Artisight

The models are the product. Computer vision interpreting what happens in a room, speech recognition, positioning and multimodal fusion across a sensor network are not a layer on top of something else here, they are the entire mechanism. Strip them out and what remains is cameras and microphones with nobody watching.

The hardware choice corroborates rather than dilutes this. Inference runs on graphics processing units installed on the premises, and the company took investment from the manufacturer of those processors. Committing to on site accelerated compute in hundreds of hospitals is an expensive decision that only makes sense if continuous model inference is the core workload.

The oversight architecture is unusually explicit for this index because it is the product rather than a safeguard bolted onto it. The system senses and interprets continuously, but the response routes to a remote human, whether a virtual nurse, a virtual observer watching several rooms, or a bedside team receiving an alert. A person remains the actor in nearly every workflow.

Held at B rather than A because the thresholds are unpublished. Nothing states the sensitivity setting for a fall alert, the false alarm rate a unit should expect, how many rooms one virtual observer can watch before attention degrades, or what happens when the network or the on site hardware fails during observation of a high risk patient. In a product replacing a human sitter, the answer to that last question is a patient safety matter rather than an operational one.

The technical approach is named at a level most vendors will not match: computer vision, deep learning, a multi sensor network, voice recognition, indoor positioning, vital sign monitoring and open integration standards, running on named accelerator hardware.

What is absent is anything a buyer could evaluate. No model architecture, no training data description, no accuracy figure for fall detection or any other detection task, no false alarm rate, and no validation methodology. Naming the technique is not the same as reporting how well it works, and for detection models the second is the only number that matters.

The architecture answers much of this axis structurally, which is the strongest form of answer available. Processing hardware sits on the client's own premises behind the customer's firewalls and physical security, and cloud components run on named providers only at the client's request, so in the default configuration content is not routed to a vendor estate at all and any chain that does exist is one the customer has opted into and can name.

That is the same class of protection this index credits for on premises deployment elsewhere, and for a product generating continuous video and audio it removes more risk than any policy could. The technical stack is named at a level most vendors will not match, including the accelerator hardware the models run on, alongside computer vision, a multi sensor network, voice recognition, indoor positioning and vital sign monitoring.

Encryption is stated by named protocol in transit and by strength at rest. Held below the top grade because the model layer itself is unnamed, with no model or provider identified for any detection task, and no sub processor list was located for the optional cloud path. Ask which providers the cloud components use when enabled, whether any third party model is invoked, and what leaves the premises in each configuration.

The outcome figures attributed to Northwestern Medicine are large and specific: an 89 percent reduction in falls, a 52 percent reduction in nursing overtime, a 76 percent reduction in nursing turnover. They are attached to a named academic health system rather than to an anonymous customer, which is better than most.

They are also unpublished, undated as to measurement period, and carry no baseline, denominator or control. A fall reduction of that magnitude would be a significant result in the patient safety literature. The company itself stated in 2024 that peer reviewed publication was anticipated, which is a candid position and the right intention, and no such publication was located in this pass. Deployment across 417 hospitals demonstrates commercial adoption, which the index does not accept as a substitute for evidence of benefit. Re verify whether the results have since been published.

Better documented than the marketing suggests, because the detail sits in the privacy policy rather than on a product page. Data and video transmissions are encrypted in transit with transport layer security 1.2 and at rest with 128 bit encryption, the processing hardware sits on client premises behind the customer's own firewalls and physical security, and cloud components run on named providers only at the client's request.

The gap is consent and retention, and it is sharper here than anywhere else in this index. This product records video and audio of patients in hospital beds, including patients who are confused, sedated, undressed or dying, and of every family member and staff member who enters the room. Nothing published states how long footage is retained, whether it is retained at all rather than processed transiently, what patients are told, whether they can decline, or whether recorded material is used to improve models. Those are the first questions to put in writing.

Compliance with United States health privacy law is asserted, and the capability list is described as compliant, but assertion is all that is available. No body certifies HIPAA compliance, so a claim of this kind reports intention rather than an audited finding.

No business associate agreement posture, notice of privacy practices or compliance documentation was located. The on premises deployment option does materially change the risk picture, since data processed behind the customer's own firewall on the customer's own site is a narrower exposure than a vendor cloud, and that is a real point in its favour rather than a formality.

The privacy policy states that systems are SOC 2 Type 2 compliant in all three deployment configurations, which is a stronger claim than most peers make because it covers on premises and hybrid installations rather than a cloud service alone. Encryption is specified concretely, with transport layer security 1.2 in transit and 128 bit encryption at rest.

Two things hold this below A. The claim is that the systems are compliant, made in a privacy policy, rather than that an audited report exists and can be requested, and those are different assertions. And 128 bit encryption at rest sits below the 256 bit standard most healthcare vendors now publish, which is not a defect but is a specific choice a security reviewer will ask about. No trust centre, penetration testing statement or HITRUST certification was located.

No clearance was located and the company does not claim one. Much of the platform sits outside device regulation as workflow, communication and observation software.

The boundary is worth naming, because parts of the capability list sit closer to it than the framing suggests. Vital sign monitoring and automated detection of a patient event in a hospital bed are functions that, described differently, would fall inside patient monitoring device territory. A buyer should establish which functions the company treats as regulated and which it does not, particularly where an alert substitutes for a human observer at the bedside.

Nothing published on subgroup performance, model monitoring or bias evaluation, and the exposure here is more concrete than in most records because the input is images of human bodies.

Computer vision and pose estimation accuracy is documented to vary with skin tone, body size and lighting, and detection of a fall or an attempted exit from a bed depends on both. The patients under continuous observation are disproportionately older, frailer, heavier or lighter than average, dark skinned in some populations, wearing casts or braces, connected to lines and monitors, or lying in positions no training set anticipated. A detection model that performs less well on some of those patients concentrates its misses in exactly the group placed under observation because they were judged highest risk. This is the second instance in this index of the pose estimation equity problem, after the Hinge Health record, and it is the more consequential one.

Two passes located no accuracy figure for fall detection or any other detection task, no false alarm rate, no validation methodology and no warranty, indemnity or remediation commitment. For detection models that omission is more serious than a missing accuracy percentage on a scribe, because the false alarm rate is the number that determines whether a system is usable at all: alarms that fire too often are silenced, worked around or ignored, and a detection product that has trained its users to disregard it has a worse safety profile than no product.

The complementary figure matters just as much and is equally absent, since a missed fall is invisible by construction, appearing only as a patient found on the floor with no record that anything should have fired. The recording surface creates a set of affected people who have no route of any kind, and this is the third product in the index where the affected party is not the patient.

Video and audio are captured in hospital rooms, so the recorded population includes patients who are confused, sedated, undressed or dying, and every family member and staff member who enters the room. Nothing published states what patients are told, whether they can decline, how long footage persists, or whether recorded material trains future models. Ask for the false alarm rate and the miss rate per detection task, and for the consent and retention position in writing.

Open integration standards are stated and ambient documentation of room activity implies a route into the record, but no named electronic health record integration, interface standard or write back mechanism was located.

The integration surface is also broader and stranger than a typical record, because the platform touches nurse call, real time location, physiological monitoring and operating room scheduling as well as clinical documentation. Each of those is a separate integration with a separate owner inside the hospital, and none of them is described in public material.

One of the better deployment positions in the index, and the reason is that the customer chooses. The platform runs on premises, in the cloud, or hybrid, at the client's election, with cloud components on named major providers only where the client asks for them. The processing hardware sits on the customer's own site behind their firewalls and physical security controls.

That gives a health system unusual control: inference on video of its own patients can stay inside its own building on its own network. Data residency effectively follows the customer's infrastructure rather than the vendor's. Held at B only because no retention schedule is published, and for continuously captured room video the retention question matters more than where the server sits.

Nothing is published. No price, no pricing mechanism, no unit of sale, no indication whether the model is per room, per bed, per facility or per module, and no separation of the hardware cost from the software subscription.

That last omission is the substantive one. This deployment requires sensors and accelerated computing hardware installed in every room it covers, so the capital component is real, potentially large, and entirely undisclosed. A buyer cannot form even an order of magnitude estimate from public material, and the capital structure adds a wrinkle worth naming: a number of client hospitals are also investors, so the reference customers and the shareholders overlap.

Broad within the hospital and absent outside it. Deployment spans inpatient rooms, clinics and operating rooms across 417 hospitals, and the functional coverage crosses nursing, patient safety, perioperative operations, clinical documentation and tele consultation.

The buyer is a health system rather than a department, which is consistent with the installation model, and the platform is specialty agnostic because it is sensing an environment rather than reasoning about a condition. It does not reach ambulatory practices without the physical estate to instrument, and nothing outside the four walls is addressed.