care.ai

Scoped to the care.ai product line, the intelligence is the point: computer vision over camera sensors is what makes a room aware, and no amount of hardware produces the capability without it.

Held at B because of where that product now sits. It is one module of a platform spanning more than 85 products from a device manufacturer, most of them physical, and the surrounding value comes from beds, badges and alarm middleware rather than from models. A hospital buying the ambient layer is buying into an ecosystem, and the further into that ecosystem it goes the smaller the share of what it owns that is a model.

A human remains in the loop and the system decides how thinly that human is spread, which is the fact worth holding onto.

Stryker states that a remote nurse can monitor, assess and support up to ten patients simultaneously using ambient data. Publishing that ratio is unusual and to the company's credit, because it makes the proposition legible: the ambient system is not replacing the nurse, it is making one nurse sufficient for ten rooms by deciding which room needs attention now.

That places the automation upstream of the human rather than beside them. Held at B because nothing published describes what happens when two rooms need attention at once, what the system does when it is uncertain, or how the ratio was arrived at.

The architecture is described more specifically than the models are. Camera based sensors carry graphics processing units on the unit, so inference happens locally, and the capability set is enumerated across virtual nursing, virtual monitoring, ambient sensing and contextual data.

No model, training description, detection accuracy or false alarm rate is published for any of it. For a system whose output determines which of ten rooms a nurse looks at, the false negative rate is the number that matters and it is absent.

The architecture bounds this axis for the sensing tier and nothing in the chain is named. Camera based sensors carry graphics processors on the unit, so inference happens locally rather than in a vendor estate, which limits what leaves the building and is a structural answer rather than a policy one. The capability set is enumerated across virtual nursing, virtual monitoring, ambient sensing and contextual data, so a buyer can see what the platform claims to do.

On enumeration there is nothing: no model or model family, no provider, no hosting arrangement for whatever platform tier exists beyond the unit, and no sub processor list was located, along with no training data description for any detection capability. The virtual nursing capability deserves a specific question that the on device claim does not cover, and it is the same one recorded for a peer in this lane.

Virtual nursing involves a remote person on a video link into the room, which is a different data path from local inference on a sensor, and nothing published describes where that stream goes, who can join it, whether it is recorded, or how long anything persists. An architectural claim made about the sensing tier should not be read as covering the tier where a human is watching. Ask what runs on the unit against what leaves it, capability by capability, and for a sub processor list covering the remote monitoring path.

No published outcome study, controlled comparison or measured deployment result was located for the ambient product.

The supervision ratio of one nurse to ten patients is an operating parameter rather than an outcome: it describes how the product is configured, not what happened to patients when it was. Corroboration is commercial rather than clinical, resting on acquisition by a major device manufacturer and inclusion at the centre of that company's platform strategy.

The comparable independent vendor in this index publishes striking figures from a named academic health system and has also not published them in the literature, so the whole segment is in the same position: widely adopted, expensively built, and unmeasured in public.

Graded on an honest basis, and the concern is identical to the one this index recorded on the comparable independent platform.

Continuous sensing of a patient room captures people who are confused, sedated, undressed or dying, along with every family member and staff member who enters. Nothing published states retention, whether footage is stored at all rather than processed transiently, what patients are told, whether they may decline, or whether recorded material contributes to model development.

On premise processing narrows the exposure, since inference at the unit means less leaves the building, and it does not answer any of the questions above.

Graded on an honest basis. No compliance statement or agreement posture specific to this product line was located in this pass.

The acquisition changes who a hospital contracts with, and a buyer should establish whether agreements signed with the independent company have been novated, what changed in them, and how the platform's data flows relate to the other Stryker systems now sharing the same interface.

Recorded honestly and provisionally: the dedicated trust and security search this index requires was not run in this pass, and no attestation specific to this product was encountered.

The parent is a listed medical device manufacturer, so its annual securities filing carries mandatory cybersecurity risk management and governance disclosure, and its device business carries quality system obligations subject to inspection. Neither was checked, and this index has noted before that regulatory rigour over device quality is not the same thing as an information security posture.

No device authorisation is claimed for the ambient intelligence and virtual care workflows themselves, which is consistent with their positioning as observation and workflow support rather than diagnosis or treatment.

The context is unusual for this index because the owner is a major device manufacturer and the platform connects to regulated equipment: beds, physiological monitors, ventilators and nurse call systems all appear in the integration list. Software that filters and prioritises alarms from regulated monitors sits closer to the regulated function than software that watches a room, and a buyer should establish which components carry clearances of their own and which do not, rather than assuming the manufacturer's overall standing covers the whole platform.

Nothing published on evaluation, monitoring, subgroup performance or error handling.

The exposure is the one this index has already recorded twice for computer vision on patients: detection accuracy varies with skin tone, body size, lighting and position, and the patients placed under continuous observation are disproportionately older, frailer, connected to lines or positioned in ways no training set anticipated. A model that performs less well on some of them concentrates its misses in the group selected for observation precisely because they were judged highest risk.

The supervision ratio sharpens it. If one nurse covers ten rooms and the system decides which room to surface, then a missed event is not merely undetected, it is undetected in a setting where the alternative human observation has already been removed on the strength of the system working.

Two passes located no detection accuracy, no false alarm rate, no validation methodology and no warranty, indemnity or remediation commitment for any capability. For a system whose output determines which of ten rooms a nurse looks at, the false negative rate is the number that decides whether the product helps or harms, and it is absent, as is the false alarm rate that determines whether staff keep trusting it. Both directions matter and neither is countable from anything published.

The recording surface is the fifth instance of a shape this index now treats as settled. Continuous sensing of a patient room captures people who are confused, sedated, undressed or dying, along with every family member and staff member who enters, and nothing published states retention, whether footage is stored at all rather than processed transiently, what patients are told, whether they may decline, or whether recorded material contributes to model development.

On device processing narrows the exposure because less leaves the building, and it answers none of those questions, since a recording made and held locally is still a recording. Ask for the false negative and false alarm rates per capability, the retention position for both sensor and virtual nursing streams, and what patients are told.

Strong, and mostly inherited. The platform's communication layer is stated to integrate with more than 150 clinical and operational systems including record systems, nurse call, ventilators and physiological monitors, and the alarm middleware exists specifically to reconcile signals across them.

That breadth came from an earlier acquisition rather than from the ambient product, which is the honest reading: the ambient layer benefits from connectivity somebody else built. Held at B because no interface standard or record vendor certification is described for the ambient and virtual care components specifically.

Processing happens on the unit, with graphics processing hardware attached to the sensors themselves, which is the right architecture for continuous video: it avoids streaming every room to a remote service and keeps the heaviest data inside the building.

That is a real answer to the residency question rather than a claim about one. Held at B because what does leave the site, what is retained centrally, and how the command centre reaches multiple facilities are not described.

Nothing published: no price, no mechanism, no unit of sale.

Two things make that harder here than for a software vendor. The cost includes hardware in every covered room, sensors and local processing, so the real question is per room installed cost, software subscription and refresh cycle as three separate figures, which is the same ask this index put to the comparable independent platform.

The second is the ecosystem. The platform is sold modularly and designed to be adopted together, so a buyer should establish what the ambient layer costs alone, what it costs alongside the communication and alarm components, and whether pricing assumes the manufacturer's beds and devices are present. The honest question to ask is whether the ambient intelligence is being sold as a product or as a reason to standardise on one manufacturer's equipment.

The hospital inpatient environment, covered continuously rather than at intervals, with the platform positioned across the whole journey from transport through treatment to recovery.

Within that it is specialty agnostic, since a sensed room works the same way whatever the service line, and the virtual command centre model scales across units and potentially across facilities. Held at B because it is inpatient only, with nothing addressing ambulatory, home or clinic settings, and because reach depends on installing hardware room by room, which bounds how quickly coverage can widen.