hellocare.ai

The safety monitoring models are real inference work: fall prevention, fall detection, and pressure injury prevention running continuously across many rooms is exactly what a human sitter cannot scale to. Held back from A because a substantial share of the platform is video infrastructure, communication tooling, digital signage, and carts, which function without AI. The company's own naming, AI Assisted Virtual Nursing, is accurate: the AI assists a nurse who remains the actor.

The staffing model is the oversight model here, and it is stated consistently. AI detects and alerts; a virtual nurse or sitter observes and decides; bedside staff intervene. The framing is explicitly capacity extension, letting a smaller team monitor more rooms simultaneously and intervene earlier, rather than autonomous action. For inpatient safety monitoring that is the correct ceiling, and it matches the design of SafelyYou, Inspiren, and Sensi in the senior care lane.

The model functions are named individually, which is better than a general claim to use AI, and nothing beneath them is described.

What is public: distinct capabilities for fall prevention, fall detection, pressure injury prevention and patient mobility tracking, plus ambient documentation, with the statement that the algorithms run on the edge of the devices. Naming four separate computer vision functions rather than referring to monitoring generally lets a buyer reason about what is being claimed, and stating that inference happens on the device is a real architectural disclosure.

What is absent is every particular. No model or method is named, no architecture described, no training data characterised, no version or update practice documented, and no evaluation published for any of the four functions. For ambient documentation, nothing states whether the underlying language model is the company's own or licensed, or where it runs, which matters because that capability is the one most likely to depend on a third party service and therefore the one where the edge processing claim is least likely to hold.

The distinction between fall prevention and fall detection is also undefined. Detection identifies an event that has occurred. Prevention implies predicting one before it happens, which is a materially harder task with a different error profile, and the company does not explain what the prevention model actually does or on what signal it acts.

Ask what each of the four models is, what runs on the device versus in the platform, what the prevention model predicts and how far ahead, and how model updates reach an installed fleet.

One architectural claim answers part of this axis and probably does not cover the part that most needs it. The company states that its algorithms run on the edge of the devices, which for the vision functions means inference happens in the room rather than in a vendor estate, and that is a real structural bound of the kind this index credits elsewhere. Four capabilities are named individually rather than folded into monitoring generally, which lets a buyer see what is being claimed.

The gap is specific and a buyer should press on it. The platform also performs ambient documentation, and nothing states whether the underlying language model is the company's own or licensed, or where it runs. That capability is the one most likely to depend on a third party service, and therefore the one where an edge processing claim is least likely to hold, so an architectural statement made about the product as a whole may not describe the component carrying the most content.

Nothing else is enumerated: no model or method for any of the four vision functions, no training data characterisation, no hosting arrangement for the platform tier and no sub processor list. Ask what runs on the device against what runs in the platform, capability by capability, whose language model produces the documentation, and how model updates reach an installed fleet.

Adoption evidence is strong and unusually well dated: more than 70 US health systems reported in February 2026 rising to more than 100 by June 2026, with three named enterprise deployments in that window including Ardent Health across more than 2,000 patient rooms, MultiCare selecting it as enterprise virtual care platform after starting with telehealth, and One Brooklyn Health.

MultiCare expanding from a narrow telehealth use to enterprise standard is the stronger signal, since it reflects a customer deepening commitment. Held back from A because no outcome data was retrieved: for fall prevention and pressure injury prevention specifically, rate reduction against baseline is the measure that matters and none was published.

This is continuous video monitoring of patients in hospital rooms, including during care activities and personal moments, which is the most privacy sensitive sensing configuration in this index. No disclosure was retrieved covering retention, whether monitoring is continuous or event triggered, patient or family consent handling, or how footage is governed and who may review it. The senior care vendors in this index treat sensing choice as a privacy decision and say so; that discussion is absent here despite a more sensitive setting.

No statement of business associate status, agreement availability, scope or subprocessors was located. Customers are health systems, which are covered entities, so business associate arrangements certainly exist contractually. Nothing is published.

Two issues here sit outside the health privacy rule entirely and are the ones a hospital's counsel should reach first.

The first is audio. The hardware includes microphones in patient rooms, and the platform provides ambient documentation, which means listening. Recording or intercepting a conversation is governed by state wiretap and eavesdropping law, not by the health privacy rule, and a substantial number of states require the consent of all parties rather than one. A patient room contains conversations between patients and families, and between clinicians, that no one has framed as a recorded interaction. Establish what is captured, whether it is recorded or only processed transiently, who is deemed to consent, and how the position varies across the states a health system operates in.

The second is the sitter substitution. Virtual sitting replaces a human assigned to watch a patient continuously, and patients placed on continuous observation are frequently those at risk of self harm, or experiencing delirium, or subject to a legal hold. Continuous video of a patient in that condition is among the most sensitive material a hospital generates, and some of those patients cannot meaningfully consent at the time.

Ask which entity signs, what retention applies to video and audio, who may view a live feed and whether that is logged, and what governs paediatric rooms specifically.

No SOC 2, HITRUST, ISO 27001 or equivalent attestation was located and there is no trust centre. The strongest security language found is marketplace copy describing added layers of data security and privacy ensuring compliance with stringent regulations, which names nothing and is verified by no one.

The reason this weighs more heavily here than the grade alone conveys is that the failure mode is different in kind from the one this axis usually contemplates. Most vendors in this index hold records. A compromise means data is exfiltrated, which is serious and is discovered after the fact. This vendor operates cameras and microphones mounted in occupied patient rooms across a reported hundred or more health systems, including one enterprise deployment spanning more than two thousand rooms. A compromise here means live audio and video access to hospital rooms containing undressed, sedated, distressed and paediatric patients, in real time. That is a surveillance capability rather than a data set, and no privacy regulation restores what it exposes.

The platform's own design partly acknowledges this: the company states its algorithms run on the edge of the devices, which limits what needs to travel. But the same devices support live two way audio and video calls with clinicians and family members, so streaming from patient rooms is a normal operating mode rather than an exception.

Ask what independent security examination has been performed and against what, how camera and microphone access is authenticated and logged, whether any remote party can activate a device without an indication visible in the room, how the fleet is patched, and what the incident process is if a device is compromised.

No clearance is claimed, none was located, and no regulatory rationale is published. That last omission is what holds this at C, because the boundary here is closer than the product's framing suggests.

The reasoning that would place this outside device regulation is available and is probably right. The platform observes patients and alerts staff. It does not diagnose, does not direct treatment, and a nurse decides what to do. Software that surfaces information for a clinician who acts on their own judgement generally sits outside device regulation, and video based fall alerting has not typically been regulated as a device.

What complicates it is what the product replaces. Continuous observation is ordered for patients at risk of falling, of self harm, or experiencing delirium, and a human sitter is a clinical intervention documented in the record. Substituting software for that intervention means the software is performing a safety function that a care plan relies on, and pressure injury prevention monitoring likewise supports a nursing intervention with mandated documentation and defined outcome measures. A missed detection in either case is a patient harm event, not a workflow inconvenience.

The company has not published where it considers the line to fall or why. For a buyer, that matters less as a compliance question than as a diligence one: a vendor that has reasoned through the boundary can explain it, and one that has not has usually not examined its own failure modes either.

Ask for the regulatory position in writing, what the intended use statement says, whether any component has been assessed against device criteria, and how the vendor characterises responsibility when an alert does not fire.

No governance framework, model documentation or performance disclosure was located, and this vendor publishes one figure that makes the absence unusually consequential.

The company states that its virtual sitting service provides continuous observation for up to thirty two rooms on a single remote clinician's monitor. A traditional sitter watches one patient. That ratio is the product's economic argument and it is only safe if the models reliably surface the rooms that need attention, because a person monitoring thirty two live tiles cannot independently catch what the software does not flag. A published supervision ratio is therefore a claim about model sensitivity, whether or not it is presented as one, and no sensitivity figure is published for fall detection, fall prevention, mobility tracking or pressure injury prevention.

The fairness question here is specific and this index has met its analogue before. These are computer vision models inferring body position, movement and posture from video in rooms with varying light. Optical measurement in health care has a documented history of performing unevenly across skin tones, and the same physical mechanism that degrades optical signal acquisition on darker skin applies to segmenting a body from a background. Nothing published reports detection performance by patient skin tone, body habitus, mobility aid use, or room lighting condition. A monitoring system that detects falls less reliably for some patients concentrates harm precisely where it is least visible, because the missed event presents as a patient who fell without an alert rather than as a model error.

The company also confirms the platform is used in paediatrics, which raises the same questions with a different evidence base.

Ask for sensitivity and false negative rates per model, and for performance broken down by patient appearance and room conditions.

Two passes located no accuracy figure, no false alarm rate, no validation methodology and no warranty, indemnity or remediation commitment for any of the four detection capabilities. One published distinction is undefined and it matters more than a missing percentage, because the two things named have different error profiles and different consequences. Fall detection identifies an event that has occurred.

Fall prevention implies predicting one before it happens, which is a materially harder task, and a prediction model carries a false positive burden that lands on staff as repeated alerts and a false negative that is invisible by construction. Nothing published explains what the prevention model actually predicts, on what signal, or how far ahead, so a buyer cannot tell whether they are purchasing a detector, a predictor, or a detector marketed as a predictor.

The recording surface adds a set of affected people with no route at all, and this is the fourth product in the index where that arises. Continuous video of patients in hospital rooms captures people who are confused, sedated, undressed or dying, together with family and staff entering the room, and nothing published states whether monitoring is continuous or event triggered, what patients are told, whether they can decline, how long footage persists or who may review it. Ask for the false alarm and miss rates per capability, and for the consent and retention position in writing.

Genuinely integrated into hospital workflow rather than sitting beside it, which is what this axis rewards.

The company describes clinicians launching virtual care sessions directly from the electronic health record into the patient room, which is the right direction of integration: the nurse stays in the system they already work in rather than opening a separate console. Ambient documentation writes clinical notes, which is a write path into the record rather than a read. Digital whiteboards and room signage display patient specific information, which requires current data from the record to be useful and safe. The platform is also described as integrating with medical devices and third party applications, and the virtual nursing module supports admission and discharge workflows, which are among the most record intensive moments in an inpatient stay.

That combination, reading from the record to drive room displays, writing documentation back, and launching from within the record, is deeper than most vendors in this index achieve, and it is the reason a hundred or more health systems can adopt it without building a parallel process.

Held at B rather than A because no interface documentation, named integration, supported standard or certification was located, and because a claim to integrate with an electronic health record covers a wide range of depth. Ask which record platforms are integrated in production, what standards are used, what the ambient documentation writes and where it lands, and how a digital whiteboard confirms it is displaying current information for the patient actually in that room.

One genuine architectural claim, and nothing else published.

The claim is worth crediting. The company states its algorithms run on the edge of the devices themselves, and presents this as protecting patient privacy while tracking behaviour, monitoring for falls and alerting staff. Processing video where it is captured rather than streaming it to a server is the right design for this problem and it is a meaningfully stronger position than centralised analysis. If the model runs in the room and only an alert leaves, the exposure is transformed.

What is not published is everything needed to know how far that holds. The same devices deliver live two way audio and video between the room, remote clinicians and family members, so video plainly leaves the room in normal operation. Whether any video is recorded or retained, and where, is unstated. So is the hosting arrangement for the platform behind the devices, the region, the tenancy separation between health systems, the retention schedule, and the subprocessor list.

The hardware dimension adds a question most vendors in this index do not raise. This is a fleet of network connected cameras and microphones physically installed in patient rooms, in one case across more than two thousand rooms at a single customer. Who owns them, who administers them, how firmware is updated, and what happens to them and their stored contents at the end of a contract are deployment questions with no software equivalent.

Ask what leaves the device and when, what is recorded and retained, where the platform runs, and how device fleet administration and patching are controlled.

No public pricing. Contact the vendor. Enterprise agreements sold by module, with named deployments describing specific module combinations, so scope is negotiable. Note that hardware is included in the offering, with smart carts and in wall fixed devices, meaning capital cost sits alongside subscription and a per room deployment across thousands of rooms carries a materially different cost structure from software alone.

Clearly bounded to the inpatient environment and its adjacencies, spanning virtual nursing, virtual sitting and safety monitoring, patient engagement, ambient documentation, hospital at home, and remote monitoring after discharge. Deployment is described across diverse inpatient settings rather than a single unit type, and the company is expanding internationally.