Apella

The models are the entire product, which distinguishes this record sharply from the other operations platforms built recently. Ambient sensors observe the operating room, computer vision and machine learning identify what is happening, and the resulting structured data drives scheduling, utilisation prediction and the record write back.

Strip out the perception layer and nothing remains that a hospital would buy. There is no workflow application, no bed management system, no electronic record and no communications platform underneath. Cameras without the models produce video nobody watches. That is a materially different proposition from TeleTracking, Alcidion, AlayaCare and Optum Integrity One, all of which retain a working business without their intelligence, and it is why this grade sits with Fathom and Nym Health rather than with the platform incumbents.

The company's own framing supports it: the founding thesis was that surgery could not be improved because it was not measured, and the product exists to create the measurement. Perception is the value, not a feature on top of it.

The scope of autonomy is stated with unusual precision and nothing published governs it.

What is disclosed is genuinely specific: the system automatically identifies up to fourteen distinct surgical case events and autonomously writes that novel data back into the electronic record. Naming the number of event types is a real boundary statement, better than the vague automation claims elsewhere in this index.

What is absent is everything that would make a buyer comfortable with it. No detection accuracy for any event type. No confidence threshold determining what gets written versus flagged. No description of whether a human reviews before write back, whether entries are marked as machine generated, or how an incorrect case time is corrected once it is in the record.

That combination is the finding. This is the only record in the index describing a system that writes machine derived clinical observations into the legal medical record without a human in the loop, and it publishes no accuracy figure for the perception producing them. Graded C because the scope is named and no oversight mechanism is. Ask for per event detection accuracy, the write back review process, and the correction workflow.

The capability description is concrete and enumerable, which is rarer than it should be. Ambient sensors in the room, computer vision and machine learning applied to the resulting stream, up to fourteen surgical case events automatically identified, structured output written to the record, and a separate forecasting product predicting case duration and utilisation ahead of the day.

Enumerating what the system detects is a genuine disclosure. A buyer can ask which fourteen events, compare them against their own documentation requirements, and judge whether the coverage matches their workflow. Most vendors in this index describe capability in adjectives.

What sits underneath is undisclosed. No model architecture, no training data description, no version, and no foundation model or vision model family named. For a computer vision company the training corpus is the central technical question, since detection performance depends entirely on the range of procedures, room configurations, equipment and patient presentations represented in it, and nothing is said about any of that.

No accuracy, precision or recall figure appears for event detection or for the duration forecasts. Graded B for specific enumerated capability with no specification or measurement behind it.

No party in the chain is named. No cloud platform, no model provider or vision model family, no sub processor list, and no hardware or sensor supplier identified despite the product depending on installed room hardware.

The training data question is the one that matters most here and is entirely unaddressed. A computer vision system detecting surgical events is trained on surgical video, and that video comes from operating rooms containing identifiable patients and staff. Whether customer footage trains the shared models, whether it is de identified first, whether hospitals consent to their procedures improving a product other hospitals buy, and whether any of it is retained for retraining are all central and all unanswered in either direction.

Competitors handling far less sensitive data answer the equivalent question explicitly. One vendor in this index commits contractually that customer data never trains any foundation model; another states it trains only on de identified data or with permission. This record contains no position at all on the most sensitive training corpus in the index.

Ask what the models were trained on, whether customer surgical video contributes to model development, and for the sub processor and hardware supplier registers.

One deployment at genuine enterprise scale, and a conflict that belongs on the record.

Houston Methodist, a nine hospital system, ran an initial 36 room pilot and has scaled to more than 200 operating rooms enterprise wide, with its executive vice president and chief innovation officer quoted on the decision. Scaling from pilot to enterprise across a system that size is the hardest signal to fake in health technology, and it is stronger evidence than any number of logos.

The conflict is disclosed in the same announcement and this record states it plainly rather than passing the reference on unqualified. Houston Methodist is an investor in the Series B, and the quoted executive appears among the company's individual investors. The deployment is real and the reference is not disinterested. This is the second instance in this session of a flagship validation carrying an undisclosed or under disclosed commercial relationship, the other being an accuracy audit conducted by a related party.

What is missing is independent measurement. The claimed average 5 percent increase in surgical volume carries no methodology, baseline definition or study. No peer reviewed publication, third party research assessment or independently validated detection accuracy was located. Ask for a reference site with no equity relationship.

Nothing operational was located. No retention schedule, encryption statement, access control description, data ownership or deletion position, data minimisation commitment, or statement on whether captured footage contributes to model development.

Graded conservatively for the same reason as the privacy axis, and the gap is the widest of any record built this session because the data is the most sensitive. Continuous surgical video captures anaesthetised patients, exposed surgical fields and the clinical team, at more than 200 rooms in the flagship deployment alone.

The questions that follow are concrete and none has a published answer. Whether video is recorded or only processed in the moment with frames discarded. Whether footage leaves the hospital. What is retained after event extraction. Who inside the vendor can view a procedure. Whether captured video trains the models other customers benefit from. How a patient who objects is handled, and whether the question is ever put to them.

A vendor placing cameras in operating rooms carries a heavier obligation to publish here than an ambient documentation vendor recording a clinic conversation, and currently publishes nothing.

No health privacy material was located in a dedicated pass: no compliance statement, control enumeration, de identification position, or business associate agreement posture, template or execution requirement.

Graded conservatively on the basis applied to the other operations records this session. Deploying across more than 200 operating rooms at a major academic health system means privacy review has been satisfied at scale and agreements demonstrably exist.

The scope question here is unusual and unaddressed. Continuous video capture in an operating room records a patient under anaesthesia, unable to consent contemporaneously or to object, alongside the entire surgical team. Staff are captured continuously as a condition of doing their job, which raises employment and works council questions distinct from patient privacy. And surgical video has an evidentiary dimension no other data type in this index carries: footage of a procedure is potentially discoverable in malpractice litigation, and whether it is retained, for how long, and who may subpoena it are questions a hospital counsel would ask before a chief information officer did.

Nothing published addresses any of it. Ask what is recorded versus processed transiently, retention, and the position on discoverability.

No credential was located. A targeted pass found no controls report, no information security certification, no health specific security framework certification and no cloud authorisation. There is no trust center, no security page, no report availability process, no penetration testing disclosure and no vulnerability disclosure policy in the material retrieved.

The absence is very likely publication rather than substance. A nine hospital academic system does not install cameras in more than 200 operating rooms without extensive security review, and an investor of that profile conducts its own diligence.

That does not change the grade, and the gap is wider here than the count suggests. This vendor captures continuous video of surgical procedures and writes autonomously into the medical record, which is a higher risk posture than any other record built this session, and it publishes less security material than vendors handling ordinary documentation. One competitor a fraction of its size publishes a live control report open to any prospective customer.

Graded D on the basis applied to TeleTracking, Alcidion and Andor Health. Ask for the controls report, its type and period, and any health specific certification held.

No device pathway is claimed and on the current product description none appears necessary. The system measures workflow rather than assessing the patient: it records when events occurred, not whether the surgery is going well, and makes no clinical determination.

The regulatory weight sits somewhere less obvious and is unaddressed. Automatically written case events become the timestamps in the medical record, and operating room and anaesthesia time drive billing. A miscounted incision or closure time is not only a documentation error, it is a claims accuracy issue, and errors originate from a perception model whose accuracy is unpublished. Nothing describes reconciliation between machine written times and the clinical team's own account, or what happens when they disagree.

The trajectory is worth watching rather than judging now. A computer vision system observing surgery that moved from counting events to assessing technique, safety or performance would be in a different regulatory category entirely, and the founding language about improving surgical quality and staff training points in that direction.

Graded C because the current position is correct and the billing accuracy implication of autonomous write back is undocumented.

Nothing is published. No bias or fairness testing, no model validation methodology, no monitoring output, no distribution reporting, no drift detection and no external audit was located.

The bias questions specific to computer vision in an operating room are concrete rather than theoretical, and none is addressed. Detection performance in vision systems is known to vary with skin tone, and an anaesthetised patient's exposed surgical field is precisely where that variation would appear. It plausibly varies with body habitus, with room lighting and configuration, with equipment generation, and with how a particular team drapes and positions. A model trained predominantly at one large academic system will encode that system's practices, and the flagship deployment is exactly that.

The consequence is not an alert a clinician ignores. It is a case time written into the record and onward into billing, so systematic detection variation would produce systematic documentation and reimbursement variation across patient populations or facilities, invisible without distribution reporting that does not exist.

Graded D because a system writing autonomously to the medical record publishes no accuracy, no breakdown and no validation. Ask for detection performance by patient and procedure characteristics.

No performance figure is published for the perception layer, so there is no stated level against which a shortfall could be measured. No detection accuracy, precision, recall or error rate appears for any of the fourteen case events, and no accuracy figure exists for the duration forecasts.

No service level agreement, warranty, indemnity or remediation commitment was located, and no pilot terms or validation offer was found beyond the fact that the flagship customer ran one.

The recourse gap is more serious here than the grade alone conveys, because of what the system does with its output. Most products in this index produce a recommendation a human accepts or rejects. This one writes machine generated observations directly into the legal medical record, where they persist, inform billing and are discoverable. An undetected error becomes part of the patient's record with no human having reviewed it, and the provider owns the record.

One pre emptive note for future passes: further volume or capacity figures cannot move this grade. A 5 percent increase in surgical volume measures commercial benefit, not detection accuracy or recourse. Only published per event detection performance, or a contractual commitment on it, will change it.

The functional claim is stronger than most and the specification is absent.

What is described is bidirectional and unusually deep. The platform consumes record system data for scheduling and case context, and writes novel structured data back autonomously, which is a materially harder integration than the read only or human mediated write patterns most vendors in this index describe. Writing machine generated clinical events into a live record requires both technical access and institutional trust that few products obtain.

What is missing is every particular. No record system is named. No interface standard is described. No connection mechanism is specified. No vendor marketplace or certification listing was located, and nothing addresses how the sensor infrastructure connects to hospital networks.

The hardware dimension is unaddressed too. Installed room sensors imply network, power and physical integration with operating room infrastructure, which is a distinct integration class from data exchange and one that determines deployment cost and timeline.

Graded C for credible and unusually deep capability with no published specification, consistent with the treatment of the same absence across this lane. Ask which record systems, through what standards, and what the room installation requires.

Nothing was located on hosting, region, tenancy or residency, and the architecture question here is more interesting than the usual silence.

The product is necessarily hybrid. Sensors sit physically in operating rooms, which is on premise hardware, while analytics, forecasting and model inference plausibly run elsewhere. Where the boundary falls is the whole question and it is unstated. If video is processed on premise and only extracted events leave, the data exposure is modest. If footage streams to a cloud environment for inference, the exposure is very different, and the bandwidth and latency implications are significant across 200 rooms.

No cloud provider, region, tenancy model or customer controlled option was located, and nothing states whether footage ever leaves the hospital network.

Graded C consistent with how silence has been treated across this lane, with the note that the on premise versus cloud inference boundary is the single most useful thing this vendor could publish and would move this grade immediately. Ask where inference runs, whether raw video leaves the building, and what the room hardware footprint requires.

No price is published and one usable economic anchor is.

The company states an average 5 percent increase in surgical volume across customers, and pairs it with a framing figure: up to 60 percent of a health system's revenue and 40 percent of its costs sit in the operating room. Those two together let a buyer model the benefit side against a number they already know precisely, which is the same construction that lifts several records in this index above D. A system with a known perioperative revenue base can calculate what 5 percent means before speaking to sales.

Everything on the cost side is absent. No pricing page, no unit of charge, no range, no implementation or hardware fee position, no minimum commitment and no pilot terms were located.

The hardware dimension makes the omission more consequential than usual. The product requires sensors installed in every operating room covered, so the cost structure has a capital component that scales with rooms, and a buyer scaling from a pilot to 200 rooms faces a materially different number from a software only purchase. Nothing addresses whether hardware is purchased, leased or bundled. Ask for the pricing basis per room, the hardware model, and what the pilot to enterprise step change costs.

Deliberately narrow, and narrow by design is not the same as broad.

Coverage is the perioperative environment: hospital operating rooms and ambulatory surgery centres. The company is explicit that this focus is the strategy rather than a limitation, on the reasoning that the operating room concentrates more revenue and cost than any other part of the health system, and the Houston Methodist deployment at more than 200 rooms demonstrates it scales within that setting.

What is not published is any breakdown inside it. No surgical specialty enumeration exists, which matters more for a computer vision product than for a software one: detection performance plausibly differs between an open cardiac case, a laparoscopic procedure and an ophthalmic one, given different room setups, equipment, draping and staff positioning. Nothing addresses whether all specialties are supported equally or whether some are excluded.

The chief executive has said the company is being pulled into other areas of the hospital, and nothing has shipped there. Graded C for a single well executed setting with no internal coverage detail. Ask which surgical specialties are supported and whether detection accuracy varies across them.