CarePredict

Genuine inference doing consequential work, inside a system where a large share of the deployed value is not model driven.

The intelligence is real and specific. Classifying activities of daily living from wrist motion is a hard signal processing and machine learning problem, and the outputs are behavioural categories rather than raw accelerometry: eating, bathing, toilet use, grooming, sleeping. Learning a personal baseline over roughly seven days and then alerting on deviation is a sound design for a population where the clinically meaningful event is change from an individual's own norm rather than crossing a population threshold.

What sits alongside it would retain most of its value with the models switched off. Resident call with two way voice, fall detection, wander management, real time location and keyless door entry are safety and access functions, and system integrators present the product as a next generation emergency call platform, which is how a good deal of it is actually bought. Location itself comes from beacon proximity rather than from a model, and the vital signs pathway is third party cleared devices uploading readings.

Graded B on the same reasoning applied to platform vendors elsewhere in this index. The predictive layer is the differentiator and the reason a buyer would choose this over a conventional nurse call system, but it is not the whole product, and a buyer should establish which half they are actually paying for.

A clean human in the loop design, with nothing published about the thing that decides whether it works in practice.

The model does not act. Insights and alerts reach care staff through push notification, web dashboards, scheduled email reports and on demand reports, and a person decides what to do. Alert channels are described as configurable, and urgent conditions such as a resident entering a restricted area or remaining in a bathroom beyond an expected period are routed immediately while slower trends surface in daily summaries. Separating urgent routing from trend reporting is the right architecture for this setting.

What is missing is calibration. No alerting threshold is published, no false positive rate, no false negative rate for fall detection, and no figure for alert volume per resident per day.

That last omission is the important one, because alert fatigue is the established failure mode of monitoring systems in senior living rather than a hypothetical risk. Understaffed communities receiving more alerts than they can action learn to dismiss them, and a system generating continuous behavioural signals across every resident in a building has an obvious path to that outcome. Nothing published describes tuning, suppression, or what the vendor advises a community to staff against.

The fall detection function carries a different oversight question, since a missed fall is an immediate physical harm and no detection rate is published.

Ask for typical alerts per resident per day at a reference community, the fall detection sensitivity, and what staffing the vendor assumes.

The approach is described more concretely than most, and no number describes how well it works.

What is disclosed is useful. The sensing stack is named at component level: a motion sensor supporting gesture recognition, combined heart rate and pulse oximetry, ultraviolet exposure sensing, and peel and stick beacons providing room level location. The learning design is stated plainly, with roughly seven days required to establish an individual's normal pattern before deviation becomes meaningful, and the training corpus is characterised as millions of activity data points accumulated over about three years. Stating the baseline period is a genuinely informative disclosure, because it tells a buyer how long a new resident produces no useful signal.

Performance is absent entirely. No accuracy figure for activity classification, no sensitivity or specificity for any prediction, no operating point, no calibration, no model card and no architecture description beyond references to deep learning.

The two published figures are lead times rather than performance. A urinary tract infection identified 3.7 days before diagnosis and depression around two weeks ahead describe how early a correct detection arrives, and say nothing about how often detection is correct, how often it fires wrongly, or in how many residents it was measured. A lead time without a detection rate is not an evaluable claim, and the decimal place implies a precision the surrounding evidence does not support.

Ask for activity classification accuracy, the detection rate and false positive rate behind each prediction, and the sample the 3.7 day figure came from.

One component of the chain is named openly, which is more than most vendors here manage, and the rest is undisclosed.

The named component is the hardware. NEXA Mobility is publicly identified as the partner that designed and built the Tempo wearable, described in that company's own account of the engagement. Hardware origin is a real supply chain fact with real consequences for firmware, component sourcing and long term device support, and most vendors selling a branded device do not disclose who makes it. The vitals pathway is similarly attributed, with readings described as coming from separate cleared third party devices rather than from CarePredict instruments.

Everything above the hardware is undisclosed. No cloud or infrastructure provider is named, no sub processor register was located, no model framework or component is described, and no position is published on whether resident data contributes to model development.

That last question carries weight given how the models were built. The training corpus is characterised as millions of activity data points accumulated over roughly three years, which is data generated by monitored residents, and nothing states the consent basis on which it was used or whether current deployments continue to feed it. Residents in memory care cannot meaningfully consent to their movement data training a commercial model, and the question is not addressed in any direction.

Ask for the sub processor register, the hosting provider, the consent basis of the training corpus, and whether current customer data trains models.

One peer reviewed pilot from 2020, and headline outcome claims that no located source supports.

The publication is real. A pilot intervention study of the platform and wearable in assisted living communities appeared in JMIR Aging in 2020. That is a genuine peer reviewed artefact and more than several vendors in this lane hold. Its limits are equally real: a pilot design, a single study, six years old against a product that has iterated since, and the company founder appears among the authors, so it is vendor participating research rather than independent evaluation.

Commercial adoption is evidenced by name rather than by number. Grand Retirement, Spring Creek Enterprise, SRI Management, Tradition Senior Living and Bickford Senior Living are identified across company and partner material, which is better than an unnamed customer count. No current figure for communities served or residents monitored was located from a primary source.

The outcome claims are the problem. Fall reduction of 69 percent and length of stay increase of 67 percent appear on distributor pages with no study, sample, comparison group, time period or citation. The detection claims carry the same weakness: a urinary tract infection identified 3.7 days before clinical diagnosis and depression around two weeks ahead are precise figures with no located methodology, no sample size and no accuracy measure beside them. A number given to one decimal place implies a study, and the study was not found.

One caution recorded deliberately. Several third party profiles of this company carry precise operational statistics traceable to no primary source and bearing the marks of machine generation. Those were excluded from this record and should be excluded from any future refresh.

Ask for the source behind the 3.7 day figure, and for a controlled fall reduction result at a named community.

The most intimate continuous dataset in this lane, with no published account of how it is handled.

A dedicated pass located no encryption statement, no key management, no retention schedule, no access control model, no audit logging, no deletion process and no data ownership position.

What is collected raises the stakes well above the usual. This is continuous, room level tracking of when a person eats, washes, uses the toilet, sleeps and moves, retained long enough to support retrospective contact tracing through the PinPoint capability. A record of how long someone spent in a bathroom, on which days, is not comparable to a lab result.

Consent is the sharpest question and it is unaddressed. A substantial share of the deployed population is in memory care, where a resident may be unable to give or withdraw meaningful consent, and the purchaser is an operator or an adult child rather than the person being monitored. Nothing published describes how consent is obtained, whether a resident can decline, whether the wearable can be removed without consequence, or what a resident is told.

Family visibility compounds it. The TouchPoint application shows relatives real time activity, so a person's daily bathroom and eating patterns become visible to family members, and nothing describes who authorises that or whether the monitored adult has any say in it.

One pre emptive note: a general privacy policy cannot move this grade. Only a published retention schedule, an access control description and a stated consent process for cognitively impaired residents will.

Ask what is retained and for how long, who can view it, and what the consent process is in memory care.

A dedicated pass located no health privacy position of any kind: no compliance statement, no business associate agreement template, no execution requirement and no description of which entity contracts with whom.

The structural point matters more than the missing badge. This product is sold into three settings with three different legal pictures, and nothing published distinguishes them. A senior living operator is frequently not a covered entity under United States health privacy law, since assisted living and independent living are largely residential rather than clinical services. A home care agency may or may not be. And the consumer at home product, bought by an adult child for a parent through a retail storefront, sits outside that framework altogether, which means the protections covering continuous behavioural data on that person are whatever the privacy policy grants rather than what statute requires.

That distinction is invisible to the buyer and consequential for the older adult. A family purchasing this expects health data protections that may simply not apply.

The data at issue makes the omission serious rather than technical. Toilet use, bathing, eating and room level location, recorded continuously, describe a person's body and private routine more intimately than most clinical records do.

Graded D because nothing at all was located on this axis after a dedicated search. Ask which deployments the company treats as covered, whether it signs agreements with operators, and what governs the consumer product.

Two dedicated passes located no security page, no external attestation, no trust center, no penetration testing statement, no vulnerability disclosure policy and no compliance documentation offered under agreement. No security credential of any kind was found, including the self asserted badges that at least appear on comparable vendors.

The exposure is broader than software alone, which is what distinguishes this record on this axis. The system includes keyless door entry by radio frequency identification driven from the resident wearable, so the same platform that holds behavioural data also governs physical access to residents' rooms. A compromise therefore has a physical dimension that a data breach ordinarily does not, in a building housing people who may be unable to respond to an intrusion.

The wearable and beacon layer adds a second surface. Body worn devices and installed radio beacons communicating continuously through a building are hardware attack surfaces, and nothing published describes device authentication, firmware update practice, or how a lost or removed wearable is deauthorised.

The monitored population raises the consequence of any exposure. A record of when a resident is asleep, out of their room, or in the bathroom is precisely the information that would be useful to someone intending harm to a vulnerable adult.

One pre emptive note: a privacy policy or a restated compliance claim cannot move this grade. Only an external attestation, or security documentation available under agreement, will.

Ask whether any external security audit has been performed, how devices authenticate, and how door access is protected if the platform is compromised.

Wellness positioning carrying two disease prediction claims, with no determination published and one attribution a reader can easily misread.

No clearance, grant or approval for any CarePredict product was located, and none is claimed. The framing throughout is caregiver awareness and preventive senior care, which is the recognised wellness route to remaining outside device regulation, and much of the product supports it: emergency call, location, door access and activity summaries are not medical claims.

Two claims sit awkwardly against that framing. Predicting a urinary tract infection ahead of clinical diagnosis is identifying a specific infection in a specific person before a clinician does, and predicting depression is identifying a psychiatric condition. Both are disease detection claims stated with numerical lead times, and neither is a general wellness statement in the way step counts or sleep summaries are. Software intended to detect or predict a specific disease is ordinarily where device regulation begins, and nothing published sets out why these fall outside it.

The attribution worth flagging concerns the vitals pathway. Company material states that measurements are collected from cleared devices and uploaded to the dashboard, which is accurate: the clearance belongs to the third party blood pressure, oximetry and glucose peripherals, not to CarePredict. Set in a paragraph about the platform, it reads to a skimming buyer as though the platform carries a clearance, and it does not.

Fall detection occupies a middle position, treated by some regulators as a safety feature and by others as a device function depending on the claim made.

Ask for the written device determination covering the infection and depression predictions, and confirmation of what the cleared devices actually cover.

A dedicated pass located no fairness testing, no subgroup performance, no calibration data, no drift monitoring, no external audit and no governance framework.

The risk here is mechanical rather than abstract, and it follows from how the system works. Activities of daily living are inferred from wrist motion through gesture recognition, which means the models encode assumptions about how a movement looks when a person performs it. Residents using a walker or wheelchair, with hemiparesis after a stroke, with Parkinsonian tremor, with contractures, or with a limb difference all generate motion signatures that differ systematically from a typical pattern. That population is not an edge case in assisted living and memory care; it is a large share of the residents.

Handedness alone raises a question no material addresses, since a wrist worn sensor on a non dominant arm sees different activity from one on a dominant arm.

The personal baseline design is genuinely mitigating and should be credited. Learning each individual's own pattern over roughly seven days and alerting on deviation avoids the worst form of population bias, because a resident is compared against themselves. It does not solve classification: if the model cannot recognise a wheelchair user's eating motion as eating in the first place, a stable personal baseline of a misclassification is still wrong.

Nothing published states detection accuracy for any subgroup, or whether mobility aids were represented in training.

Ask for activity classification accuracy among residents using mobility aids, and whether atypical movement patterns were represented in the training population.

Nothing allocates responsibility, and this product carries functions where a miss is an immediate physical event rather than a delayed one.

A dedicated pass located no service level agreement, no accuracy warranty, no performance guarantee, no indemnity, no uptime commitment and no remediation position. Consumer subscription terms were not reachable from the storefront listing.

The distinctive exposure is that this is partly an emergency system. Fall detection and resident call with two way voice are life safety functions, and integrators present the platform as an emergency call system. A missed fall, an alert that does not reach staff, or a device that has lost connectivity produces harm within minutes, and no detection rate, no uptime figure and no escalation commitment is published for any of it.

The predictive layer carries a quieter version of the same problem. A claim to identify infection days before diagnosis invites a community to rely on the system, and no false negative rate is published, so nobody can say what proportion of infections it misses. Reliance without a published miss rate is the condition under which a monitoring system quietly displaces vigilance it was meant to supplement.

The consumer channel sharpens it again. A family purchasing directly has no procurement function, no counsel and no leverage, and is buying a safety product on standard terms they cannot negotiate.

One pre emptive note: further customer names or outcome marketing cannot move this grade. Only a contractual commitment, or published detection and uptime performance for the emergency functions, will.

Ask what the community agreement commits to on fall detection and alert delivery, and what recourse a consumer subscriber has.

Integration categories are named, integration products are not.

What is stated has substance. Vital sign readings are described as uploading automatically to the dashboard or to the record system, which asserts an outbound clinical data path rather than a closed application. The company also describes integrating electronic door systems, resident management systems, nurse call vendors and other alert sources into a single unified workflow, and in this setting that is the harder and more valuable integration work: a community running separate nurse call, access control and care management systems has a genuine consolidation problem.

Not one product is named. The systems a senior living operator actually runs, whether care management platforms, electronic medication administration records or nurse call hardware, appear nowhere by name, and no interface standard is described. Whether a resident management integration is bidirectional, or whether care staff must work in two systems, is unstated.

One question is specific to this product and unaddressed. Behavioural insight is only clinically useful if it reaches whoever makes the care decision, and in assisted living that may be a community nurse rather than a physician. Whether an alert or a trend ever reaches a treating clinician outside the community, or stops at the operator dashboard, determines whether the earlier detection claim can translate into earlier treatment.

Ask which resident management and nurse call systems are integrated in production, through what standard, and whether findings reach a treating clinician.

A dedicated pass located no hosting provider, no region, no tenancy model, no residency commitment, no segregation description and no backup or recovery position.

What can be established concerns only the customer side of the boundary. The deployment has a substantial physical footprint: a wearable per resident, Context beacons installed through a building for room level location, and integration with door access hardware. That equipment sits in the community. Where the resulting data travels, and where it comes to rest, is unstated.

Two factors make the omission more consequential than a generic absence. Location history is retained long enough to support retrospective contact tracing, which means a persistent movement record exists somewhere for every resident, and nothing describes where or under whose control. And distribution through a Canadian reseller means residents outside the United States are generating that record, with cross border transfer and residency obligations that no published material acknowledges in either direction.

A related question follows from the architecture and is also unanswered: whether activity inference runs on the wearable, on premise, or in the cloud, which determines whether raw motion data leaves the building at all.

Graded D because nothing was located, not because what exists is thin. Ask where data is hosted and in which region, where inference runs, what location history is retained, and what governs Canadian deployments.

A published price with a published unit, which is rare enough in this index to lift the grade, covering the smaller half of the business.

What is open is genuinely open. The company's own storefront sells the consumer at home subscription at 69.99 dollars per month, states a 45 day trial, and states no annual contract, so a family can establish total cost before speaking to anyone. A Canadian reseller publishes a two person annual arrangement at 138 dollars per month. Earlier reporting recorded roughly 169 dollars for the wearable and about 30 dollars monthly, which indicates the historical shape of hardware plus subscription. Return arguments are also published rather than withheld, including a claim that a single prevented fall covers a year of system cost.

What is dark is the larger half. Senior living operator pricing appears nowhere. That omission carries more weight here than for a software only vendor, because this deployment has physical components: a wearable for every resident, Context beacons installed throughout a building to achieve room level location, and integration with door access and nurse call systems. Hardware quantity, installation labour and building survey are all real costs and none is indicated even as a range.

Nothing states whether the community price is per resident, per bed, per building or per unit of hardware, whether beacons are purchased or leased, or what happens to the hardware at contract end.

The published return claims also lack the inputs to test them, since fall reduction and length of stay figures carry no method.

Ask for the community pricing basis, the beacon and installation cost for a representative building, and the contract term.

Coverage across the senior care continuum is real and evidenced by named deployments, with depth outside that continuum absent.

The breadth is genuine. Assisted living, memory care and independent living are addressed with different emphases, home care agencies are a stated buyer, and families purchase directly for a relative aging in place, which is a materially different context from an operated community. Named communities span several operators, and a Canadian reseller channel indicates deployment beyond the United States.

Condition coverage within the population is coherent rather than scattered. Urinary tract infection, depression, malnutrition, fall risk and elopement are all conditions where the earliest signal is a change in ordinary daily behaviour, which is exactly what the sensing layer measures. Choosing conditions that match the signal is a sign the scope was reasoned rather than assembled.

Independent living and memory care being served by the same product raises a question that is not addressed. A cognitively intact independent resident and a memory care resident with advanced dementia differ completely in mobility, routine regularity and capacity to consent, and nothing describes whether models are tuned differently or validated separately across those groups.

Skilled nursing is not addressed, and neither is any acute or clinical setting, which is consistent with the product rather than a gap in it.

Ask whether models are validated separately for memory care residents, and how the system handles residents with limited or atypical mobility.