Introduction

Physical activity (PA) of nursing home residents is researched in the context of increasing health outcomes and well-being [1]. Examples are the association of PA with sleep [2], number of falls [3], or energy expenditure and intake [4]. Dementia and PA influence each other. Depression and apathy might manifest in passivity and low PA, while restless behaviors will manifest as higher PA [5]. Any measurement tool for PA will hence have to be able to account for a large range of activity levels. Additionally, residents with dementia have more non-psychiatric comorbidities than other residents [6]. If not taken into account, comorbidities affecting heart function might make PA estimation based on heart rate less reliable. As in other geriatric patient populations, any measurement tool will have to account for the use of walking aids, such as walkers or canes.

Here, we report on the feasibility and the limitations of measuring PA with a consumer wearable in a Dutch nursing home. The presented data were collected as part of a research project that assessed the impact of providing unobstructed access to green spaces to residents with dementia. The main research question was how more access to green spaces affects the residents’ quality of life, as well as the embedding of dementia care in the local community. We therefore assessed the PA of the residents with different measurement tools, collected data with installed sensors and conducted observations at the nursing home.

Commonly used assessment tools for PA of nursing home residents without cognitive impairment can be divided into three categories: observations of functional fitness tests, questionnaires, and technical measures [1]. The first two types of tools require the participant to understand and follow instructions. For residents with dementia, this might pose a problem if the test is performed at the wrong moment, e.g., after a troubling experience for the resident. Furthermore, a participant’s willingness to cooperate can fluctuate by day and time of the day. For our study, we opted for a consumer-grade fitness wearable, which was easy to put on and remove as well as water proof, which was an important requirement for the nursing staff. Furthermore, commercial wearables are often outfitted with more sensors, such as magnetometers, barometers, or GPS receivers. Those additional sensors collect useful data for activity recognition methods [7]. The option of tracking location outside by GPS was used to assess the use of green spaces by residents in another part of the study. To anchor the sensor data in real-life observations, we used data collected with the Medlo observational tool. The Medlo tool [8] was developed to capture the daily life activities of residents in nursing homes and daycare centers. Observations are logged per minute in 30-min sessions and observers use an observation sheet covering four dimensions of daily life activities in ten variables: activities (type, engagement, PA), physical environment (location, interaction with environment), social interaction (occurrence, type, interaction partner), and emotional well-being (mood, agitation). For the aim of this paper, we focused on the variable “physical activity”, which is scored on a 7-point Likert scale. We used off-the-shelf Samsung wrist-worn wearables to record accelerometer data. Additionally, we recorded manually the “24hrs log” of the watch, which are movement categories that are calculated and displayed on the device (low activity, medium activity, high activity, sleep, run, swim, cycle). We wanted to find out if those movement categories are useful in a mostly sedentary patient population. Furthermore, we wanted to compare PA estimates based on raw accelerometers with the modeled movement categories and observed activity levels.

Materials and methods

Results

Summary statistics

For both the Medlo and the “24hrs log” categories, we see mainly sedentary behaviors. Residents spent the most time in the categories “inactive” for the wearable categories (Figure 1) and “sitting quietly” for the Medlo (Figure 2). The wearables detected approximately the same amount of “light” activity as the Medlo observations of “ML04 = standing” and “ML05 = standing activity/walking around” combined. The category “not on wrist” refers to moments when participants wore the watch, but the watch classified those moments wrongly as “not on wrist”. We collected almost no observations for the Medlo categories “ML03 = light to moderate sitting”, “ML06 = active walk/cycling”, and “ML07 = sports/full body movement” as well as for the “24hrs log” categories of “heavy”, “biking”, and “running”. We therefore removed those variables from further analysis. The MAD has an overall mean of 11.15 milligravity units (range: 4.5 to 20.4 milligravity units). This is in accordance with Bakrania et al. [11] who found a mean between 6 and 12 milligravity units (depending on the device) in their sample for sedentary behaviors.

Display full size

Figure 1. 

Boxplots of the collected “24hrs log” data of all participating nursing home residents (N = 16). Per activity, we plotted boxplots of the percentage of measurement time that residents were active in one category. The “not on wrist” category refers to moments that residents wore the watch but it was categorized as if the watch was not worn

Display full size

Figure 2. 

Boxplots of the collected Medlo data of all participating nursing home residents (N = 13). Per Medlo activity category, a boxplot shows the percentage of time that residents were active in one category. Categories are: ML01 = lying/no movement, ML02 = sitting quietly, ML03 = light to moderate sitting, ML04 = standing, ML05 = standing activity/walking around, ML06 = walking activity/cycling, ML07 = sports/whole body movement

Correlation analysis

We correlated all variables with each other after preprocessing the raw data. The correlation matrix in Figure 3 is sorted by hierarchical clustering using the complete linkage method. The clusters show clearly that the MAD-derived metrics correlate somewhat positively with the more active categories of the Medlo (ML04, ML05) and the “light” activity of the watches’ activity classification. There is also a positive correlation with the category “not on wrist”. They correlate negatively with the less active categories of the watches’ model (sleeping, inactive) and the Medlo (ML01, ML02). Unsurprisingly, all MAD-derived parameters correlate highly with each other, however, there are differences in association strength with the other variables. The SD and 95th percentile seem to differentiate best between the more active and more inactive Medlo and watch categories. There are larger positive and negative correlations with the watch categories than with the Medlo categories. The directions of the correlations are as expected; positive with the more active categories and negative with the less active categories. The watches’ and Medlo variables also correlate with each other in the expected way. This includes the watch category “sleeping” that correlates highly (0.66) with the ML01 category, which describes the sleeping and resting of participants. The “not on wrist” category has overall low correlations with other variables but a high correlation with ML05 (0.65), which describes light-standing activities.

Display full size

Figure 3. 

Correlation matrix of all variables. The correlation matrix is sorted hierarchically, which yields two clusters, one with the derivates of the mean amplitude deviation (MAD) and three active categories plus “not on wrist” and another cluster of the least active movement categories. MADrel: time spent above 100 milli gravitational units of the mean amplitude deviation; MADfrag: fragmentation rate of the mean amplitude deviation; MADsd: standard deviation of the mean amplitude deviation; MADqant: 95th percentile of mean amplitude deviation

Within the Medlo categories, there are high negative correlations between ML02 (sitting quietly) and ML04 (–0.81) as well as ML05 (standing and walking around) (–0.83). The more a resident is sitting the less they are standing or walking around. Similarly, within the Samsung categories, we can see a high negative correlation (–0.89) between inactivity and light activity. Again, the more a resident is inactive, the less light activity is recorded. The sleep categories of both the Medlo and the watch are somewhat puzzling as they do not correlate with other variables of the same category, apart from a negative correlation (–0.41) between ML01 (sleep) and ML04 (standing) as well as sleep and inactive (–0.42). The inactive category also correlates negatively with “not on wrist” (–0.39). Unfortunately, we cannot explain those correlations without knowing more about the proprietary Samsung activity classification.

Discussion

Our main question was if we can use consumer wearables to estimate activity levels in nursing home residents with dementia, and how those estimations would relate to observational data. The summary statistics confirmed our expectation of a rather sedentary sample. For the Medlo scores, this is also an artifact of the data collection process, where Medlo observations took place in shared living rooms, which invite residents to sit down and generally do not encourage physical activities. Activities outside the observer’s field of view were registered as “non-observable”, which might cause an underestimation of PA [12]. The wearable data confirm low activity levels over the entire measurement period, which includes time spent outside shared living spaces and outside registered observations. We did not collect data during morning and evening routines, which we would expect to come with a big share of PA, such as walking to and from the living room, moving in and out of bed, and bathroom routines. In future research, one could use wearables during those moments to collect data without breaching residents’ privacy. This would give a more complete picture of the residents’ PA throughout the day. Overall, the found correlations are according to our expectations, with higher scores on MAD metrics being associated with more time spent in higher activity categories. We found positive correlations between the MAD metrics and the active Medlo categories (ML04 and ML05). It is noteworthy that the correlations between the watches’ “light” activity category and the Medlo categories are more pronounced than the correlations between the MAD metrics and the Medlo categories. In our sample, conclusions based on any of the collection tools will be very similar. Apart from PA levels, researchers might be interested in how engaged participants are when performing certain activities or in which context the activity takes place. Observational tools can generate such additional information. We argue that PA scores can be estimated similarly well with observations and wearables. A combination of both however would allow researchers to set the results of the observation period into the context of an entire day of PA. Data from wearables can provide this context without having to rely on carers’ or residents’ memory of that day. For our analysis, we derived several MAD metrics. From the correlation matrix, we see that there are differences in how those metrics correlate with the activity categories. Especially, the SD and the fragmentation rate seem to differentiate better between categories. We suggest including derivatives for the spread of MAD when using raw accelerometer data in predictive models to increase predictive performance.

The presented data are limited by a rather small sample. However, we collected per person a long interval of movement data with the wearables. This makes the associations between the watch-derived data points more robust. Our sample includes almost exclusively women. Future research should ensure to have a gender-diverse sample and register the use of mobility aids throughout the data collection. In an experiment with more participants, it could also be investigated whether setting the watches to the participants’ true characteristics would positively affect the data. A bigger and more diverse sample would allow for subgroup-specific conclusions. At the level of recorded sensor data, future work might consider extending the recordings to other in-built sensors such as the gyroscope or barometer as well as recording physiological measurements such as heart rate. It would be interesting to compare those measurements to observations or as input for machine learning.

Regarding limitations of the used wearable, we would like to point out that this watch was not designed with nursing home residents in mind. The wrists of most residents were very slim, which might add movements of the watch around the wrist to the data. The watches were however comfortable enough to be worn for long stretches of time by most participants, which makes them a good tool for ubiquitous measurement.

We would like to point out that Medlo is not the only observational tool to assess activity levels amongst dementia patients. Any future research should include other tools (surveys, diaries, medical records) to compare the wearable data with and to ground the conclusions drawn based on wearable data.