Using Embedded Sensors in Independent Living to Predict Gait Changes and Falls

Abstract

This study explored using big data, totaling 66 terabytes over 10 years, captured from sensor systems installed in independent living apartments to predict falls from pre-fall changes in residents' Kinect-recorded gait parameters. Over a period of 3 to 48 months, we analyzed gait parameters continuously collected for residents who actually fell (n = 13) and those who did not fall (n = 10). We analyzed associations between participants' fall events (n = 69) and pre-fall changes in in-home gait speed and stride length (n = 2,070). Preliminary results indicate that a cumulative change in speed over time is associated with the probability of a fall (p < .0001). The odds of a resident falling within 3 weeks after a cumulative change of 2.54 cm/s is 4.22 times the odds of a resident falling within 3 weeks after no change in in-home gait speed. Results demonstrate using sensors to measure in-home gait parameters associated with the occurrence of future falls.

Full text

Using Embedded Sensors in Independent Living to Predict Gait
Changes and Falls
Lorraine J. Phillips1, Chelsea B. DeRoche1, Marilyn Rantz1, Gregory L. Alexander1, Marjorie
Skubic1, Laurel Despins1, Carmen Abbott1, Bradford H. Harris1, Colleen Galambos1, and
Richelle J. Koopman1
1University of Missouri, Columbia, USA
Abstract
This study explored using big data, totaling 66 terabytes over 10 years, captured from sensor
systems installed in independent living apartments to predict falls from pre-fall changes in
residents’ Kinect-recorded gait parameters. Over a period of 3 to 48 months, we analyzed gait
parameters continuously collected for residents who actually fell (
n
= 13) and those who did not
fall (
n
= 10). We analyzed associations between participants’ fall events (
n
= 69) and pre-fall
changes in in-home gait speed and stride length (
n
= 2,070). Preliminary results indicate that a
cumulative change in speed over time is associated with the probability of a fall (
p
< .0001). The
odds of a resident falling within 3 weeks after a cumulative change of 2.54 cm/s is 4.22 times the
odds of a resident falling within 3 weeks after no change in in-home gait speed. Results
demonstrate using sensors to measure in-home gait parameters associated with the occurrence of
future falls.
Keywords
sensors; falls; gait speed; stride length; older adults
The ability of nurses to detect changes in health conditions, functional decline, or increasing
fall risk in aging residents who live in the community is a critical part of patient assessment.
Historically, these types of assessments have included direct observation, surveys, and
interviews to evaluate a resident’s condition change or declining functional status
(Rubenstein et al., 2004; Tideiksarr, 2003; Tinetti, 2003). Although many of these
assessments used by nurses have been validated and are reliable sources of data for trending
status changes in community-dwelling residents, they require a health care provider to be
present for the data collection and evaluation. Cutting edge models of care using sensor
technology that collects data 24/7 and results in many terabytes of data over several years
are being incorporated into residents’ living quarters in elder living communities to facilitate
faster decisions about nursing care needed and to make selected living environments safe for
Reprints and permissions: sagepub.com/journalsPermissions.nav
Corresponding Author: Lorraine J. Phillips, Sinclair School of Nursing, University of Missouri, S414 Nursing, Columbia, MO
65211, USA. phillipsl@missouri.edu.
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
HHS Public Access
Author manuscript
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Published in final edited form as:
West J Nurs Res
. 2017 January ; 39(1): 78–94. doi:10.1177/0193945916662027.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

older people (Dawadi, Cook, & Schmitter-Edgecombe, 2014; Reeder et al., 2013; Skubic,
Alexander, Popescu, Rantz, & Keller, 2009). However, because of their age, frail state, and
often times declining health, these residents may require enhanced nursing care
coordination. Sensor technology is being used to detect changes in a resident’s condition or
an episode of functional decline sooner, enhancing nurses’ ability to react to episodic
condition changes, perform a resident assessment more quickly, and possibly suggest
treatment options sooner (Rantz et al., 2013). We describe prediction of older adult fall risk
using big data sets captured over many years by multiple sensing devices and nurse care
coordinators’ use of these data in the care of these adults.
Fall Detection and Activity Monitoring Using Sensor Technology
Falls are the primary cause of injury in older adults, occurring at an annual rate of 33% in
persons above the age of 65 years (Centers for Disease Control and Prevention [CDC],
2015a). Early and efficient identification of older adults at risk for falls could prevent costly
injuries and loss of quality of life. The extant literature shows that people who walk slower
with shorter strides are more likely to fall (Barak, Wagenaar, & Holt, 2006; Quach et al.,
2011), but older adults do not receive regular gait evaluations. In an effort to increase fall
risk screening efforts, the CDC developed the Stopping Elderly Accidents, Deaths, and
Injuries (STEADI) toolkit for health care providers to use in clinical practice (CDC, 2015b).
Toolkit materials instruct providers how to conduct specific tests known to predict fall risk,
such as the Timed Up and Go, 30-s chair stand, and four-stage balance tests (Guralnik et al.,
1994; Jones, Rikli, & Beam, 1999; Podsiadlo & Richardson, 1991). However, providers face
time limitations, competing demands, and reimbursement constraints in evaluating and
managing fall risk among their patients (Tinetti, Gordon, Sogolow, Lapin, & Bradley, 2006).
Moreover, most falls go unreported, further limiting opportunities for providers to assess risk
and recommend preventive measures (Shumway-Cook et al., 2009).
One solution to the problem of underreported falls and existing gaps in fall risk screening
among older adults is in-home monitoring systems that passively and continuously capture
gait parameters from which fall risk can be predicted (Barak et al., 2006). Until recently,
human motion analysis using video required the attachment of body markers, as with the
Vicon system or the use of multiple cameras, and intensive computations to extract enough
silhouettes to fit a skeletal model (Stone & Skubic, 2011). In 2011, the Center for Eldercare
and Rehabilitation Technology research team at the University of Missouri (MU) developed
a methodology to obtain measurements of temporal and spatial gait parameters from the
Microsoft Kinect sensor. Using both a Vicon motion capture system and a web-camera-
based system to provide ground truth, researchers validated Kinect-recorded three-
dimensional (3D) depth images against standardized gait and balance measures. In addition,
fall-detection algorithms were initially developed with stunt actors performing falls in the
laboratory and further validated using stunt actors in the apartments of residents of
TigerPlace, an independent living community and MU research partner. Following iterative
refinement of the fall-detection algorithms, the Kinect fall-detection system has been
installed in a total of 39 TigerPlace apartments for continuous automated monitoring of gait
parameters for fall risk as well as automated fall detection in the apartments.
Phillips et al. Page 2
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

As interest in smart home technologies burgeons, other applications of Kinect sensors are
being tested (Reeder et al., 2013). Ejupi and colleagues (2016) used the Kinect in a
prospective study of fall events in community-dwelling older adults. They demonstrated that
upper extremity reaction time was significantly slower for older adults who experienced a
fall in the subsequent 6 months compared with those with no fall events. Ejupi et al.
envisioned the system could eventually be installed in clinical and home settings for ongoing
monitoring of fall risk. In other research, a Kinect sensor was paired with inertial sensors
attached to eating utensils for the purpose of tracking movements of interest to therapists
(Hondori, Khademi, & Lopes, 2012). In addition, rehabilitation exercise games using the
Kinect sensor are under development for use in the home setting (Shapi’i, Bahari, Arshad,
Zin, & Mahayuddin, 2015). Although this research demonstrates a variety of clinical uses
for the Kinect sensor, our research has advanced smart home technology by real-world
operation of the Kinect sensor in older adults’ residences.
Currently, novel sensor systems at TigerPlace use multimodal data sources to collect an
array of data 24/7 about residents. Some types of sensor systems are listed and described in
Table 1. These data sources extend the ability of TigerPlace’s nurses to detect baseline
changes in a resident’s activity levels, without having to be physically present for the
assessment. This is a shift from traditional models of assessment. For example, wall-
mounted sensor systems help nurses detect when a resident is not as active in their whole
apartment, and are spending more time in bed than usual.
Figure 1 illustrates motion density sensor data visualized through an electronic interface
used by nurses to evaluate activity in resident apartments in independent living. This specific
illustration also shows that an alert was generated based on a computation using a pre-
determined number of standard deviations change from the resident’s normal activity level.
An alert email was also sent to a nurse indicating that the resident was having increased
motion in their apartment on September 30, 2015. This increased motion can be seen by the
light gray highlighted areas between 2:00 and 6:00 a.m. on September 30, 2015. In addition,
the resident appears to be leaving the apartment (black areas between 2:00 and 6:00 a.m. on
September 30, 2015) during this same time frame. Based on the patterns observed in this
interface, this activity is irregular activity for this person compared with other nights in this
density map. This irregular activity is a situation that a nurse may want to review. Case study
analyses indicate that increasing periods of inactivity (decreased motion density) could
indicate the resident is experiencing higher levels of depression (Galambos, Skubic, Wang,
& Rantz, 2013), or it could indicate that the resident is experiencing an illness that is causing
greater fatigue and weakness, prompting more rest. Inactivity (black areas) could also mean
that the resident has left the apartment at unusual periods of the night or day which may be
an indicator of confusion or altered cognitive state.
The sensor system used in this research provides a complex and challenging opportunity to
combine different modes of sensor data input that can be triangulated to help nurses detect
changes in health conditions and functional decline in frail elders. Not only do sensor system
modalities help build a better picture of a resident’s current health status, but with the
addition of the Kinect sensors, they provide automated assessment of in-home gait
Phillips et al. Page 3
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

parameters and actual falls. Obtaining gait pattern data provides supplemental information
about residents’ health status because research has shown that—down to the 10th of a meter
per second—an older person’s pace, along with their age and gender, can predict their life
expectancy just as well as the complex battery of other health indicators (Harmon, 2011).
Therefore, to investigate whether Kinect sensor data could be used to identify individuals at
increased risk of falling, we analyzed the association between pre-fall changes in Kinect-
recorded gait parameters and known fall events in TigerPlace residents with Kinect sensors.
Method
Setting
TigerPlace is a senior housing community in Columbia, Missouri, with 54 independent
living apartments. Named for the MU mascot, it is an innovative living environment built
and operated by Americare Senior Living, in affiliation with the MU Sinclair School of
Nursing. Infrastructure is in place to support sensor networks in TigerPlace apartments.
TigerPlace provides a unique opportunity in which to develop and evaluate technology in a
collaborative setting with researchers from MU departments of Electrical and Computer
Engineering, Computer Science, Health Management and Informatics, and Schools of
Nursing, Health Professions, Social Work, and Medicine. These researchers form an
interdisciplinary team to develop and implement projects that improve the quality of life and
care of seniors. No other setting in the United States offers the population of subjects,
research infrastructure, faculty, and resources like TigerPlace; research projects are
encouraged, and residents who choose to participate enjoy the experience of developing new
technologies to help seniors’ age in place. TigerPlace residents, like other seniors in
Missouri and the United States, are concerned about maintaining independence and dignity.
TigerPlace is a key component of the Sinclair School of Nursing Aging in Place project that
was designed with MU faculty working with Americare Systems, Inc., Sikeston, Missouri,
to promote the independence of older adults. Nurses, physical therapists, occupational
therapists, and architects with an emphasis in environmental design participated in the
building plan. At TigerPlace, residents can age in place without fear of being moved to a
traditional nursing home unless they choose to leave. Findings indicate that with the right
supportive and restorative services, it is possible to help elders improve their health and
well-being and delay or prevent nursing home placement (Rantz et al., 2005). Registered
nurse (RN) care coordinators direct the wellness center that offers exercise and other health-
related classes, and conduct regular health assessments to promote residents’ well-being and
vitality. An electronic health record is used to document day-to-day care. Aging in Place
staff provide an array of home care services such as medication management, assistance
with activities of daily living, care coordination of health conditions, communication with
residents’ physicians and other health care providers, psychosocial services, and Medicare
home health care when residents need and qualify for that service.
Sample
Data required for the analysis in the present study were available for a convenience sample
of 23 TigerPlace participants with Kinect sensor systems and were collected over intervals
Phillips et al. Page 4
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

ranging from 3 to 48 months, the length of which depended on the duration of participants’
enrollment in the study and length of stay in the community. We excluded participants living
with another person because when co-habiting individuals have similar height and gait
velocities, their unique clusters of gait data are not discernable in the collective data. All
participants were Caucasian and 70% were female. At the time of admission to TigerPlace,
participants were, on average, 85.2 years of age, and at the time of data analysis, had a mean
length of stay of 49 months. As with TigerPlace research, the institutional review board at
MU approved the research and participants provided written informed consent.
Sensor Data Collected
The Kinect sensor continuously monitors and records residents’ in-home movement
occurring within sensor range. The Kinect sensor is typically located on a small shelf above
the front door to maximize the camera’s view of activity in the main living area. A computer
that logs and transmits the Kinect depth image data is located in a nearby kitchen cabinet.
Data are processed as depth images that appear as silhouettes to protect privacy. If the Kinect
system detected a potential fall, an alert was sent to staff who responded to check on the
resident. If the resident was found on the floor in response to a fall alert or the resident self-
reported a fall in response to the alert, the fall was recorded in the resident’s health record.
In the present study, a fall was defined as an observed or reported unexpected event in which
the resident came to rest on the ground or lower surface. In addition, falls that did not
generate an alert but that TigerPlace staff observed or falls that were otherwise reported, for
example, by residents themselves or family members, were recorded in residents’ health
records. After extracting fall event data from health records, we were able to determine the
frequency of falls for TigerPlace residents with Kinect system installations and then analyze
the association between participants’ fall events and change in in-home gait speed and stride
length. Figure 2 provides an illustration of the interface developed to display gait parameter
sensor data, longitudinally for 30 days. Figure 3 represents fall events for a participant
plotted along weekly gait speed estimates during an 87-week period.
Data Collection
Skeletal tracking with the Kinect Software Development Kit is limited to a range of 1.5 to 4
m, which is insufficient to capture walking sequences in some areas of TigerPlace
apartments. Therefore, raw disparity values from the Kinect depth stream, which is a
subtraction technique to separate foreground 3D objects (e.g., the person walking) from
background stationary objects, such as furniture, are processed to yield walk data for
distances of up to 8 m from the Kinect (Stone & Skubic, 2013). Algorithms developed for
the Kinect sensor automatically identify, segment and analyze walking segments of at least
1.2 m occurring within sensor range for gait speed, stride time, stride length, and height of
the individual walking. The height of each resident was measured a priori so that a cluster, or
mode, in the dataset is recognized as belonging to that resident. Modes that closely match
the known height of a resident are used for resident model initialization. Typically, a range
of 4 to 35 walks per day are required to identify walks belonging to a resident and not
another person. Assuming a dataset of walks exists for a resident, the model is updated using
data from a given time interval, or model window size, which may be as short as 2 weeks or
as long as 3 months. A longer model window size may be required to initialize a model for a
Phillips et al. Page 5
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

resident who walks infrequently. Stride parameters are extracted only for walks for which at
least five sequential steps can be identified. Further information on data science methods can
be accessed in related papers at the following website: https://www.eldertech.missouri.edu/.
A recent subset of the data (e.g., 1–2 weeks) used in model initialization is used to calculate
gait parameters. A 2-week gait parameter window size reflects longer term changes
compared with short-term fluctuations that may be represented in a 1-week gait parameter
window. Gait parameter estimates for any given day are calculated as a weighted average of
gait data from all walks identified in that resident’s apartment within the gait parameter
window. A weight is applied to each walk based on the likelihood that it reflects the resident
model. The window for both the model estimation and gait parameter estimation slide
forward one day and the process of model estimation repeats itself using the previous model
estimate for the update procedure. The process of identifying walking segments, extracting
gait parameters, updating resident models, and calculating gait parameters was and
continues to be completely automated. Data recorded by the sensor systems are continuously
logged 24/7 by participant identification number and stored on a secure server, of which
Kinect data account for 66 terabytes being used to develop algorithms for gait speed change
and fall detection. In totality, sensor network data account for 82 terabytes of storage space.
Depending on the length of time over which gait data are collected, daily gait parameter data
may be available in model window sizes of 30, 60, and 90 days and gait parameter window
sizes of 7 and 14 days, yielding up to six possible estimates each day for each gait parameter
(i.e., gait speed, stride length, stride time). For this study, window sizes of 30 days and 14
days were used for the model and gait parameter estimation steps, respectively. These
window sizes were selected so that a sufficient number of walks were included to allow
accurate resident modeling but not excluding cases lacking data that would be needed to
compute model window sizes of 60 or 90 days.
Data Analysis
Daily gait parameter estimates were available for 23 participants. For those residents that
had falls during the study period, we calculated cumulative change in gait parameters from
30 days before a fall. For those residents that had no falls during the study period, we chose
a random 30-day window to serve as a control. For a fall to be included in the model, the
resident had to have at least 14 days of gait parameter estimates (not necessarily
consecutive) or no more than 16 days missing for the entire 30-day window before a fall
event. A total of 69 falls met the aforementioned criteria, which resulted in 2,070
observations of gait parameters. We estimated a logistic regression model to predict the odds
and the probability of a fall event based on cumulative change in gait parameters at intervals
of 7, 14, 21, and 28 days. The outcome was whether a resident had a fall at the end of the
30-day window, and the predictor variable was cumulative change stratified by time
(numerical day in the 30-day window). Two separate models were run for cumulative change
in in-home gait speed and cumulative change in in-home stride length stratified by time. All
analyses were run in SAS 9.4 with a significance level of .05.
Phillips et al. Page 6
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

Results
Ten of the 23 participants had no falls during the monitoring period, which varied by length
of residence from 2011 to 2015. Participants with no fall events were, on average, 85 years
of age on admission to TigerPlace and had resided there for an average of 44 months at the
time of data collection. Residents who had fallen (
n
= 13) had between 1 and 12 fall events,
were 85.4 years of age, and had resided at TigerPlace for an average of 52 months.
Preliminary results indicate that cumulative change in speed over time is significantly
associated with probability of a fall (
p
< .0001). The odds of a resident falling within 3
weeks after a cumulative decline of 2.54 cm/s over 7 days is 4.22 (95% confidence interval
[CI] = [2.14, 8.30]) times the odds of a resident falling within 3 weeks after no change in in-
home gait speed. The model estimates that a cumulative decrease of 5.1 cm/s over 7 days in
in-home gait speed is associated with an 86.3% probability of falling within the next 3
weeks compared with a 19.5% probability for those with no change. The area under the
curve (ROC) is 0.86 indicating that cumulative change in gait speed is good at separating
residents that have fallen from non-fallers.
Similarly, preliminary results show that cumulative changes in stride length over time are
significantly associated with probability of a fall (
p
value < .0001). The odds of a resident
falling within 3 weeks after a cumulative change of 2.54 cm over 7 days is 6.78 (95% CI =
[2.69, 17.07]) times the odds of a resident falling within 3 weeks after no change in in-home
stride length. This model estimates that a cumulative decrease of 7.6 cm over 7 days in in-
home stride length is associated with a 50.6% probability of falling within the next 3 weeks
compared with 11.4% probability for those with no change. The ROC is 0.88 which
indicates that cumulative change in stride length is good at separating residents who have
fallen from non-fallers.
Discussion
Our results demonstrate the feasibility of using environmentally embedded sensors to
measure in-home gait parameters associated with the occurrence of falls. The Kinect sensor
system holds promise for unobtrusively monitoring older adults in their homes while
maintaining privacy and eliminating the burden of additional monitoring procedures. The
Kinect sensor system does not require the older adult to perform gait tests, wear devices, or
push a button after a fall. In addition, in-home gait parameters can be measured continuously
so that changes over time can be automatically and promptly detected, alerts generated, and
appropriate medical or therapy referrals arranged. In response to deterioration in gait speed
or stride length, best practice guidelines for screening and assessment and fall prevention
interventions could be initiated before a fall occurs and significant injury or disability ensues
(Kenny et al., 2011). Residents report feeling more secure in their home environment
knowing that the sensors will detect falls and that assistance will be provided when needed
(Jacelon & Hanson, 2013; Rantz et al., 2015).
Our model estimated that a 5.1 cm/s (i.e., 0.051 m/s) decrease in in-home gait speed was
associated with a high probability of falling within 3 weeks, and this is consistent with
Phillips et al. Page 7
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

published thresholds of 0.05 m/s for a small meaningful change in gait speed assessed over
distances of 10 foot, 10 m, or 4 m (Perera, Mody, Woodman, & Studenski, 2006). Similarly,
Kwon and colleagues’ (2009) analysis of gait speed change for 424 older adults in a clinical
exercise trial showed that a change of 0.03 to 0.05 m/s corresponded to an effect size of 0.2
and a change of 0.08 m/s corresponded to an effect size of 0.5. More importantly, that we
can detect similar declines in gait speed by recording and analyzing walking segments as
older adults go about their everyday activity means that decline may be detected earlier and
clinical interventions initiated sooner than would occur by intermittent clinical evaluation.
Benefits of early interventions in the presence of declining walk parameters extend beyond
reducing fall risk. Improvement, as opposed to decline, in gait speed of 0.1 m/s over 1 year,
has been associated with a 58% reduction in relative risk of mortality over the subsequent 8
years (Hardy, Perera, Roumani, Chandler, & Studenski, 2007).
Another approach to unobtrusive and ongoing gait assessment involves using passive
infrared motion sensors (Kaye et al., 2012). Kaye and colleagues (2012) used sensors
positioned in a sensor line (e.g., attached to the ceiling of a hallway or corridor) in the homes
of 76 community-dwelling older adults to estimate gait speed from the pattern and time
intervals of sensor firings and validated these estimates against performance-based mobility
tests. Furthering this work, Rana, Austin, Jacobs, Karunanithi, and Kaye (2013)
demonstrated that transition time between rooms, measured with infrared sensors positioned
throughout the dwelling, not only correlated highly with gait velocity measured along the
“sensor line” (
R
2 = .98) but yielded a greater number and variety of gait velocity estimates.
Although infrared sensors offer a compelling alternative to camera-based sensors for
measuring in-home gait speed, they lack the fall-detection functionality that our team has
developed with Kinect sensors.
The relevance of continuous in-home gait speed analysis is to be able to apply these findings
to the ongoing, real-time data and to automatically send alerts about increasing risk of a fall
to the older adult, nearby family members, and/or health care staff. With 3 weeks of lead
time, the older person could seek assistance to improve their functional capacity or health
status and possibly avoid the fall. Having lead time to alert the older person and others has
enormous potential to change the loss of functional abilities, trauma, and potentially death
that result from falls.
In addition, by identifying the in-home gait speed change threshold at which fall risk is
substantially increased, alert algorithms can be further refined for greater specificity. Rather
than use the pre-determined standard deviation multiplier developed initially through
retrospective analysis and clinician expertise, new alert parameters based on absolute change
in gait speed and stride length can be tested alongside the current alert algorithm. Refining
alerts for greater specificity is important not only to promptly identify decline in elders’
physical function but also to prevent false alarms that cause clinicians to become alert-
fatigued and desensitized to safety alerts (Agency for Healthcare Research and Quality,
2015). Our team is committed to ongoing efforts that advance sensor technology research
and optimize clinical relevance to support aging in place for the older adult population.
Phillips et al. Page 8
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

Our analysis is not without several limitations. One set of limitations is related to
generalizability of the findings. Although our sample of fall events included 69 falls, these
fall events came from a participant sample of 23 individuals and were restricted to
TigerPlace residents. Because TigerPlace residents receive RN care coordination services
and may age in place as their health and function decline, it is unclear how representative the
sample is of other independent living community residents.
Second, as noted above, not all falls were directly observed or captured by the Kinect sensor.
The accuracy of fall history could not be confirmed in some cases. That being said, many
studies on fall risk factors and fall incidence rely on self-reports of fall events as most falls
go unwitnessed (Barak et al., 2006; Rosen, Mack, & Noonan, 2013; Shumway-Cook et al.,
2009). However, “falls” were not explicitly defined within the health record at TigerPlace
and thus were subject to clinicians’ interpretation of post-fall circumstances or residents’
self-report.
Third, our analysis was a retrospective analysis of previously recorded sensor data. Use of
existing data not only has the advantages of reduced cost and burden but also, in the present
study, has the disadvantages of missing data related to sensor system failure and missing
data related to inadequate number of walks for resident modeling. It is not possible to know
whether residents with too few walks to compute individual models had fall events that were
not included in the analysis. Finally, apartments with greater than one resident were
excluded from the analysis because resident models could not be computed accurately at the
time the data were analyzed.
Future research could involve intervention trials to determine whether fall rates differ for
senior housing residents who receive therapeutic interventions based on real-time gait
parameter data from Kinect sensors compared with senior housing residents who receive
standard of care. Assisted living would be an ideal setting in which to test the Kinect sensor-
alert system because 24-hr staff are available to respond to fall alerts and falls are a major
problem in this setting, affecting 21% of residents in any 90-day period (Sengupta, Harris-
Kojetin, & Caffrey, 2015). In addition, future research to refine the alert algorithms as
described above could occur in tandem with a clinical trial or large-scale deployment to
yield greater accuracy in algorithmic predictions.
Sensor systems that assist nurse care coordinators to detect functional decline including
changes in gait that indicate an increased fall risk and actual falls are sorely needed to
facilitate earlier detection, treatment, and prevention of these costly events. Traditional
systems of assessment that require nurses to complete unidimensional instruments are risky,
because they are not as timely and they typically do not incorporate multiple assessment
modalities in a living environment like the sensors described in this study. Another benefit of
sensor data collected over months and years on the same residents is the longitudinal nature
of this big data that enhances the nurse’s ability to detect changes in baseline over time
much more effectively than do traditional assessment methods.
Acknowledgments
Funding
Phillips et al. Page 9
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of
this article: This work was supported by the National Institutes of Health (R01NR014255).
References
Agency for Healthcare Research and Quality. Patient safety primer: Alert fatigue. 2015. Retrieved
from https://psnet.ahrq.gov/primers/primer/28
Barak Y, Wagenaar RC, Holt KG. Gait characteristics of elderly people with a history of falls: A
dynamic approach. Physical Therapy. 2006; 86:1501–1510. [PubMed: 17079750]
Centers for Disease Control and Prevention. Important facts about falls. 2015a. Retrieved from http://
www.cdc.gov/homeandrecreationalsafety/falls/adultfalls.html
Centers for Disease Control and Prevention. STEADI materials for health care providers. 2015b.
Retrieved from http://www.cdc.gov/steadi/materials.html
Dawadi P, Cook DJ, Schmitter-Edgecombe M. Proceedings of the 2014 ACM International Joint
Conference on Pervasive and Ubiquitous Computing: Adjunct publication. New York, NY:
Association for Computing Machinery; 2014 Sep. Smart home-based longitudinal functional
assessment; 1217–1224.
Ejupi A, Gschwind YJ, Brodie M, Zagler WL, Lord SR, Delbaere K. Kinect-based choice reaching
and stepping reaction time tests for clinical and in-home assessment of fall risk in older people: A
prospective study. European Review of Aging and Physical Activity. 2016; 13(1):2. [PubMed:
26865881]
Galambos C, Skubic M, Wang S, Rantz M. Management of dementia and depression utilizing in-home
passive sensor data. Gerontechnology: International Journal on the Fundamental Aspects of
Technology to Serve the Ageing Society. 2013; 11:457–468. [PubMed: 24049513]
Guralnik JM, Simonsick EM, Ferrucci L, Glynn RJ, Berkman LF, Blazer DG, Wallace RB. A short
physical performance battery assessing lower extremity function: Association with self-reported
disability and prediction of mortality and nursing home admission. Journal of Gerontology. 1994;
49(2):M85–M94. [PubMed: 8126356]
Hardy SE, Perera S, Roumani YF, Chandler JM, Studenski SA. Improvement in usual gait speed
predicts better survival in older adults. Journal of the American Geriatrics Society. 2007; 55:1727–
1734. [PubMed: 17916121]
Harmon K. Walking speed predicts life expectancy of older adults. 2011 Retrieved from http://
www.scientificamerican.com/article/walking-speed-survival/.
Hondori HM, Khademi M, Lopes CV. Monitoring intake gestures using sensor fusion (Microsoft
Kinect and inertial sensors) for smart home tele-rehab setting; 2012 1st Annual IEEE Healthcare
Innovation Conference; 2012. Nov, Retrieved from http://www.ics.uci.edu/~mkhademi/files/
Publications/HIC2012.pdf
Jacelon CS, Hanson A. Older adults’ participation in the development of smart environments: An
integrated review of the literature. Geriatric Nursing. 2013; 34:116–121. [PubMed: 23276642]
Jones CJ, Rikli RE, Beam WC. A 30-s chair-stand test as a measure of lower body strength in
community-residing older adults. Research Quarterly for Exercise and Sport. 1999; 70:113–119.
[PubMed: 10380242]
Kaye J, Mattek N, Dodge H, Buracchio T, Austin D, Hagler S, Hayes T. One walk a year to 1000
within a year: Continuous in-home unobtrusive gait assessment of older adults. Gait & Posture.
2012; 35:197–202. [PubMed: 22047773]
Kenny RA, Rubenstein LZ, Tinetti ME, Brewer K, Cameron KA, Capezuti L, Peterson EW. Summary
of the updated American Geriatrics Society/British Geriatrics Society clinical practice guideline
for prevention of falls in older persons. Journal of the American Geriatrics Society. 2011; 59:148–
157. [PubMed: 21226685]
Kwon S, Perera S, Pahor M, Katula JA, King AC, Groessl EJ, Studenski SA. What is a meaningful
change in physical performance? Findings from a clinical trial in older adults (the LIFE-P study).
The Journal of Nutrition, Health & Aging. 2009; 13:538–544.
Phillips et al. Page 10
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

Perera S, Mody SH, Woodman RC, Studenski SA. Meaningful change and responsiveness in common
physical performance measures in older adults. Journal of the American Geriatrics Society. 2006;
54:743–749. [PubMed: 16696738]
Podsiadlo D, Richardson S. The timed “up & go”: A test of basic functional mobility for frail elderly
persons. Journal of the American Geriatrics Society. 1991; 39:142–148. [PubMed: 1991946]
Quach L, Galica AM, Jones RN, Procter-Gray E, Manor B, Hannan MT, Lipsitz LA. The non-linear
relationship between gait speed and falls: The MOBILIZE Boston Study. Journal of the American
Geriatrics Society. 2011; 59:1069–1073. [PubMed: 21649615]
Rana R, Austin D, Jacobs PG, Karunanithi M, Kaye J. Gait velocity estimation using “time
interleaved” between consecutive passive IR sensor activations. 2013 Retrieved from http://
arxiv.org/pdf/1310.4880.pdf.
Rantz MJ, Dorman-Marek K, Aud M, Johnson RA, Otto D, Porter R. TigerPlace: A new future for
older adults. Journal of Nursing Care Quality. 2005; 20:1–4. [PubMed: 15686069]
Rantz MJ, Skubic M, Miller S, Galambos C, Alexander GL, Keller J, Popescu M. Sensor technology to
support aging in place. Journal of the American Medical Directors Association. 2013; 14:386–391.
[PubMed: 23562281]
Rantz MJ, Skubic M, Popescu M, Galambos C, Koopman RJ, Alexander GL, Miller SJ. A new
paradigm of technology-enabled “vital signs” for early detection of health change for older adults.
Gerontology. 2015; 61:281–290. [PubMed: 25428525]
Reeder B, Meyer E, Lazar A, Chaudhuri S, Thompson HJ, Demiris G. Framing the evidence for health
smart homes and home-based consumer health technologies as a public health intervention for
independent aging: A systematic review. International Journal of Medical Informatics. 2013;
82:565–579. [PubMed: 23639263]
Rosen T, Mack KA, Noonan RK. Slipping and tripping: Fall injuries in adults associated with rugs and
carpets. Journal of Injury and Violence Research. 2013; 5:61–69. [PubMed: 22868399]
Rubenstein LZ, Solomon DH, Rothe CP, Young RT, Shekelle PG, Chang JT, Wenger NS. Detection
and management of falls and instability in vulnerable elders by community physicians. American
Geriatrics Society. 2004; 52:1527–1531.
Sengupta M, Harris-Kojetin LD, Caffrey C. Variation in residential care community resident
characteristics, by size of community: United States, 2014 (NCHS data brief, No. 223).
Hyattsville, MD: National Center for Health Statistics; 2015 Nov.
Shapi’i A, Bahari NN, Arshad H, Zin NAM, Mahayuddin ZR. 2nd International Conference on
Biomedical Engineering (ICoBE), 2015. Piscataway, NJ: Institute of Electrical and Electronics
Engineers; 2015 Mar. Rehabilitation exercise game model for post-stroke using Microsoft Kinect
camera; 1–6.
Shumway-Cook A, Ciol MA, Hoffman J, Dudgeon BJ, Yorkston K, Chan L. Falls in the Medicare
population: Incidence, associated factors, and impact on health care. Physical Therapy. 2009;
89:324–332. [PubMed: 19228831]
Skubic M, Alexander GL, Popescu M, Rantz MJ, Keller J. A smart home application to eldercare:
Current status and lessons learned. Technology in Health Care. 2009; 17:183–201.
Stone E, Skubic M. Evaluation of an inexpensive depth camera for in-home gait assessment. Journal of
Ambient Intelligence and Smart Environments. 2011; 13:349–361.
Stone E, Skubic M. Unobtrusive, continuous, in-home gait measurement using the Microsoft Kinect.
IEEE Transactions on Biomedical Engineering. 2013; 60:2925–2932. [PubMed: 23744661]
Tideiksarr R. Best practice approach to fall prevention in community-living elders. Topics in Geriatric
Rehabilitation. 2003; 19:199–205.
Tinetti ME. Preventing falls in elderly persons. The New England Journal of Medicine. 2003; 348(1):
42–49. [PubMed: 12510042]
Tinetti ME, Gordon C, Sogolow E, Lapin P, Bradley EH. Fall risk evaluation and management:
Challenges in adopting geriatric care practices. The Gerontologist. 2006; 46:717–725. [PubMed:
17169927]
Phillips et al. Page 11
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

Figure 1.
Motion sensor data density visualization including night-time motion density: Increase alert.
Phillips et al. Page 12
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

Figure 2.
Gait parameter interface showing calculated stride length and time, gait speed, and alerts.
Note.
TUG = Timed Up and Go.
Phillips et al. Page 13
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

Figure 3.
Weekly gait speed averages with falls plotted on in-home gait speed graph.
Note.
○ = average gait speed; ★ = fall event.
Phillips et al. Page 14
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.
Author Manuscript Author Manuscript Author Manuscript Author Manuscript

Author Manuscript Author Manuscript Author Manuscript Author Manuscript
Phillips et al. Page 15
Table 1
Sensor System Modalities and Descriptions.
Type of sensor Description
Passive infrared
(PIR) Installed to detect presence in a particular room (e.g., living
room, by the door or bathroom) as well as for specific
activities. For example, a motion sensor installed on the
ceiling above the shower detects showering activity.
Bed sensor The bed sensor is a set of four hydraulic transducers installed
under the mattress which captures a ballistocardiogram,
respiration, and bed restlessness as someone lies on the bed.
Specialized signal processing is used to compute heart rate,
respiration rate, and restlessness every 15 s.
Depth camera A depth camera is used to compute gait parameters (gait
speed, walks/day, and stride length). The depth camera also
uses a fall-detection algorithm to detect falls in real-time and
send alerts to designated individuals (e.g., clinical staff) with a
link to a depth video (shadow-like silhouette images) showing
the fall.
Doppler radar Doppler radar systems have been tested to measure gait
parameters and to detect falls. A Doppler radar transmits an
electromagnetic wave at a specific frequency and measures
the shifts in the reflected waves, like weather radars. These
frequency shifts can then be used to obtain the velocities of
the person’s body parts in the radar’s direction.
West J Nurs Res
. Author manuscript; available in PMC 2018 January 27.