, while explicit evidence levels distinguish offline performance, streaming feasibility, and intraoperative evaluation. Event-level accuracy, false-alert burden, and response time connect technical outputs to clinical use. Generalization, reliable integration, and surgeon interaction remain barriers; prospective comparisons are needed to establish whether alerts improve surgical response.