In this study, we examined smartwatch-derived characteristics of everyday behavior (physical activity) and subjective frailty status in older adults. The main objective was to investigate and better understand the relation between sensor-based measures of daily physical activity, separated into upper limb activity and gait (ambulation), and the self-estimation of frailty levels. The correlation of the derived components gait and activity with the self-reported frailty scores were weak to moderate. Cluster analysis resulted in two clusters with either low (cluster 3) or high activity (cluster 2) in the dimensions gait and activity. The third cluster (cluster 1) was characterized by high activity and low extent of gait. The odds of being female and frail was significantly increased for cluster 1. Additionally, the odds of having a walking aid as women compared to men was increased, too.
The correlation analyses between the frailty level of our cohort (exclusively self-reported), showed moderate negative associations with the gait parameter ‘STEP95’ (cadence-based). Additionally, weak to moderate negative correlations were found between the activity parameter ‘MADmedian’ (MAD-based) and both scores. The difference between the two frailty scores was that we excluded the criteria ‘muscle strength’ and ‘weight loss’ for the reduced version (0:3) to examine whether the explained variance increased when we omitted these parameters, which may not be directly measurable with a wrist-worn sensor. In general, however, both scores revealed comparable results. A more direct comparison between the objective physical activity data and the subjective frailty components was done by separate analyses of variance. Both, the parameter ‘STEP95’ and the dimension ‘gait’ showed differences with medium effects for the criteria ‘weakness’ and ‘physical activity’. The upper limb related activity (‘MADmedian’ and ‘activity’) showed differences with a medium effect for the ‘physical activity’ item. These results seem to be consistent with the content of the questions. The ‘physical activity’ question includes all types of physical activity, regardless of whether the upper or lower limbs are affected. However, the ‘weakness’ question specifically asks about the ability to climb one or more flights of stairs. Furthermore, the dimension ‘gait’ showed additional small effects for the criterion ‘weight loss’ and ‘exhaustion’. Our findings are consistent with previous studies stating that the level of activity (respectively inactivity) is associated with different frailty levels [18,19,20,21, 23, 24] and that especially gait (ambulation) related parameters seem to be more sensitive [18, 19]. In 2012, Theou and colleagues [18] presented a wide range of comparisons between different PA measures (e.g., accelerometer, heart rate (HR), electromyography (EMG), global positioning system (GPS), and Minnesota Leisure Time Physical Activity Questionnaire (MLTAPQ)) with each other and with the Frailty Index (FI). They found that the FI was significantly correlated to all PA measures. For accelerometry, ‘total steps’ and ‘total activity minutes’ were most strongly correlated to FI [18]. Our study extends this body of evidence by showing that there is also a correlation, however only weak to moderate, between an exclusively self-reported frailty score, including five short and easy questions, and gait and activity-related parameters as assessed by accelerometry. Furthermore, for our parameters related to gait, a significant difference was found for almost all frailty criteria, suggesting that mobility might be the driving parameter related to frailty, although the relationship between frailty and behavior might be multimodal, as seen, for instance, in the relation with falls [36]. Still, about 60% of the variance of behavior and the FI frailty assessment [18] (mixed assessment) as well as approximately 75% of the comparison with the self-reported frailty questionnaire remains unexplained. In addition, it should be considered that the association between pedometry and self-reports may be due to the fact that the self-report explicitly asks for gait function. However, R2 did not change when non-PA questions were removed, indicating a generalizability of ambulation (via pedometry) to other domains.
The debate about the relationship between objective and subjective activity assessments in older adults is well known (e.g., [13, 37, 38]). The lack of complete agreement could indicate insufficient validity of self-reports or objective assessments, or (alternatively) low sensitivity of self-reports and objective measures [39]. This may include various aspects, such as detail of the questionnaire, extent of supervision, or the length of the recall period. Questionnaires may cover periods from one to seven days, or even up to several months. Answers are dependent on the subjects’ age, their living environment as their health/behavioral reference (e.g., when an old person compared him/herself with the younger neighbor), and the context of questioning [40]. In contrast to self-reports, objective assessments using wearable sensors can provide accurate documentation of daily activities such as walking, standing, sitting, and lying down [24, 25], which in turn, may allow to identify frailty-specific patterns in peoples’ natural environment [23]. Different levels of frailty may be characterized by differences in daily PA patterns, such as fragmented walking distances (e.g., due to exhaustion, declining strength, or walking indoors instead of outdoors) or lower PA complexity [23]. However, in addition to the simple comparison of objective and subjective measures, cluster analyses might contribute to the validation of self-reports as a new approach and, in this sense, this may lead to a better understanding of what the self-report measures assess.
Based on the behavioral data (acceleration data), our cluster analysis resulted in three clusters that differed in terms of their upper extremity pronounced activity level and extent of gait (ambulation). Cluster 2 appeared to represent the most active group showing the highest extent of ambulation and activity, followed by cluster 1 (high activity and low extent of ambulation) and the least active cluster 3 (low activity and low extent of ambulation). Participants allocated to cluster 2 did not only show a different extent of gait and activity level, but also had a better self-rated health status (mean = 3.0; i.e., ‘good’) and less frequent use of walking aids. In contrast, participants in cluster 1 and cluster 3 showed a comparable use of walking aids. Therefore, the odds ratio of using walking assistive devices was significantly increased for cluster 1 and 3 compared to cluster 2. Yet, 33% of the participants classified into cluster 2 were categorized as frail and 45% as pre-frail. Our analyses of variance further revealed significant differences between the clusters and the frailty scores (both full and reduced version). Post hoc analyses showed that the clusters in which participants showed lower extent of ambulation also contained subjects with a higher frailty score (cluster 1 and cluster 3). This led to one robust group (cluster 2) and two frailty subtypes. However, for elderlies allocated to cluster 1, not only the risk of being frail was significantly increased, but also the probability of being female. Only the parameter ‘gait’ showed clear differences between all three clusters with a large effect. Consequently, cluster 1 contained the participants with the highest frailty scores, most of whom were female, and on average the oldest group. Interestingly, this finding seems to be consistent with the ‘male-female health-survival paradox’, which states that women live longer than men, but with poorer health [41] as they usually experience more functional limitations, co-morbidities, and poorer self-rated health [42, 43]. Studies using the FI consistently show that women have higher FI scores than men at all ages even though they tolerate a higher degree of frailty (lower mortality rate at any given FI score or age). Therefore, they can be seen as more frail (due to a poorer health status) and less frail (lower risk of mortality) at the same time [41]. Gender differences also seem to be reflected in activity behavior. In a study of Li et al. [44] PA was assessed by the CHAMPS (Physical Activity Questionnaire for Older Adults) and accelerometry for 7 consecutive days. They analyzed the data of 114 older adults (mean age 74.0 ± 6.0) and observed that preferences for level, type, and location of PA differed substantially between gender [44]. In addition, there is evidence based on sensor measures that men engage in more MVPA than women (e.g., [45, 46]), and that there might be also gender differences in the amount of time spent in lower intensity domains, such as sedentary and light activities [47]. Accordingly, even though men might achieve a greater amount of MVPA, they spend more time sedentary, whereas women may accumulate a greater number of light activities [17, 46, 47]. This fits quite well to the results of our cluster analysis showing that cluster 1 (most frail participants and 71% female) still performed activities connected to upper limb movements. Light housework (e.g., dusting, sweeping), or even cooking, body hygiene, and knitting, for example, may therefore have led to increased hand activity in women in our cohort. Additionally, potential gender issues may be present in the use of walking aids as well. There is evidence showing that predominantly women use walking aids, e.g., [48]. Consequently, this may have had an effect in the subjective frailty scoring, too. The frail group, based on the subjective frailty score, included 67% women and 72% walking aid users. However, for the results of the cluster analysis, there was no significant difference between cluster 1 and 3 regarding the use of walking aids.
Considering cluster 2 as the robust group – in terms of gait (ambulation) and upper limb PA, the differences between the two frail groups (1 and 3) remain to be discussed. Cluster 3 showed what the stereotype of frailty suggests: inactivity in both the gait (ambulation) and the upper limb PA dimension. The odds ratio analysis of each separate criterion revealed that especially participants within cluster 3 reported low PA and showed a trend towards increased weakness i.e., ‘climbing stairs’ and tendency towards low muscle strength i.e., ‘lifting a heavy bag’. Cluster 1 was prone to both a high risk of low PA and weakness. Moreover, participants in cluster 1 showed a tendency towards exhaustion i.e., ‘lack of energy’. Additionally, the parameter MADfrag, which reflects the relation of long and short activity bouts, and therefore displays the specific changes in intensities, was particularly low for people allocated to cluster 3. This is in line with our previous study in which we found that higher frailty scores were associated with more monotonous behavior during two activities of daily living (gardening and tea task) [35]. To what extent the use of walking aids might be seen as a consequence of frailty or might lead to the classification as frail, could be answered by cluster 1. Participants in this cluster walked even less than the other frail group (cluster 3), while showing the same upper limb related activity rate as the robust group. This group (cluster 1) had walking aids in 80% of cases. Thus, assuming that pedometers have problems detecting steps in people using walkers [26], analyzing walking activity alone might not provide valid information about frailty. Additionally, the questions related to ‘weakness’ in the questionnaire used in our study (equivalent to ‘walking speed’ of the Fried frailty score [9]) was about difficulties going up one or several flights of stairs without resting. This tends to basically exclude people with walking aids as it becomes more difficult to climb stairs and consequently might result in a positive rating for this criterion. This is in line with existing evidence. Nunes et al. [15], for example, showed in their study that comparing the Fried frailty criteria with a self-reported frailty questionnaire, ‘decreased walking speed’ showed a rather low specificity of 31.4%. Therefore, gait-related self-reported information may be misleading when it comes to frailty. Barreto and colleagues [25] investigated a self-reported frailty screening tool at baseline and one year later. Their analysis showed that frail individuals were older, predominantly female, had more co-morbidities, a greater decline in physical function, suffered more often from chronic pain, and reported decreased health status. Furthermore, they stated that frailty is a single entity, different from co-morbidity and physical limitation [25]. The question that arises from the train of thoughts is whether a person remains active, regardless of physical or neurological impairments that might impair walking. Although the values of the metabolic equivalent of task (MET) differs between gait and upper limb related activities, motor function of upper extremity has been identified as an important predictor of, e.g., disability [49]. In general (in terms of METs), lower body activities should have higher energy expenditure in comparison to upper body activities because they involve major muscle groups and the whole body mass is moving (instead of just an arm). Therefore, it could be meaningful to investigate changes in the different categories (ambulation and upper body activity), as changes in physical behavior may represent the first sign of frailty. This leads to the idea of a parameter for early detection of frailty. If we assume that cluster 1 is self-classified as frail solely due to ambulatory impediments, more specific actigraphy assessments of everyday behavior could help to get more information on the overall activity level of the person as well as to differentiate between frailty and possible frailty-biases (e.g., using walking assistive devices or questions, which disadvantage people with walking aids). This could ultimately help to identify causes and mechanisms of what we consider physical frailty. However, in our study, the number of robust older adults was far less than the pre-frail and frail older adults. This might be due to the fact, that the frailty classification was solely based on the self-estimation of the individuals. Consequently, the participant may have considered themselves in a worse condition than they are.
This study has several strengths, but also limitations that need to be addressed and therefore, results should be interpreted with caution. First, we intended to increase the compliance and wear time of the smartwatch by using the wrist of choice of the participants [24, 26]. This further allows for continuous data collection and the capturing of more commonly performed tasks of daily living [26], including gait activity. This resulted in significant group differences between placement on the left or right wrist between cluster 1 and cluster 2. However, this may not have been of relevance, as these clusters differed in the extent of ambulation (gait), but not in the amount of hand-related activity. It is possible that this is a consequence of wearing the watch in a balanced proportion on the non-dominant or dominant hand between both groups, or more bimanual hand activity [50]. While this is solely based on speculation, as we had no information about hand dominance, activity in cluster 1 might have been overestimated due to more frequent measurement on the dominant hand compared to cluster 2. With regard to future studies, we recommend to assess handedness and define where the sensor should be worn. Second, while most clinical tests are geared towards measuring maximum physical performance (capacity), the use of wearables in everyday life, in contrast, is aimed at recording actual (submaximal) behavior. However, individuals’ submaximal behavior and its relation to capacity could deviate. This is usually the case unless the capacity is greatly reduced and thus no longer different from actual everyday behavior. Therefore, the interpretation of the actual behavior remains difficult. In addition, accelerometry (in the proposed form) is not able to detect the movement of non-body masses (e.g., carrying groceries) and could therefore underestimate the energy expenditure. Third, devices worn on the wrist are known to be susceptible to interference with the use of assistive devices such as walkers [26, 51], which is increasingly the case in older age. In this context, especially the use of rolling walkers leads to a rather stationary wrist, and it is still unclear how much movement is actually registered by the device [26]. Although we assessed whether participants used an assistive device, we did not further subdivide the assistive devices (e.g., walking sticks or rolling walkers). For future studies, the simultaneous use of hip or ankle pedometers and wrist-worn bands might be an attractive solution, since wearables are becoming smaller and cheaper. If such instruments were combined, this could also provide more insights into the importance of energy consumption during certain activities. Another limitation refers to the smartwatch itself. As we had some major issues with the charging process and data collection, some participants were not able to handle the smartwatch independently and dropped out of the study. Consequently, the dropout rate was around 40%, which is quite high. This also affected the wear time of the watch drastically and we had to adapt the valid measurement hours to decrease data loss and the burden for the participants. The current wear time recommendation for valid accelerometer measurements is ≥10 hours/day for 6-9 days (e.g., [28]). Since we had issues with the frequent charging process of the watch, we adapted the amount of valid measurement hours down to ≥8 hours to decrease the burden for the participants. However, since we had an extended wearing period of at least 10 days for 90% of the sample, this might still achieve good reliability in capturing PA. Another limitation which needs to be discussed is the use of a convenience sample and therefore the lack of generalizability. Due to the potential bias of the sampling technique, subgroups might be under-represented, e.g., those who were not interested in the topic (public dissemination). In this regard, care managers were involved in the recruitment process to reach out for people in nursing homes as well. For future research, we would strongly recommend the use of devices with higher battery life and greater robustness towards unintentional adjustment of measurement setting.