CarePredict

The models drive the predictive module, inside a system where much of the deployed value is not model driven. Classifying activities of daily living from wrist motion (eating, bathing, toilet use, grooming, sleeping) and alerting on deviation from a personal baseline learned over about seven days is real inference, and the privacy policy (updated 2 December 2025) describes training AI models on de-identified data.

Alongside it, resident call with two way voice, fall detection, wander management, real time location from beacons and keyless door entry are safety and access functions that keep their value with the models switched off, and integrators present the product as an emergency call platform. The home page sells care plan optimization, resident safety and staff tools beside predictive health analytics.

The system alerts and a person decides, and the calibration that decides whether that works is missing. Insights and alerts reach care staff through push notification, dashboards, scheduled email reports and on demand reports, with urgent conditions such as a resident entering a restricted area or staying in a bathroom beyond an expected time routed immediately and slower trends in daily summaries.

The company publishes no alert thresholds, false positive rates, fall detection sensitivity or alert volume per resident per day, and nothing describes what happens after a missed alert. Alert fatigue is the established failure mode of monitoring in senior living: understaffed communities receiving more alerts than they can act on learn to dismiss them.

Sensing, method and training data are described; versioning and performance are not. The company names its sensing at component level (a wrist motion sensor for gesture recognition, heart rate and pulse oximetry, ultraviolet exposure, and beacons for room level location), describes deep learning classification of activities of daily living with a personal baseline learned over about seven days, and characterizes its training corpus as millions of activity data points gathered over about three years; the privacy policy states that training data excludes identifying information.

The baseline period is informative: a new resident produces no useful signal for about a week. Model versions, update practice, architecture beyond deep learning and any accuracy figure all go unpublished; the 3.7 day infection lead time says how early a correct detection arrives, not how often detection is correct.

The models are the company's own and the service providers are named. The privacy policy (updated 2 December 2025) describes the company training its own AI models on de-identified data, and names its providers: Amazon Web Services for storage and service email, Atlassian and Jira, Monday.com, QuickBooks, Firebase, Zendesk, HubSpot, Zoho, ClickHouse and shipping carriers. The Tempo wearable is built by NEXA Mobility, named publicly as the hardware partner. The list does not say which providers receive resident data rather than business data, and no terms governing data once it arrives are published.

A peer reviewed pilot and named communities with outcome figures, none with a method on the company's pages. A pilot study of the platform and wearable in assisted living communities appeared in JMIR Aging in 2020, with the company's founder among the authors, which makes it published research short of independent validation, and six years old against a product that has changed since.

The home page attributes outcomes to named communities: 60 percent improved urinary tract infection recovery times at Bickford, 39 percent lower hospitalizations at Lifewell Senior Living, 37 percent higher staff efficiency at Paradise Pointe, more than $320,000 of unbilled care captured at Edgewood Healthcare, and a 67 percent increase in length of stay attributed to a National Institutes of Health study, each without sample, period or comparison. The detection lead times (a urinary tract infection 3.7 days before clinical diagnosis, depression about two weeks ahead) carry no accuracy measure.

Training use and de-identification are stated as a commitment, and no retention schedule is. The privacy policy (CarePredict, Inc., updated 2 December 2025) covers the wearables, mobile and web applications used by care homes, home care agencies and consumers; it lists the data collected (motion state, steps, location, heart rate, and ambient humidity, pressure, temperature and light), states that data used to train AI models will not include identifying information, and allows sharing of anonymized data with third parties; the terms grant the company a perpetual license to anonymized user data.

Retention after termination is a commercially reasonable time for backup and archival, with no period. The dataset is intimate and continuous (where a resident is, when they bathe, eat and use the toilet), and nothing addresses consent for residents who lack capacity beyond registration by the customer or a guardian. Safety engineering, such as how a false or missed fall alert is handled, is not published.

A service level privacy document is published and no HIPAA status is stated. The privacy policy (updated 2 December 2025) covers the wearables and the applications used by care homes, home care agencies and consumers, describes the company as a processor acting for its customers, and names its service providers, but does not mention HIPAA, business associate status or protected health information; the terms of service are silent on it too. Whether HIPAA applies depends on the customer: many assisted living operators are not covered entities, while home care agencies that bill insurers and any clinical partner receiving data are.

No security attestation or outside check is published. The privacy policy (updated 2 December 2025) says the company takes reasonable steps to protect information but cannot ensure or warrant its security, and names no SOC 2 report, HITRUST or ISO/IEC 27001 certification; there is no security page, trust center, penetration testing statement or vulnerability disclosure policy.

The holding is a continuous record of where each resident is, room by room, and when they eat, bathe and use the toilet, with keyless door entry tied to the same wearable, so compromise reaches physical access as well as data.

No device claim is made, and two disease prediction claims sit against the terms' position. The terms of service state that the company is not a healthcare provider and cannot diagnose or treat conditions, and that the service is not for emergencies; the framing throughout is caregiver awareness, and emergency call, location, door access and activity summaries are not medical claims.

The company has also described identifying a urinary tract infection 3.7 days before clinical diagnosis and depression about two weeks ahead, and its home page credits earlier detection with faster infection recovery at a named community. Software intended to detect a specific disease is ordinarily where device regulation begins, and nothing published sets out why these fall outside it. Vital signs come from third party cleared devices uploading to the dashboard; that clearance belongs to those peripherals, not to the platform.

Nothing is published on how the models are governed or tested across residents. No fairness testing, subgroup performance, calibration data, drift monitoring, external audit or governance framework appears in the company's materials, and the privacy policy's commitment concerns de-identification of training data rather than how the models behave.

The risk follows from the method: activities are inferred from wrist motion, so residents using a walker or wheelchair, with hemiparesis after stroke, Parkinsonian tremor, contractures or a limb difference produce motion that differs systematically from a typical pattern, and in assisted living and memory care they are a large share of residents. Which wrist the device sits on changes what it sees. Learning each resident's own baseline avoids the worst population bias, but a stable baseline of a misclassified activity is still wrong.

A person acts on every alert and insight, and the terms put the risk of a wrong one on the customer. Alerts and trends reach care staff by push notification, dashboards and reports, and staff decide what to do, so an alert can be checked against the resident before care changes.

The terms of service (updated 7 August 2024) state that the service is not intended for emergencies and that a person in danger should call 911, that the company is not a healthcare provider and cannot diagnose, and disclaim all warranties of accuracy or error free operation, including responsibility for delayed response times. Liability is capped at the greater of two months' fees or $100, and the user indemnifies the company. Fall detection and wander management carry the sharpest exposure, since a miss is an immediate physical harm and no detection rate is published.

Integration categories are named and no product is. The home page lists integrations with electronic health records, safety and access systems, fall detection and engagement applications, and the company describes vital sign readings from third party devices uploading to its dashboard or a record system, and bringing door systems, resident management systems and nurse call vendors into one workflow, which is the harder consolidation problem in senior living.

No care management platform, medication administration record or nurse call product is named, and no interface standard is described, so whether staff work in one system or two is unstated. Behavioral insight helps only if it reaches whoever makes the care decision, and nothing says whether an alert or trend reaches a treating clinician outside the community.

A single hosted service on a named cloud, with location not committed. The privacy policy (updated 2 December 2025) names Amazon Web Services for data storage and states that data will be stored and processed in the United States and any other country. The physical deployment is described: a wrist wearable for each resident, peel and stick Context beacons for room level location, and dashboards and apps for staff and families. No region, residency commitment, tenancy model, segregation between communities, or backup and recovery position is published.

A consumer price is published, and the community business the product is mainly sold into is not priced. The company's storefront has sold the at home subscription at $69.99 per month with a 45 day trial and no annual contract, and a Canadian reseller lists a two person annual arrangement at $138 per month, so the running cost for a family is public before any conversation.

Senior living operators, whom the home page addresses first, get no price: a community needs a wearable for every resident, Context beacons throughout the building for room level location, and integration with door access and nurse call systems, and nothing states whether pricing is per resident, per bed or per building, whether beacons are bought or leased, or what installation costs.

Coverage across the senior care continuum is named, with named deployments behind it and validation behind part. The home page addresses senior living communities (assisted living, memory care, independent living, residential assisted living homes), home care agencies and people aging in place at home, and names communities including Mapleknoll, Edgewood Healthcare, Bickford, Westminster Communities, NuCare, Lifewell Senior Living and Paradise Pointe; a Canadian reseller extends it beyond the United States.

The conditions addressed (urinary tract infection, depression, malnutrition, fall risk, elopement) are ones where the first signal is a change in daily behavior, which is what the sensing measures. A cognitively intact independent resident and a memory care resident with advanced dementia differ in mobility, routine and capacity to consent, and nothing says whether models are validated separately across them.