This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in JMIR Aging, is properly cited. The complete bibliographic information, a link to the original publication on http://aging.jmir.org, as well as this copyright and license information must be included.
Wearables and artificial intelligence (AI)–powered digital health platforms that utilize machine learning algorithms can autonomously measure a senior’s change in activity and behavior and may be useful tools for proactive interventions that target modifiable risk factors.
The goal of this study was to analyze how a wearable device and AI-powered digital health platform could provide improved health outcomes for older adults in assisted living communities.
Data from 490 residents from six assisted living communities were analyzed retrospectively over 24 months. The intervention group (+CP) consisted of 3 communities that utilized CarePredict (n=256), and the control group (–CP) consisted of 3 communities (n=234) that did not utilize CarePredict. The following outcomes were measured and compared to baseline: hospitalization rate, fall rate, length of stay (LOS), and staff response time.
The residents of the +CP and –CP communities exhibit no statistical difference in age (
The AI-powered digital health platform provides the community staff with actionable information regarding each resident’s activities and behavior, which can be used to identify older adults that are at an increased risk for a health decline. Staff can use this data to intervene much earlier, protecting seniors from conditions that left untreated could result in hospitalization. In summary, the use of wearables and AI-powered digital health platform can contribute to improved health outcomes for seniors in assisted living communities. The accuracy of the system will be further validated in a larger trial.
Advances in public health and medical treatment over the past century have increased the average life expectancy in the United States by 30 years [
Older adults are disproportionally affected by chronic conditions, where 77% have at least two, and 65% have four or more chronic diseases [
Assisted living is a long-term care option that combines housing, personal assistance with activities of daily living, and supportive specialized services and therapy. In the US, approximately 812,000 older adults live in nearly 29,000 assisted living communities [
Many caregivers in assisted living communities rely solely on their observational powers to detect health changes in older adults in their care. As the number of residents requiring more assistance is increasing in these communities, the number of available caregivers is decreasing dramatically. Technologies can be utilized to augment and force-multiply human observation and provide quality care. Such solutions can provide continuous observation, detecting changes in activity and behavior patterns that may be indicative of a change in health status—information that cannot be provided by intermittent human observation. Artificial intelligence (AI) can be used to bridge the caregiver-senior ratio gap and augment occasional human observation with continuous machine observation and deep learning neural nets to predict when interventions are needed.
A growing body of evidence demonstrates that sensor-laden wearables utilizing AI, and in particular machine-learning algorithms, can detect an individual’s daily activity and behavior [
In this study, we tested whether the use of the CarePredict system could effectively improve the care provided in senior living communities. Specifically, we assessed the impact on hospitalization rate, fall rate, length of stay, and staff response time in each assisted living community.
The study was designed to assess facility-level and resident-level outcomes for communities that utilized CarePredict’s AI-powered digital health platform, wearable device, and real-time location system. Retrospective analysis of anonymized resident data was collected from six assisted living communities in three states over 24 months. A study flow chart is provided in
Workflow of data collection and analysis.
All de-identified data analyzed in this study were collected and reported by facility staff using each community’s online electronic healthcare software platform. The same software was used in all communities. CarePredict employees were provided access to the extracted anonymized data for scientific evaluation. No identifiable resident information has been or will be shared.
The CarePredict system consists of a wrist-worn wearable device, context beacons for room location, and a cloud-based AI-powered platform (
AI-powered digital health platform, wearable device, and room location system. A. Wearable device and sample representation of gesture recognition and activity detection. B. System architecture and overview of the data collection process. C. Summary of the product’s primary functions.
The real-time location or context beacons enhance the accuracy of the wearable’s gesture recognition engine by bringing in room type data and permitting accurate room-level location tracking in an indoor setting [
The system collects unique and rich data sets to train deep learning neural nets to surface crucial insights that correlate with an increased risk for a fall, UTI, or depression. A few correlates used in the system include the following: increased fall risk due to malnutrition, skipping meals, increased nightly get up count, reduced sleep duration, and decreased physical activity level; the increased probability for a UTI due to increased frequency or duration in bathroom visits, unusual toileting patterns, increased nighttime bathroom patterns, and reduced activity level; early signs of depression due to increased frequency of skipping meals, restless sleep, avoidance of bright lights and sunshine, and reduced physical activity levels. Further details on this system and established correlations are provided here [
Operators and staff both benefit from the use of an accurate real-time location system. First, staff can know the location of a resident who has pressed the button on their wearable to call for assistance, enabling the closest staff member to attend to a resident quickly. Second, historical insights allow the operators to assess previous shifts’ activities to improve staffing efficiencies. Such information may serve to facilitate improved response times, care coordination, and optimal workforce distribution. In addition, geofence alerts provide an added safety measure against wandering and elopement risks of residents with Alzheimer’s and dementia.
With a surge in acuity levels across senior living communities, providers need to have visibility into the amount of care required and provided to its residents. This solution allows caregivers to document at the point of care what services were rendered and a suite of reports that provide response time to alerts, time spent with residents on various direct care activities, and insights regarding quantity and the quality of care provided.
The wearable provides two-way voice communication that allows residents to communicate directly with the caregiving staff. Staff can prioritize alerts and respond appropriately. As a single communication platform for residents and staff, the wearable eliminates the need for multiple devices and provides tracking and reporting capabilities for staff efficiency.
The wearable is integrated with passive RFID technology so it can be used for keyless door entry, providing convenience and safety to residents and staff and assuring consistent adherence of use.
Facility staff at each community collected and reported the following data daily: occupancy, headcount, number of vacant units, unit move-ins and move-outs, staff service counts, duration, and type (such as dressing, bathing, grooming, transferring, and toileting), length of stay, fall incidents, emergency department and/or hospital admissions/discharges. Resident incident reports were utilized to document hospitalizations and fall incidents. The headcount, hospitalization, and fall incident numbers were used to compute both a hospitalization and fall rate. The hospitalization rate was defined as the number of hospitalization incidents per headcount in the facility, and the fall rate as the number of falls recorded per headcount per year. The average “baseline” rates for each community were measured from the first quarter of the study; the average “end of study” measurement was collected in the 8th quarter of the study. The average rate of change between these periods was computed between these two periods. Staff response times were automatically measured in this study using the CarePredict system. Residents trigger an alert and call for staff assistance by depressing the button on their wearable, and a staff member acknowledges the alert using the CarePredict software. We analyzed both the duration of time the staff required to acknowledge an alert and then to reach the resident. The residents’ length of stay in each community was also measured. Length of stay is defined as the number of months a resident resides in a given community. The average, geometric, and median length of stay were analyzed, and detailed descriptions are provided in the supplementary materials section (
Descriptive statistics, including means, standard deviations, and distributions, were provided for all study variables and compared across groups (+CP vs. –CP communities). Study variables were compared to baseline measurements for each group. A two-sample, two-tailed, t-test was applied for metric variables to test for significant differences between groups. A
Informed consent was obtained from the communities and participants included in the study.
The resident demographic data (age, sex) and facility staff service time were assessed. The average resident ages were 87.3 years (SD 1.2 years) for the +CP communities and 88.1 years (SD 1.6 years) for the –CP communities (
Resident age in the CarePredict and control assisted-living communities.
Age group | CarePredict (N=252), n (%) | Control (N=220), n (%) | |
Below 75 years | 12 (4.76) | 21 (9.55) | .64 |
75-80 years | 40 (15.87) | 26 (11.82) | .47 |
81-85 years | 74 (29.37) | 55 (25.00) | .44 |
86-90 years | 69 (27.38) | 60 (27.27) | .62 |
Over 90 years | 57 (22.62) | 58 (26.36) | .72 |
Average staff service time (hours) spent per headcount per month. There was no significant difference between groups (
|
Hours | ||
|
76.7 (20.9) | ||
|
1 | 81 | |
|
2 | 54 | |
|
3 | 95 | |
|
77.6 (2.6) | ||
|
1 | 59 | |
|
2 | 71 | |
|
3 | 103 |
The hospitalization and fall rates for the six assisted living communities are provided in
Outcomes: hospitalization and fall rates for six assisted living communities.
Community | CarePredict (+/-) | Hospital incidents per headcount, N=490 | Falls per headcount, N=490 | ||||
|
|
Baseline, n (%) | Change from baseline, (%) | End of study, n, % (SD) | Baseline, n, fall rate | Change from baseline | End of study, n, fall rate (SD) |
|
|
|
|
|
|
|
|
1 | – | 70 (42.0) | 18.7 | 74, 60.7 (11.2) | 70, 2.46 | 1.38 | 74, 3.84 (0.5) |
2 | – | 70 (37.2) | 57.1 | 78, 94.3 (12.7) | 70, 2.31 | 0.81 | 78, 3.12 (0.8) |
3 | – | 80 (38.1) | 27.8 | 82, 65.9 (3.1) | 80, 2.11 | 0.28 | 82, 2.39 (0.1) |
4 | + | 80 (45.1) | –18.2 | 84, 26.9 (6.7) | 80, 1.92 | –0.65 | 84, 1.27 (0.8) |
5 | + | 84 (57.2) | –19.4 | 84, 37.8 (18.9) | 84, 2.40 | –1.67 | 84, 0.73 (0.5) |
6 | + | 88 (44.0) | –7.3 | 88, 36.7 (10.1) | 88, 1.62 | –0.72 | 88, 0.90 (0.7) |
|
|
|
|
|
|
|
|
Mean (SD) | – | 39.1 (2.5) | 34.5 (7.24) | 73.6 (18.1) | 2.29 (0.18) | 0.82 (0.55) | 3.11 (0.75) |
Mean (SD) | + | 48.8 (7.3) | –15.0 (8.51) | 33.8 (6.0) | 1.98 (0.39) | –1.01 (0.57) | 0.97 (0.28) |
Delta |
|
9.7 | –49.5 | –39.8 | 0.16 | –1.83 | –2.14 |
|
.21 | .04 | .02 | .30 | .05 | .01 |
The median, geometric, and mean length of stay in the CarePredict and control communities are provided in
Length of stay in CarePredict and control communities.
Community | CarePredict | Control | Difference in CarePredict vs control (%) | |
Median length of stay (SD) | 214 (38) | 128 (8.7) | 67 | .03 |
Geometric mean length of stay (SD) | 178 (46) | 92 (8.6) | 93 | .04 |
Mean length of stay (SD) | 268 (42.4) | 192 (18) | 40 | .03 |
The average time to alert acknowledgment improved by 230 seconds (
Average acknowledgment and response times at baseline and the end of the study.
Response | Baseline, seconds, mean (SD) | End of study, seconds, mean (SD) | Improvement (%) | |
Acknowledge alert | 580 (42) | 349.5 (82) | 40 | .03 |
Reach resident | 763.5 (78) | 500 (35) | 37 | .02 |
In this pilot study, we assessed whether the use of a wearable device and AI-powered digital health platform could provide improved health outcomes for older adults in an assisted living community. We found that the communities with CarePredict (+CP) exhibited a 40% lower hospitalization rate, 69% lower fall rate, and 67% greater length of residence stay compared to control communities (–CP). Overall, the use of CarePredict technology in assisted living communities appears to contribute to improved outcomes and shows promise as an effective tool to provide a higher quality of care.
There are several possible explanations for these findings. First, since both the residents and staff wear the CarePredict device, it functions as both an effective communication platform and resident alert system allowing for the coordination of prompt care. The system also provides robust staff performance metrics, which can be used to encourage continually improving staff response times to residents who need aid. The alert system prevents minor situations from escalating to emergent situations requiring hospitalization. Residents desire prompt, attentive care in assisted living communities, and the CarePredict system helped contribute to the facility staff acknowledging alerts 40% faster and reach residents in response to those alerts in 37% less time [
Second, the system provides the community staff with detailed information regarding each resident’s activities and behavior. Changes in an adult’s activity and behavior are well-characterized to precede health declines; therefore, staff can use this information to quickly identify older adults that are at an increased probability for a health decline and intervene much earlier [
Fall rates also appear to be positively impacted by the use of the CarePredict solution. The data shows that +CP communities exhibited a 69% lower fall rate than the –CP communities. Fall rates are known to increase steadily with age [
By identifying older adults whose activity and behavior pattern indicates decreasing mobility, staff can take pre-emptive action to mitigate senior fall risk, UTIs, or other incidents that may have required hospitalization. Reducing hospital admissions also helps to maintain census, reduce resident turnover, and increase resident LOS in the community. The data shows that +CP communities exhibited a nearly 40% lower average hospitalization rate than the –CP communities (33.8% versus 73.6%). Several senior living outcomes studies by Zimmerman et al and Hedrick et al reported average annual hospitalization rates of 51% and 40%, respectively [
There are several limitations to this pilot study. First, the study was conducted at six assisted living communities with less than 500 total residents. This study needs to be replicated and results confirmed using a larger sample size of individuals. Second, the –CP communities did not use an alert response technology system in this study, and thus staff response times could not be collected and analyzed for the –CP communities. As a result, the impact that the CarePredict technology had on staff response times was only measured and analyzed for the +CP communities. We, therefore, could not compare the staff response times between the +CP and –CP communities, and rather only measured the response times for the +CP communities at baseline and end of the study. Third, staff in the +CP communities used the CarePredict technology system for multiple purposes: to acknowledge and respond to resident alerts, to communicate to residents and other staff members, and to autonomously collect resident activity and behavior data. Since Carepredict served multiple functions, it is difficult to attribute which of these system capabilities and data sets directly contributed to the improved outcomes.
To better understand the mechanisms by which these improvements were provided, in future studies, we plan to include a control group of communities that only utilize the CarePredict system for alerting and communication purposes. The added value provided by the predictive analytics feature will then be easier to assess and quantify directly. These results will also allow us to assess the impact that the proactive, actionable data generated by the CarePredict system will have on identifying and preventing high-risk residents from being hospitalized. Finally, although the six communities assessed in this study had comparable resident demographics (age, gender) and staff service hours per resident, other factors like residents’ hospital and fall history, and quality of care indicators may also have contributed to the differences observed in the measured outcomes.
The leading cause of residents moving out of assisted living communities is unplanned hospitalization [
Overview video on AI-powered digital health platform and wearable.
Definitions and methods for calculating length of stay.
artificial intelligence
community with CarePredict
community without CarePredict (control)
length of stay
radio-frequency identification
urinary tract infection
GJW, KD, JG, GZ, and SM are employees of CarePredict. HMF is an advisor to CarePredict.