Egocentric medical and surgical procedure datasets A research and sourcing guide to clinician-worn procedural video, including camera perspective, authenticity, expert labels, permissions and evaluation. Egocentric surgical datasets show a procedure from a clinician-worn viewpoint. They differ from endoscopic, microscope, overhead and lecture recordings, even when all are described as surgical video. Appropriate source selection depends on the specialty, visible actions, model task, authenticity and required expert annotations. Real patient care and emulated procedures should remain distinguishable. Source authority, participants, handling and permitted uses require review before sharing material. Public EgoSurgery-Phase and EgoExOR papers illustrate different research designs rather than a general commercial inventory. This guide explains how to evaluate phase, hand/tool and workflow evidence without implying clinical validation or unrestricted access. Primary data Egocentric medical and surgical procedure Commercial model Project-specific source and permission review Price evidence Named public comparisons; no invented market average Reviewed 12 November 2026 What this category includes A surgeon-worn view is distinct from an endoscope, overhead camera or narrated lecture; compare phase, hand/tool and team-context evidence without implying unrestricted clinical data. Surgical workflow research and assistive procedural understanding. Phase recognition and clinically informed hand/tool interaction evaluation. Identify the actual camera perspective Clinical video can come from an endoscope, microscope, overhead camera, wall camera or a clinician-worn device. Those viewpoints are not interchangeable. An egocentric surgical dataset specifically follows the wearer’s perspective and can include head movement, changing attention and occlusion from hands or assistants. A narrated procedure is not necessarily an egocentric recording. Establish the source device and what it actually sees before selecting a task or assigning a category label. The perspective determines what can be annotated. A head-mounted view may show instruments and workflow while failing to expose a decisive anatomical detail. An endoscopic image may show internal anatomy but omit the clinician’s surrounding actions. If a project needs both, specify paired streams and their relationship. Do not merge unrelated views into a single presumed clinical experience. The delivery manifest should distinguish each camera, procedure and recording segment. Define the model task before collecting more footage Surgical phase recognition, hand/tool tracking, interaction analysis and procedural question answering require different evidence. A phase label describes where an operation sits in its workflow; it does not necessarily identify a precise action or confirm a clinical outcome. Hand/tool labels may be useful without patient diagnosis information. Start with the intended research or product task and the minimum information needed to evaluate it. Define the specialty, procedure type and setting. An open-surgery benchmark is not a general corpus of dentistry, endoscopy or emergency care. If a model is intended for a particular procedure, the source should demonstrate its phases, instruments and meaningful variation. Avoid using the word medical as a substitute for a specification. The page describes a category for scoped sourcing and research comparison, not an assurance of access to every clinical environment. Separate authentic care from emulated procedures [EgoSurgery-Phase](https://papers.miccai.org/miccai-2024/264-Paper0627.html) provides a real open-surgery research comparison. [EgoExOR](https://arxiv.org/abs/2505.24287) studies paired views of emulated spine procedures. That distinction should remain prominent. Emulated work can offer controlled conditions, repeatable events and flexible instrumentation, but its distribution and consequences differ from actual patient care. A buyer should know which kind of source is being proposed. An evaluation can include both authentic and emulated data if it reports them separately and explains the intended role of each. A successful result on an emulated procedure does not establish clinical effectiveness. Likewise, a retrospective recording does not automatically provide ground truth for all decisions made during the operation. Specify what the record supports, what expert annotation adds and what remains unknown rather than treating clinical realism as a single yes-or-no quality score. Permissions and handling are part of the source definition A source review must establish who controls the recordings, which participants and sites are involved and which uses are proposed. Possession of a video, public accessibility or the ability to narrate a procedure does not itself establish authority to license it for model training. Discuss the proposed use and appropriate handling before transferring recordings or participant documentation. The relevant requirements depend on the actual source and jurisdiction; a category page cannot decide them in advance. Potentially sensitive information may appear outside the central procedure: names spoken aloud, charts, room screens, labels, staff conversations or incidental people. Review the entire retained source, not only a small anatomical crop. Redaction can also change learning value or timing evidence, so document preparation decisions and meaningful exclusions. Do not publish private case details, supplier correspondence or sample links as proof of expertise. Public research comparisons can explain the category without exposing private leads. Expert labels need a clinical meaning and a review protocol A phase taxonomy should define entry and exit conditions, ambiguous transitions and events that can occur more than once. Instrument presence is not always equivalent to instrument use. If hand/tool contact, anatomical targets or decision rationale are required, specify their relationship to visible evidence. Use unknown or uncertain states when the video cannot support a confident label. Describe annotator qualifications and how disagreements are handled without implying a clinical endorsement. An expert can supply retrospective interpretation that was not observable in the original view. Keep that interpretation distinguishable from direct visual labels and from actual care records. If questions and answers are included, preserve supporting intervals and document when an answer depends on external information. A valid training record should reveal these dependencies rather than making every confident sentence look equally grounded. Temporal and procedure-level evaluation matters Split at the level of procedures, teams or sites according to the intended generalization. Neighboring frames from one operation are strongly related; randomly distributing them between train and test can create an easy but uninformative task. If the same recording appears as a full procedure, shorter clip and narrated edit, those derivatives should remain in one lineage group. A phase benchmark also needs an explicit treatment of incomplete recordings. The operation may begin before capture starts, pause during a critical interval or end after recording stops. A model’s inability to find a missing phase is not necessarily a recognition failure. Evaluation should distinguish absent observation, unknown status and visible but misclassified events. Report the source restrictions of a test set alongside the metric so readers understand what the result can and cannot imply. Clinical capture quality is task dependent High nominal resolution cannot compensate for an unsuitable view. Inspect the decisive moment for motion blur, occlusion, lighting, lens contamination and attention changes. A stable overview may support broad phase recognition while being inadequate for fine hand/tool pose. Audio and narration can add context, but their availability should be tested rather than assumed from the word narrated. For multimodal work, specify clock alignment, camera calibration, pose representation and any external measurement. Headset tracking is not a clinical outcome label. A synchronized microscope or ultrasound stream has its own coordinate and annotation requirements. Source files, derived outputs and corrections should have stable identifiers so an expert review can trace a disputed label back to the actual observation. Keep technical validity separate from clinical relevance and from permissions. There is no defensible generic clinical-data price here No verified public comparable commercial price for this exact category was established in the checked sources. The cost of an ordinary video archive cannot be applied to clinical recording access, governance, expert labels or de-identification. A commercial inquiry should identify the procedure, viewpoint, authenticity, allowed uses and preparation before a meaningful quote can be compared. A public research benchmark may be accessible without a purchase while still restricting commercial use or redistribution. That is different from sourcing authorized material for a specific commercial project. Likewise, a company investment or medical AI acquisition is not a dataset licensing transaction. Where contract consideration is not public, say undisclosed rather than filling the gap with a speculative per-hour rate. The appropriate next step is a bounded source review and representative acceptance plan. A careful first inquiry Send Fiund a high-level description of the model task and procedure before sending clinical media. Include the desired viewpoint, whether authentic or emulated material is acceptable, required annotations, intended uses and the evidence a sample must demonstrate. State any security, governance or review requirements that affect handling. Do not include patient identifiers or private participant files in an initial informal brief. A feasible sourcing discussion can then distinguish source access, preparation, expert review and commercial terms. It may identify missing information or show that a proposed source does not fit. That is a useful outcome: it prevents a broad category claim from becoming a misleading promise. This guide does not claim that Fiund currently holds a clinically authorized ready-to-license corpus, and it does not establish diagnostic performance, clinical safety or regulatory acceptance for any model. Example: a phase label with incomplete evidence If recording starts after an operation has begun, the first visible action may occur in the middle of a phase. The label can describe the observed interval without implying that the complete phase was captured. A review record should indicate the missing beginning and any uncertainty about transitions. The same principle applies to narration: a later explanation can clarify context while remaining different from direct observation at the moment of action. Keeping those distinctions visible prevents a technically neat dataset from overstating what its source proves. It also helps a buyer decide whether additional expert review or a different recording view is necessary. A specification your sample can demonstrate Write the acceptance record before reviewing a large archive. The checks below identify the evidence to request; they do not mean that a proposed collection has already passed. Record exceptions at the file or episode level so a useful sample cannot hide problems elsewhere in the release. Exact viewpoint, real versus emulated procedure, specialty and phase taxonomy. Clinical/participant/site authorization and allowed uses reviewed before samples. Patient identifiers, speech, charts, background screens and incidental persons addressed. Expert labeling, uncertainty, procedure/session-level splits and per-phase coverage. Published comparison points and access terms These are external research or supplier references. They illustrate technical scale and access conditions; Fiund does not claim ownership of these datasets or a partnership with their publishers. Counts use the named release and unit rather than a claim to measure the entire market. Reference Published scope or offer Important comparison limit [EgoSurgery-Phase paper](https://papers.miccai.org/miccai-2024/264-Paper0627.html) 15 hours across nine open-surgery phases Head-mounted research source; commercial rights not established here. [EgoExOR paper](https://arxiv.org/abs/2505.24287) 94 minutes of two emulated spine procedures Emulated procedures must not be described as authentic patient care. Pricing and commercial evidence No verified public comparable commercial price. Specialist access, governance, de-identification, expert annotation and authorized use drive quote; public benchmark access is not a clinical-data sales offer. A commercial comparison should show what is being licensed, the permitted model uses, the acceptance unit, delivery preparation, exclusivity if any, update obligations and the term. A public benchmark fee, hardware price, annotation invoice and purchase of an entire archive are different numbers. No reliable category-wide average is asserted here; request comparable written scopes before treating two offers as substitutes. What public deals do and do not establish The checked sources do not establish a disclosed commercial purchase value for this precise category. Public research releases show what has been collected, not what an enterprise paid. Bristol’s [EPIC-KITCHENS commercial license listing](https://express-licences.bristol.ac.uk/product/epic-kitchens-100---dataset) supplies a concrete named-dataset comparison; it should not be converted into a custom-capture market average. Discuss sourcing with Fiund Fiund reviews the proposed source, technical specification, permissions, preparation and commercial scope for the project. Send the target task, required fields, intended use and a description of what a representative sample must show to [jaeden@fiund.com](mailto:jaeden@fiund.com). Availability and permitted uses need to be established for the selected material; a category guide is not a promise of ready-to-download inventory. Is a narrated dental procedure automatically an egocentric dataset? No. Narration describes an audio or annotation property, while egocentric describes the camera perspective. Establish whether the camera follows the clinician’s view, what actions are visible and whether the recording includes the timing and labels needed for the task. Permissions for the proposed use remain a separate source-review question. Can a public surgical research dataset be used commercially? Review the exact publisher agreement and the relevant artifact terms. Public access does not automatically permit commercial model development, redistribution or sublicensing. A research paper’s dataset description is not a license grant. Ask the publisher for clarification where the intended use is not established; Fiund does not claim authority to override a third party’s restrictions. What is a sensible sample acceptance standard? Use a procedure-level sample with the actual viewpoint, labels and preparation proposed for delivery. Review critical event visibility, phase definitions, uncertain labels, recording gaps, file lineage and the appropriate permission/handling evidence. The standard should test the stated model task and source conditions. It should not imply clinical validation from a small technical data sample. Sources EgoSurgery-Phase paper: https://papers.miccai.org/miccai-2024/264-Paper0627.html EgoExOR paper: https://arxiv.org/abs/2505.24287 Bristol EPIC-KITCHENS-100 commercial license pricing: https://express-licences.bristol.ac.uk/product/epic-kitchens-100---dataset