SafelyYou

Computer vision detection is the mechanism. The hardware exists to run the model, and the clinical services exist to act on what it finds; without the detection model there is no product.

Structurally sound and unusually well supported. The system detects and alerts; staff respond and decide. Beyond the alert, the video review loop is what distinguishes it: clinical staff including RNs, OTs, and PTs review incidents against a stated 50 data point analysis in fall huddles to identify root cause and adjust care plans. The AI informs a human clinical process rather than replacing one, and the vendor supplies clinical expertise to run that process.

A performance figure is published, and it is the least informative figure available.

The claim is over 99 percent accuracy for fall detection, with the same figure quoted separately for care measurement. Accuracy is the wrong metric for a rare event and it is close to meaningless here. Falls are infrequent relative to the hours a room is monitored, so a detector that never alerted at all would also score above 99 percent accuracy. The numbers a buyer needs are sensitivity, meaning what share of real falls are caught, specificity and positive predictive value, meaning how often an alert is a real fall, and the false alarm rate per room per week. None is published.

This connects directly to the regulatory position. Because the category sits outside device regulation, nothing compels the decomposition, and the marketing figure survives unchallenged. The vendor is not unusual in this; it is worth naming because the vendor does publish a figure, which is more than its peer, and the form of the figure is the problem rather than its absence.

What is described elsewhere is more substantial. Event based capture triggered by a detected fall rather than continuous streaming, a stated 50 data point analysis applied in clinical review, and more than 100,000 fall incidents analysed over seven years. The scale of the operating dataset is credible.

Absent: training data description, any statement of the room conditions, lighting, camera placements and resident profiles the models were built against, and any model update or versioning policy.

Ask for sensitivity, specificity and false alarm rate, and for what the 99 percent figure is a percentage of.

One deliberate design decision bounds what exists to be handled, and nothing in the chain is named. Capture is event based, triggered by a detected fall, rather than continuous livestreaming, so the video that exists is a set of incident clips rather than a running record of a resident's bedroom, and no wearable is required. That materially limits what any party in the chain could hold, and it is a disclosed architectural choice rather than a retention policy.

It should still be read against the sensing method, which is a camera in a private bedroom and is inherently more invasive than the camera free skeletal approach a peer uses for the same task; a buyer should see that tradeoff rather than have it obscured.

On enumeration there is nothing: no model or model family, no provider, no hosting arrangement and no sub processor list was located, and no description of training data, room conditions, camera placements or resident profiles the models were built against. Retention of the incident clips is also unaddressed, which matters because those clips are footage of a fall, sometimes of a resident undressed or incontinent, and they persist by design rather than transiently. Ask for clip retention, who may view them, where they are stored, and for a sub processor list.

The strongest evidence provenance in the senior living segment of this index, because the foundational research was government funded rather than vendor commissioned. NIA Small Business Innovation Research grants supported a study reporting a 41 percent reduction in falls and a 69 percent reduction in emergency room visits, and the company received three subsequent NIA grants totaling more than $2.8 million.

Operator reported results corroborate the direction, including residents at one operator staying roughly six months longer. Scale is substantial at more than 800 communities over seven years, and the clinical team reports analyzing over 100,000 fall incidents. Graded on provenance and design; this index does not re verify the underlying study.

The design decisions are deliberate and disclosed: event based video capture triggered by a detected fall rather than continuous livestreaming, secure password protected review portal, and no wearable required. That is a meaningfully better posture than always on recording.

Graded B rather than A because the sensing method is a camera in a resident's bedroom, which is inherently more privacy invasive than the camera free skeletal imaging Inspiren uses for the same task, and the record should let a buyer see that tradeoff rather than obscure it. Operators should confirm consent processes and video retention terms, particularly for residents with dementia who cannot meaningfully consent.

Nothing published that sets out the posture, and a scoping question the vendor does not raise.

No business associate agreement, provider facing addendum, role statement, subcontractor flow down, breach notification timetable or review cadence was retrieved across two differently phrased searches.

The underlying question is whether the federal privacy regime reaches this data at all, and the answer depends on the buyer rather than on the sensitivity of the material. A covered entity is a health care provider transmitting health information electronically in connection with a standard transaction. Skilled nursing facilities almost always qualify. Assisted living and memory care communities often do not, because many bill through mechanisms that never touch the standard code sets, and a substantial share of residential care communities maintain no electronic health record at all. Where the operator is not a covered entity, this vendor is not a business associate, and the Health Insurance Portability and Accountability Act does not attach.

The consequence is sharper here than for the camera free peer, because what is held is video of residents in their bedrooms. Whether that footage carries federal protection turns on the billing configuration of the company that installed the system, a fact the resident cannot see and did not choose.

A second population sits outside the regime by construction. Every caregiver entering a monitored room is captured, and staff activity is sold as a product feature. Workforce monitoring data is not protected health information under any reading, so no federal health privacy obligation attaches to it regardless of the operator's status.

Ask whether a business associate agreement is signed by default, and what governs the footage when the answer is that nothing federal does.

One real enterprise control published, and no independent attestation behind it.

Single sign on is offered and named, which is a genuine control rather than a claim and matters for an application whose users are shift staff across many sites with high turnover. Access to the review portal is described as password protected.

Beyond that, two differently phrased searches returned no SOC 2 report of either type, no HITRUST certification, no ISO 27001, no trust centre, no penetration testing statement, no encryption specification and no subprocessor disclosure. Absence of a retrieved document is not proof that none exists and a buyer should ask, but this material is findable when vendors publish it.

The holding is the reason to press. This is video of cognitively impaired people inside their bedrooms, retained long enough for clinical staff to review incidents in fall huddles, accumulated across more than 800 communities and more than 100,000 analysed incidents. It is difficult to identify a more sensitive corpus anywhere in this index. Video of a resident falling, undressed or distressed, is materially different from a database record about them, and a breach would not be remediable by credit monitoring.

Retention is the specific unknown that should be settled in the contract rather than assumed: how long footage is held, who may view it, whether access is logged, whether a resident or family may request deletion, and what happens to accumulated video at termination.

Ask for the attestation, the retention schedule and the access log design together.

No clearance, probably none required, and the regime that does govern is the one nobody addresses.

The Food and Drug Administration's general wellness policy, revised January 2026, leaves low risk products of this kind either outside the device definition or under enforcement discretion. Published analysis of the fall detection category notes the direct consequence: manufacturers are not required to publish clinical performance data, so sensitivity, specificity and false alarm rates go unreported across the category.

What makes this record different from its peer is a second regime that applies specifically because the sensing method is a camera. More than twenty states have enacted electronic monitoring statutes governing recording devices in long term care and assisted living rooms, with requirements covering written resident consent, consent from any roommate with a right of revocation, signage at the room entrance, and written notice to the facility. Several states, including California and Florida, additionally require all party consent for audio recording of private conversations.

The critical nuance, and it cuts in the vendor's favour on the letter of the law: those statutes were written for resident and family initiated monitoring, cameras installed to watch the facility. They grant residents a right to observe. They were not drafted for operator deployed sensing installed in every room as part of the care model, which points the other way. So the detailed consent architecture that legislatures built for cameras in these rooms may not formally bind this deployment at all.

That is a gap, not a permission. The consent, signage and roommate provisions in those statutes are exactly the protections a resident would want here, and their non application is an artefact of drafting rather than a judgement that they are unnecessary.

Nothing published states the company's regulatory position or offers operators consent and notice tooling. Graded C on that silence.

No subgroup data, in an operating environment that concentrates the known weaknesses of computer vision.

No performance breakdown by any resident characteristic was retrieved, no bias testing statement, no governance framework, no model update policy and no third party audit.

The conditions are the concern. Falls concentrate at night, so the operating condition that matters most is a darkened bedroom, which is where camera based detection is hardest. Computer vision performance variation across skin tone is documented and well established, and low light and infrared conditions do not remove that variation. The index already records this pattern for pose and vision systems elsewhere.

The population compounds it. Memory care residents commonly use walkers and wheelchairs, have atypical or flexed posture, and move in ways that a model trained on ordinary ambulation may classify poorly. Bedding, furniture and mobility equipment occlude the body. Those are the residents at highest fall risk, so any performance gap lands on the people the system exists to protect.

There is a partial counterweight worth stating. The foundational research was supported by National Institute on Aging grants, which means the underlying science passed federal peer review rather than internal review alone. That is better provenance than most of this category and it is credited on the evidence axis. It does not substitute for published subgroup performance, which is a different question.

Ask for detection performance by ambient light level, by skin tone and by mobility aid use.

A figure is published and it is the least informative figure available, while a genuine verification mechanism sits alongside it. Take the mechanism first, because it is the reason this is not lower. Capture is event based, triggered by a detected fall rather than streaming continuously, and the resulting clip is reviewable through a protected portal.

That means an alert arrives with the evidence for it attached, so a reviewer can establish whether a real fall occurred rather than trusting the classifier, and a disputed alert can be settled by looking. Almost nothing else in this lane offers contestability at that level. Now the figure.

Better than 99 percent accuracy is claimed for fall detection, and accuracy is the wrong statistic for a rare event to the point of being close to meaningless: falls are infrequent relative to the hours a room is monitored, so a detector that never alerted at all would also score above 99 percent.

The numbers a buyer needs are sensitivity, meaning what share of real falls are caught, positive predictive value, meaning how often an alert is a real fall, and the false alarm rate per room per week. None is published, and because this category sits outside device regulation nothing compels the decomposition. Ask what the 99 percent is a percentage of, and for sensitivity and false alarm rate separately.

The best integration position in this small lane, and thinner than it looks at first reading.

Three senior living records of account are named: Yardi, ALIS and PointClickCare. Yardi's own materials independently list this vendor among its integration partners, so the claim is corroborated outside the vendor's site rather than only asserted. Single sign on is supported, which matters in an environment of shift staff and high turnover.

Alerts route through a dedicated integration with a senior living notification platform, and the stated purpose of that integration is to consolidate alerts from multiple systems and reduce alert fatigue. That is the correct problem to be solving. Continuous monitoring in a staffed building fails through desensitisation long before it fails technically, and addressing alert delivery as an integration problem rather than a product problem is a mature choice.

The qualifications matter. The published position includes more integrations coming soon, which is a roadmap rather than a capability and should not be counted. No application programming interface, interface specification, data export path or standards based interoperability is described, so these are bilateral partnerships and portability depends on the vendor. And as with the peer record, what is not evidenced is whether detected events and the clinical findings from incident review are written back into the resident record, or whether they remain in the vendor's own portal while the record stays silent.

Ask what field level data flows into the record of account, in which direction, and whether an operator can extract its own history on exit.

Hardware in every room, a cloud review portal, and no published answer on where the video lives.

The physical estate is substantial and it is the deployment. Wall mounted devices in resident rooms, plus the sensor, pendant and button hardware family, plus a next generation emergency call layer. That is capital equipment across the community, with installation per room, a refresh cycle and a switching cost that scales with the number of units. It also means an operator changing vendors is changing hardware, not a subscription.

On residency, nothing was retrieved. No hosting provider, no region, no tenancy model, no subprocessor list, no statement of whether detection runs on the device or in the cloud, and no retention period for captured video.

The on device question is the one that determines how much the event based design is actually worth. The architecture's central privacy claim is that it captures a short clip when a fall is detected rather than streaming continuously. If detection runs locally and only triggered clips leave the room, that claim is structural and strong. If a continuous stream leaves the room and clipping happens downstream, the same claim describes what is stored rather than what is transmitted. Those are materially different, and the published material does not distinguish them.

Retention deserves the same scrutiny. A clip reviewed in a fall huddle days later is a clip that persisted, so the practical retention window is not momentary and nothing states what it is.

Ask where inference runs, what leaves the room, how long clips persist and where they are stored.

No pricing published, and a headline return figure whose study design does not support the way it is used.

Nothing is disclosed on price: no rate card, no basis, no separation of hardware, software and clinical consultation, no term, no per room or per community structure, and no device refresh terms across an estate that spans hundreds of communities.

The return is framed substantially as occupancy economics rather than cost avoidance. The most prominent figure is that residents at one operator using the service stayed roughly six months longer than those not using it. In senior living, length of stay is revenue, so this is presented to the buyer as a top line argument.

The comparison is the problem, and it is stated in the vendor's own phrasing. The contrast is between residents who opted into the service and residents who did not. Those groups are not randomised and there is no reason to expect them to be alike. Families who opt into monitoring may be more engaged, more financially able to continue paying, or making different decisions about when to move a relative to a higher level of care. Any of those would produce a length of stay difference on its own. The figure may still reflect real benefit, and nothing published allows a buyer to separate the two.

That is worth naming precisely because the underlying clinical evidence here is strong and federally funded. A well evidenced product does not need a selection biased comparison in its marketing, and using one invites doubt about figures that would survive scrutiny.

The supporting cost figures come from the company's own state of falls research, so treat them as a vendor publication rather than an independent benchmark.

Ask for pricing basis first, and for whether the length of stay comparison has been repeated with any adjustment for who opts in.

Tightly and deliberately bounded: senior living communities with a stated concentration in memory care, where fall risk is highest and residents are least able to self report. Scope is stated as a focus rather than a limitation, and the product design follows from it.