AvaSure

Real computer vision doing consequential work, layered onto a platform that existed and functioned before it.

The honest starting point is that virtual sitting worked without artificial intelligence for a decade. A remote observer watching a bank of feeds and speaking to a patient through two way audio is the original product, and it delivered outcomes on that basis alone.

What the models change is the ratio, and the ratio is the economics. Computer vision alerting an observer to potential falls and other high risk events is what makes one person covering 16 patients defensible rather than negligent, because sustained human attention across sixteen simultaneous video feeds is not something a person reliably provides. The intelligence is buying attention allocation, and attention allocation is the whole safety argument. Ambient sensing, predictive modelling and a dedicated agent for virtual care workflows have since been added on top.

What sits around it would retain substantial value with the models removed. Cameras, two way audio, nurse call routing, Epic integration, device form factors and the clinical education programme delivered by the company's nursing team are infrastructure, workflow and services rather than inference.

Graded B on the same reasoning applied to platform vendors elsewhere in this index: genuine intelligence performing work that matters, inside a product that is more than the intelligence. The company's own language, describing AI augmented and AI enabled workflows rather than AI native ones, is an accurate self description.

The clearest and most measurable oversight model recorded in this index, because the loop actually closes and someone published the time it takes.

The design is human in the loop by construction rather than by policy. Computer vision alerts; a human observer assesses; that observer speaks to the patient and summons staff. The observer cannot physically intervene, which means the system's entire value depends on how fast something seen becomes someone arriving, and the architecture is built around that fact rather than around detection alone.

The closing of the loop is what earns the grade. Urgent alarms route in real time to nurse call systems, and caregiver acknowledgements sync back into the platform. That makes response auditable rather than assumed: a hospital can see which alerts were acknowledged, by whom and how quickly, which is the accountability layer that monitoring products in this index almost universally lack.

And it produces a published number. One health system reported an 11 second average response time alongside a 38 percent reduction in urgent alarm rates. Reducing alarm volume while measuring response speed is the correct pair of metrics for this problem, because it demonstrates the system is addressing alert fatigue rather than generating it, which is the recognised failure mode of continuous monitoring.

The ratio is stated plainly at up to 16 patients per observer, so a buyer knows the supervision density being purchased rather than inferring it.

What is still missing is detection calibration: no sensitivity or false alarm rate is published for the computer vision itself, and no vendor stated escalation requirement exists for what a customer must staff.

Ask for detection sensitivity, the alert volume per observer shift, and the acknowledgement rate across the installed base.

Hardware specified in genuine detail and the model layer described only by label.

The physical disclosure is thorough. Cameras are 1080p with 10x optical zoom and 360 degree pan, tilt and zoom, with infrared illumination for night vision. Form factors are enumerated with their purposes: fixed ceiling units that are tamper resistant and available ligature safe, wall or television mounted dual camera units pairing a wide angle ambient lens with a consult lens, and mobile carts. Power, privacy and wireless indicators are described. A buyer can determine what the equipment physically does before purchase.

The software is characterised rather than specified. Computer vision, human in the loop alerting, ambient sensing, predictive modelling and a dedicated agent for virtual care workflows are all named as capabilities, and none carries a description of what it detects, how, or how well. Ambient sensing in particular is used repeatedly without any statement of which signals are captured or derived.

No performance figure exists anywhere in the model layer. No detection sensitivity, no false alarm rate, no operating point, no model card, no architecture, and no statement of which detections are model generated versus rule based.

That last distinction matters commercially. A hospital evaluating whether to pay for AI enabled applications on top of the base platform needs to know what the models add over threshold and motion logic, and nothing published separates them.

Ask which events are model detected, at what sensitivity, and what measurable difference the AI enabled tier makes over the base platform.

A platform explicitly built to host other parties' models, naming none of them.

A dedicated pass located no cloud or infrastructure provider, no base or foundation model, no computer vision framework, no sub processor register, no device manufacturer and no position on whether patient video contributes to model development.

The open platform positioning is what makes this a genuine gap rather than a routine omission. The architecture is promoted as open and extensible, connecting AI enabled applications so that each new integration extends the smart room, which means third party models are intended to run against continuous video of patients in hospital beds. Who those parties are, what they receive, where they process it and under what terms is entirely undisclosed. A hospital adopting the platform is adopting an extension point whose future occupants it cannot evaluate.

The hardware chain is likewise unnamed. Devices carry the vendor's own brand across several form factors with detailed optical specifications, and no manufacturer, camera module or silicon supplier is identified, which for connected cameras in patient rooms is a supply chain question with security and provenance dimensions.

The training question carries particular weight given the material. Video of patients under continuous observation, including behavioral health patients and confused or agitated patients, is recorded in circumstances where meaningful consent is often impossible, and nothing states whether any of it has been used to develop the detection models.

Ask for the sub processor register, how third party applications are vetted and what data they receive, the device manufacturer, and whether patient video trains models.

The largest evidenced deployment base in this index, with a peer reviewed longitudinal study and named quantified customer outcomes behind it.

Scale first, because it is verifiable. More than 1,200 hospitals, fifteen years in market, and participation in federal health system procurement. A deployment base of that size across that duration is itself evidence that the product survives contact with real acute care operations, which many products in this index have never been tested against.

The publication is the part that matters most. A peer reviewed study published in early 2026 evaluated a continuous virtual video monitoring programme in a long term acute care hospital over 40 months, reporting sustained improvement in both patient safety and financial performance. Forty months is long enough to capture programme decay, staff turnover and novelty effects, which is precisely what short pilots miss, and the setting had previously been noted as lacking long term evidence.

Named customer outcomes carry specific figures rather than adjectives. One health system reported a 38 percent reduction in urgent alarm rates with an 11 second average response time. An urban safety net academic hospital reported emergency department holding time down 16 percent and discharge turnaround improved 31 percent. Another reported 3.2 million dollars in measurable impact, and a separate academic centre reported more than a million dollars saved in a year. More than 150 customer generated presentations and studies exist, and a third party analyst firm has recognised the company in cost of care.

The qualification is provenance. Most of that body is customer generated and vendor hosted rather than independently designed, and headline claims of adverse event reduction above 50 percent carry no single stated method.

Ask for the peer reviewed study directly, and for outcome data broken out by application rather than aggregated.

Tangible patient facing privacy controls built into the hardware, with the retention question left open.

The controls are physical and visible, which is rarer and better than a policy commitment. Devices carry power, privacy and wireless indicators so the patient can see when the camera is active rather than being asked to trust that it is not. Privacy modes allow the feed to be blanked when a clinician performs an examination or a patient uses the bathroom, which addresses the single most objectionable moment in continuous room observation. Two way audio means the observer's presence is announceable rather than covert. Ceiling devices are tamper resistant, and ligature safe configurations exist for the population where a device itself could become a hazard. A federal hardware variant carries validated cryptography.

Infrared illumination for night vision is disclosed rather than buried, which matters because a patient who believes darkness confers privacy is mistaken and ought to be told.

The unanswered question is retention, and in this category it is the central one. Whether video is observed live and discarded, or recorded and stored, determines the storage security burden, the retention limits, the access logging obligations and the discovery exposure. Nothing published states the default, whether it is configurable, or who can retrieve historical footage. No encryption, access control or deletion description was located either.

The training position is also unstated, which matters for a platform now described as hosting AI enabled applications on patient video.

Ask whether the default is live only or recorded, retention length, who can retrieve footage, and whether video contributes to model development.

Agreements demonstrably exist at scale and nothing about the posture is published.

The situation here differs from the consumer facing sensing vendors in this index, and the difference is worth stating rather than glossing. Every customer is a hospital, which means every customer is a covered entity, which means business associate agreements are executed across more than 1,200 institutions and have been for fifteen years. There is no ambiguity about whether the regime applies, as there is for a product sold through retail channels to families. Federal health system procurement adds a further layer of contractual and privacy review that the vendor has evidently satisfied.

What is absent is disclosure. No agreement template, no negotiation stance, no execution requirement, no subcontractor flow down position and no statement of scope was located.

One dimension specific to this product is unaddressed and matters more than the template. Continuous video of patients in beds, including behavioral health patients under ligature risk observation and patients during personal care, is among the most sensitive material any vendor in this index handles. Whether video is retained at all, and for how long, determines the entire disclosure and discovery picture, and nothing published states the default. Live only observation and recorded observation are very different products from a privacy and legal standpoint.

Ask for the agreement template, the retention default, and whether recording can be disabled at the organisation level.

One externally validated credential visible through a procurement record, and no attestation covering the platform.

The credential is real and specific. A federal variant of the hardware carries validated cryptography under the United States federal standard for cryptographic modules, which is an independently tested certification rather than a self assertion, and it appears in a government solicitation rather than in marketing. Federal health system deployment implies the vendor has passed a security review process, and the procurement record shows the buyer specifying support tiers and software versions in a way consistent with managed patching.

What is absent is any attestation covering the platform as a whole. No information security report, no health specific certification, no trust center, no penetration testing statement, no vulnerability disclosure policy and no documentation offered under agreement was located.

The surface deserves that attention. These are network connected cameras with microphones, permanently installed in patient rooms across more than 1,200 hospitals, integrated into nurse call systems that are life safety infrastructure. A compromised camera in a patient room is a surveillance device in a hospital, and a compromised alarm path is a failure of an emergency system. Firmware update practice and device authentication across an installed base of that size are first order questions and are unaddressed publicly.

The open platform positioning compounds it, since third party applications gain access to the same video and device layer.

Ask whether a security attestation exists and can be shared, how devices authenticate and receive firmware, and how third party applications on the platform are vetted.

A defensible non device position, undeclared, in a product whose highest stakes application sits under a different regulator entirely.

No clearance, grant or approval was located and none is claimed. The positioning supports that: this is observation and communication infrastructure where a human observer assesses every alert and a clinician acts, so the software neither diagnoses nor treats. Video monitoring with a person in the loop is a considerably cleaner non device story than autonomous detection would be.

The complication is that the models increasingly do more. Computer vision flagging fall risk, predictive modelling and an AI agent for care workflows move toward clinical decision support, and no determination is published for any of them.

The more significant regulatory dimension here is not the device regime at all, and this record should say so plainly. Video monitoring of patients at high risk for suicide operates under hospital accreditation standards, with published guidance on ligature risk reduction and continuous observation, and the company's material references that framework. Accreditation compliance is what a hospital's risk officer will actually examine, and a failure in that application is a sentinel event rather than a regulatory filing question. Ligature safe device configurations indicate the standard has been engineered for.

Federal health system procurement implies a further layer of review the vendor has satisfied.

Ask for the written device determination covering the newer predictive applications, and for documentation of conformance with the accreditation guidance on continuous observation.

A dedicated pass located no fairness testing, no subgroup performance, no calibration data, no validation methodology and no external audit.

The bias mechanism for camera based detection is well documented and applies directly. Computer vision systems reading human bodies carry differential error by skin tone, and that effect is amplified in exactly the conditions this product operates in: low light rooms at night under infrared illumination, where contrast and reflectance behave differently again. Body habitus, bedding, positioning aids and restraints all alter what the camera sees. Patients with contractures, tremor, hemiparesis or involuntary movement present motion patterns unlike a typical training subject, and those patients are disproportionately the ones placed under continuous observation in the first place.

The population selection makes this sharper than it would be in a general setting. Patients are put on virtual observation precisely because they are confused, agitated, at high fall risk or at risk of self harm, so the deployed distribution is systematically the atypical one.

The highest stakes application carries no published performance at all. Behavioral health observation for ligature risk is where a detection failure is catastrophic and unrecoverable, and no sensitivity figure, subgroup breakdown or validation description exists for it publicly.

Graded D because nothing was located, notwithstanding the mitigating fact that a human observer sits between every model output and every action, which limits how far a biased detection can propagate unchallenged.

Ask for detection performance by skin tone and under infrared conditions, and for validation specific to the behavioral health application.

Nothing contractual is published, in a product deployed against falls, self harm and workplace violence.

A dedicated pass located no service level agreement, no detection warranty, no uptime commitment, no indemnity and no remediation position. A federal solicitation references support tiers and a help desk requirement, which indicates service commitments exist in contracts, and none of their content is public.

One structural feature genuinely does clarify responsibility here, and it should be stated because it distinguishes this record from others graded the same. A hospital employee observes every feed, assesses every alert and decides every intervention. Clinical responsibility therefore rests unambiguously with the hospital rather than being distributed between a model and an absent vendor, which is a cleaner allocation than autonomous detection produces. The published customer figures on acknowledgement and response times exist precisely because that human layer is accountable and measured.

That clarity does not extend to the vendor's own obligations. If the computer vision fails to flag an event, if a device is offline, or if the alarm path to nurse call breaks, nothing published states what is warranted or what recourse follows. No detection rate exists, so residual risk cannot be characterised, and a hospital replacing one to one sitters with a 1 to 16 ratio is making a staffing decision on unpublished performance.

One pre emptive note: further customer outcome studies cannot move this grade. Only a contractual commitment, or published detection and alarm delivery performance, will.

Ask what the agreement warrants on detection, device uptime and alarm delivery, and what happens when the alarm path fails.

Named systems, a named standard, a stated direction of flow and a closed loop, which together make this the strongest interoperability disclosure in this index.

The record system integration is specific rather than gestural. The virtual visit application launches directly from within the Epic clinical desktop, and devices pair with the Epic monitoring module for virtual observation. Launching inside the record system rather than beside it means a nurse does not leave their workflow to start a virtual interaction, which is the difference between an integration and a second application.

The nurse call path is described with mechanism. Urgent alarms route in real time over HL7 interfaces, application programming interfaces or middleware, which names the standard and acknowledges the reality that many hospitals sit behind an integration engine.

The acknowledgement path is what lifts this to the top of the axis. Caregiver acknowledgements sync back into the platform, closing the loop. Almost every monitoring vendor in this index can push an alert outward; this one can tell a customer whether a human received it and responded, which converts an alerting product into an auditable accountability system and makes the published response time figures possible at all.

The architecture is described as open and extensible, hosting third party applications on the same device and data layer.

What is absent is documentation depth: no public interface specification, no marketplace listing was located, and no statement addresses whether observation events are written to the record as discrete entries.

Ask for the interface specification, and whether observation events and interventions post to the chart.

The on premise footprint is substantial and evidenced, and the hosted side is undescribed.

What can be established is more concrete than for most vendors, largely because a federal procurement document exposes it. Devices are physically installed in patient rooms in fixed and mobile forms, monitor stations sit in the facility, and at least one device software version is documented as requiring a Microsoft server operating system, which indicates real on premise server components rather than a purely hosted architecture. Flexible deployment options are referenced, and a federal hardware variant carries validated cryptography.

A substantially on premise architecture is a meaningful residency answer in itself for a product handling continuous patient video, because video that never leaves the building raises no cross border question.

What is missing is everything about what does leave. No cloud provider is named, no region is stated, no residency commitment is made, no tenancy model is described for whatever is centrally hosted, and no statement distinguishes which components run locally from which run remotely. Analytics reporting, an AI agent and predictive modelling all imply some centralised processing, and none is located.

The open platform positioning adds a further question, since third party applications running on the same data layer may process elsewhere entirely.

International deployment is referenced without any accompanying statement about how non United States installations are handled.

Ask which components are on premise versus hosted, where hosted processing occurs, and whether video ever leaves the facility.

No rate card exists, and the commercial structure is nonetheless visible in the open because a federal buyer published it.

A United States Department of Veterans Affairs solicitation sets out the line items a hospital actually buys. Software is licensed per hardware device on an annual term at a named tier, with a support level carried separately and bundled inclusions enumerated: an analytics reporting tool, pre recorded announcements in 30 languages, clinical programme materials and eLearning modules. Hardware is a separate line, with a federal variant carrying validated cryptography. Clinical services are a distinct line covering policy and procedure planning, programme development and onsite go live support. A monitor station is separately priced. Installation runs 180 days from receipt of order.

That is the pricing basis, the term, the unit and the unbundling, all without a dollar figure, and it is materially more than most vendors in this index disclose deliberately. A buyer can construct a line item budget and knows to ask about the software tier, the support tier and the clinical services scope as three separate negotiations.

The return side is also better served than usual. Published customer results carry dollar figures at named institutions, and the cost of falls basis is attributed to a cited external source rather than asserted, which lets a buyer substitute their own fall cost and recompute.

What is missing is any price. No figure appears for the per device licence, the hardware, the support tiers or the clinical services, and nothing indicates volume behaviour across a 500 bed deployment.

Ask for the per device annual licence, what separates the support tiers, and whether clinical services are mandatory in year one.

Breadth across acute care that few vendors match, with the distribution of applications across it unstated.

The range is real and differentiated at the hardware level, which is stronger evidence than a list of served settings. Ceiling mounted tamper resistant devices in ligature safe configurations exist specifically for behavioral health. Mobile carts serve the emergency department and spaces outside the patient room. Dual camera fixed units serve rooms needing both ambient capture and consults. Building distinct form factors for distinct settings indicates the settings were actually engineered for.

Named populations and use cases span medical surgical, intensive care, emergency, behavioral health, long term acute care, pediatrics, isolation precautions and workplace violence, with the peer reviewed evidence coming specifically from a long term acute care setting rather than a general one. Applications span continuous observation, virtual nursing, episodic consults, virtual visits and ambient monitoring. Federal health system deployment adds a distinct procurement context.

What is missing is the breakdown. More than 1,200 hospitals are claimed and nothing states how many run virtual nursing versus observation only, how many have progressed beyond a single unit, or what proportion use the AI enabled applications rather than the base platform. The company's own published research notes that only a small minority of hospital leaders have reached a stage where virtual care is standard practice, which suggests much of the installed base sits earlier in that curve than the headline number implies.

Ask what proportion of the installed base runs each application, and how many units a typical customer has deployed.