The use of wearable sensor technology (e.g., accelerometers) for tracking human movement has allowed for efficient measurement of activity of the upper limb (UL) in daily life (1–6). Accelerometry has become an established, valid and reliable methodology to directly measure performance of UL activity in daily life in neurologically intact adults (7, 8) and adults with stroke (9–13). Per the World Health Organization International Classification of Functioning, Disability and Health (ICF) model (14), activity performance, defined as what a person does in the unstructured, free-living environment, is a different but related construct to the capacity for activity (i.e., functional capacity), which is measured by standardized assessments in the structured clinical or laboratory setting. Clinicians and researchers typically assess a person's functional capacity for activity in the structured clinic or laboratory environments with standardized assessments. However, people seek out rehabilitation services because they want to be able to perform better in their daily lives (15), and improvements in UL capacity seen in the clinic do not necessarily translate to improvements in UL performance in daily life (13, 16–19). Therefore, assessment of UL activity performance in an individual's unstructured, free-living environment is critical to evaluating effectiveness of rehabilitation services and determining if the services provided have achieved the goal of improving performance in daily life.

Data extracted from bilateral, wrist-worn accelerometers can be used to quantify variables measuring different aspects of UL performance in one or both limbs. These variables collectively inform clinician scientists about the real-world activity performance. The numerous variables calculated from accelerometers measure different aspects of UL performance, such as: (1) duration (7, 20); (2) magnitude (12, 21, 22); (3) variability (12, 23); (4) symmetry or laterality (3, 7, 9); and (5) quality of movement (6, 24–26). Each UL performance variable conveys slightly different information about the collective nature of UL use, with a single variable providing only part of the picture (6). Furthermore, some variables are narrowly distributed in neurologically-intact (adult controls) individuals (e.g., use ratio, an index of duration of activity of one limb vs. the other), while other variables are widely distributed (e.g., bilateral magnitude, a measure of magnitude of bilateral UL activity) (3). Thus, multiple variables quantifying different aspects of movement along with heterogeneous distributions of those variables can make it difficult to interpret UL performance data for clinical decision-making.

One reason wearable sensor technology (e.g., accelerometry) for measurement of UL performance has remained largely confined to rehabilitation research with limited ventures into clinical practice is because the current output from accelerometers is not easily accessible for rehabilitation professionals (4). A potential solution to the multi-variable problem would be the formation of categories (or groups) of UL performance in daily life. If there were natural groupings that occur among multiple UL performance variables calculated from accelerometry data (27), then these groupings could help to facilitate clinical decision making and implementation of UL performance data into routine rehabilitation care. In other biomedical science fields, formation of categories which encompass multi-dimensional measures have facilitated clinical decision making for persons with health conditions such as, spinal cord injury (28), heart failure (29, 30), and chronic obstructive pulmonary disease (31).

The purpose of this study, therefore, was to identify categories of UL performance in daily life in adults with and without stroke using data from previously collected cohorts. Cluster analyses were performed with variables of UL performance calculated from 24 h accelerometer recordings from three cohorts, two samples of persons with stroke and one from neurologically-intact adult controls. We hypothesized that at least three categories (low, medium, and high) of UL performance would be identified across the UL performance variables quantified by accelerometer data, spanning the possible ranges of UL performance in daily life. We also anticipated that the emerging categories would group individuals with similar ranges of the performance variables and provide a simpler method to interpret UL performance in daily life for clinicians and persons with health conditions whom they treat.

This study analyzed accelerometer data from adults from three previous cohorts, using the same accelerometry methodology (32). Data used in these analyses were UL performance variables calculated from accelerometer data over 1 day, associated clinical capacity measures, and participant characteristics.

A sample of 211 participants were included in the analyses. Demographic and participant characteristics for the three cohorts are provided in Table 2. UL capacity was measured by the ARAT and indicated that both stroke cohorts were and had moderate deficits in UL functional capacity.

The Hopkins statistic was H = 0.78, indicating that clusters exist in the sample. Table 3 summarizes the range of solutions evaluated including 12, 9, 7, and 5 UL performance variables in either a 3-, 4-, or 5-cluster solutions. Across the different numbers of input variables, two principal components explained the majority of the variance, PC1 and PC2. There were similar loadings of the input variables onto these principal components, regardless of the number of variables entered. Interestingly, adding more performance variables (e.g., 12 vs. 5) was associated with both PC1 and PC2 explaining less of the total variance (see the first column of Table 3). Thus, across different numbers of input dimensions, the number of principal components was relatively stable. PC1 and PC2 appeared to be explaining similar variance in all models. We therefore proceeded with including only 5 input variables. When including 5 UL performance variables, the first principal component (PC1) explained the most variance (76.4%) and was comprised of variables that all had moderate to strong, positive loadings, including; paretic/non-dominant hours, median acceleration of paretic/non-dominant limb activity, acceleration variability of paretic/non-dominant limb activity and the use ratio. The second principal component (PC2) explained less variance (17.6%) and was comprised of primarily the non-paretic/dominant hours, a single variable that had a strong, negative loading. See Supplementary Table 1 for the loadings of all factors of PC1 and PC2 for the final chosen solution.

Across the models with differing numbers of UL performance variables, a 5-cluster solution explained the most overall total variance when compared to a 3- or 4-cluster solution as seen in the middle portion of Table 3 and visually in Figures 1A,B (including 5 performance variables). We then examined several metrics to determine how many clusters were appropriate for the 5-variable solution. Figure 1A supports that there are ≥ at least two clusters in this dataset and the flattened slope on Figure 1A indicates that the reduction of within-cluster variance is minimal and there are no further improvements after 5-clusters for this dataset. We therefore explored a 3-, 4-, and 5- cluster solutions. Figure 1B displays the effect of increasing numbers of clusters on the total explained variance when including 5 UL performance variables and confirms that a 5-cluster solution explains more total variance than the 3- or 4-cluster solutions. Examining the AIC values seen in the last three columns of Table 3, also confirmed that a 5-cluster solution produced the best model fit compared to the 3- and 4-cluster solutions across the different number of input variables (5, 7, 9, and 12 UL performance variables). Although each solution was statistically feasible, the chosen final solution was 5-clusters, from 5 UL performance variables including: (1) hours of use of paretic/non-dominant limb; (2) hours of use of non-paretic/dominant limb; (3) median acceleration of paretic/non-dominant limb; (4) acceleration variability of paretic/non-dominant limb activity; and (5) use ratio. Figure 1C presents the location of the 5-clusters across the two dimensional space. Dimension 1 (x-axis) is the first principal component and dimension 2 (y-axis) is the second principal component. The two clusters with the lowest overall UL performance are represented by clusters numbered 1 and 2 with the highest in number 5. Figure 2 shows a scatterplot matrix of how the 5 input variables relate to each other and to the 5-clusters.

The means and ranges of each UL performance variable, percentage concordant, and UL capacity for each of the 5-clusters in the final solution are presented in Table 4. The clusters are presented with the “lowest” overall UL performance within the first column and the “highest” overall UL performance in the last column. The 5-clusters become categories of UL performance and are named based on a synthesis of information from other publications that have described UL performance in daily life (45, 56, 57) not on the underlying PCA dimensions. The cluster names were chosen as intuitively as possible and represent the overall amount of UL activity and integration of the ULs into daily life activities (see Discussion for further interpretation). We refer to these clusters/categories as: (1) Minimal Activity/Rare Integration; (2) Minimal Activity/Limited Integration; (3) Moderate Activity/Moderate Integration; (4) Moderate Activity/Full Integration; and (5) High Activity/Full Integration. The cluster with the lowest UL performance is the Minimal Activity/Rare Integration, this cluster has the lowest mean values on variables that quantify duration, magnitude and variability of UL activity. People in this cluster use their non-paretic UL ~2.5 times more than their paretic UL and have little to no magnitude or variability of their paretic UL activity in daily life. People in the Minimal Activity/Limited Integration cluster use both the paretic and non-paretic limb for more overall hours than the Minimal Activity/Rare Integration cluster, but the non-paretic limb is still active twice as much as the paretic UL. Additionally, people in this cluster have slightly higher mean values on performance variables that quantify both the magnitude and variability of the paretic limb when compared to the Minimal Activity/Rare Integration cluster. Both of these clusters have little integration of the ULs into activity, as suggested by a mean use ratio below 0.50 the Minimal Activity/Rare Integration cluster and a mean use ratio just above 0.50 in the Minimal Activity/Limited Integration cluster. The cluster with overall, moderate UL performance is the Moderate Activity/Moderate Integration cluster. In this cluster, people have more symmetrical UL use compared to the two lower clusters, which is reflected in the in the use ratio (0.85) and the mean values of both duration variables (4.5 vs. 5.3 h). People in this cluster have moderate values on variables that quantify both the magnitude and variability of paretic/non-dominant limb activity. The two clusters with the highest overall UL performance are the Moderate Activity/Full Integration and the High Activity/Full Integration clusters. These clusters have progressively higher mean values of variables quantifying duration, magnitude and variability of UL activity with those in the High Activity/Full Integration cluster having the highest mean values compared to the other clusters. Both of these clusters however, have similar mean values of the use ratio, which is approaching 1.0 indicating that people in these two clusters have relatively equal contributions of both ULs. Interestingly, if only the use ratio was used to examine these two clusters it could be assumed that they are relatively equal, but the other variables show they are not. The two clusters with the highest overall UL performance also had the highest % of people with concordant stroke. It is also noteworthy that participants within each of the 5 clusters have wide, overlapping ranges of UL capacity, as indicated by the mean and ranges of ARAT scores in the bottom row of Table 4, consistent with the premise that UL capacity and UL performance are different, but related constructs. Figure 3 presents how the three included cohorts separated into the 5-clusters. The two clusters with the lowest overall UL performance (Minimal Activity/Rare Integration and Minimal Activity/Limited Integration) are comprised of only persons from the stroke cohorts. The cluster with moderate UL performance (Moderate Activity/Moderate Integration) contains mostly people with stroke but there are also a few neurologically intact adult controls in this cluster too. The two clusters with the highest overall UL performance (Moderate Activity/Full Integration and High Activity/Full Integration) contains the neurologically intact adult controls and some persons with stroke. Finally, there was a statistically significant omnibus effect of cluster in each of the univariate ANOVAs for the 5 UL performance variables (p-values for each variable < 0.001). Note that not all clusters were statistically different from all other clusters in each variable, based on post-hoc t-tests. However, this speaks to the multivariate nature of the cluster analyses; across all dimensions, these clusters group similar to observations together, but along any single dimension there will likely be overlap in the neghiboring clusters.

Figure 4 presents the group and individual coxcomb charts for each of the 5-clusters. The rows (A, B, C, D, E) are presented in order of increasing overall UL performance, with group data in the first column in dark gray and then individual examples of people in that cluster in columns two and three (A-E, numbered 2 and 3 respectively) in light gray. Each of the 5 standardized UL performance variables are represented by wedges within the plot, and the area of the wedge reflects the standardized value on that single variable. Figures 4A,B present the two clusters with the lowest overall UL performance (Minimal Activity/Rare Integration and Minimal Activity/Limited Integration), the wedges in these two clusters are small with the exception of the non-paretic/D hours of use, indicating that people in these two clusters use their non-paretic UL out of proportion to their paretic UL. As you move down each row from Minimal Activity/Rare Integration (Figure 4A) to High Activity/Full Integration (Figure 4E) one can see that the wedges get larger and begin to fill more area of the radial plot, however some variables are still out of proportion to the others as seen in Figures 4C,D. By the final group plot in Figure 4E1, the wedges for each variable span the largest area and almost form a perfect circle, compared to the clusters with lower UL performance (Figures 4A,B), indicating people in this cluster have the highest values across all 5 performance variables.

In a large sample of persons with and without neurological UL deficits, we used a k-means cluster analysis with multiple UL performance variables, captured via accelerometry, to derive a 5-cluster categorization that included 5 UL performance variables. Two principal components explain most of the variance in the input variables and 5-clusters explained the most total variance and had the best model fit. In this 5-cluster solution, two groups with what might be considered “normal” UL performance (Moderate Activity/Full Integration and High Activity/Full Integration) emerged, as indicated by the presence of many neurologically intact adult controls in those categories. One category in the middle had moderate UL performance (Moderate Activity/Moderate Integration), while two categories had low, overall UL performance (Minimal Activity/Rare Integration and Minimal Activity/Limited Integration). The names of each of the 5 categories were chosen for their overall UL activity and integration, with the future goal that these categories could be evaluated for their application to other clinical populations, not just persons with stroke.

The 5-category solution from 5 UL performance variables, derived from this statistical analysis, leads to a clinically-logical interpretation of UL performance in daily life. In this analysis we purposefully included three cohorts of persons with and without stroke in order to capture a wide range of the variables, extracted from accelerometer data, that quantify different aspects of UL performance in daily life. In Figure 3, the two categories with the highest overall UL performance (Moderate Activity/Full Integration and High Activity/Full Integration) contain most of the neurologically intact adult controls indicating that people without neurological impairments display a wide range of UL activity that can be considered unimpaired or normal. This is important because these people have integrated their ULs, as indicated by the use ratio variable, but people in these categories have different levels of overall UL activity, ranging from moderate to high UL activity. This is not unusual when we consider the wide range of activities and behaviors of people (58–60). For example, when walking performance is quantified by pedometers, neurologically-intact adults walk symmetrically but present with a wide range of variability in the total number of steps-per-day that can all be considered “normal” walking performance (58, 61–65). Based on the current results, it appears that people without neurological UL impairments similarly display a wide range of UL activity that can also be considered unimpaired or normal. For example, two neurologically intact older adults may have very different activities of daily living and leisure activities (e.g., swimming vs. knitting) but would both be considered to have “normal” UL performance. In other efforts to categorize UL activity, some groups have found four categories (54, 55, 57), and others have found six (56). These analyses however tended to examine only the separation of UL activity of persons with stroke. In this analysis, the goal was not to form categories to differentiate between those who had a stroke and those who did not. Instead, the goal was to categorize people based on their overall UL use in daily life. In the 5-category solution here, we see that the two categories with the lowest UL activity and integration are comprised of only persons with stroke, but there are also people with stroke in the three categories with the highest overall UL performance too. This is a positive finding, showing that some people with stroke use their ULs similarly to neurologically intact adults. Persons with stroke who ended up in the two categories with the highest overall UL performance have likely experienced either full recovery of their ULs following their stroke, or have figured out how to use the wide range of capacity that they have to integrate their paretic limb and be active in daily life (19). An example of this is shown in Figure 4E2 which is an individual from stroke cohort 1 who ended up in the High Activity/Full Integration category.

Categories of UL performance have tremendous research and clinical potential. Within other biomedical science fields, formation of categories which encompass multi-dimensional measures have facilitated clinical decision making for persons with health conditions (see section Introduction). Specific to rehabilitation, categories of ambulation (based on the capacity measure of walking speed) have been validated, shown to be sensitive to change (66–68), used to set goals in clinical practice, and have been used as a primary outcome in a Phase III clinical trial (69). In that trial, the primary outcome was the percentage of people who changed (leaped) to a higher ambulation category after the intervention. The identified categories of UL performance that emerged in this analysis could be useful for future trials of persons with UL impairments following subsequent, future validation studies. Categories that emerged in this analysis have stratified participants into groupings with similar overall UL performance, representing a profile of arm activity in daily life (38, 55, 70). Individuals within each category have similar ranges of each performance variable included (e.g., duration, magnitude, variability and symmetry) that formed the 5-clusters. Interestingly, in this analysis people with stroke within each of the five clusters display a wide range of UL capacity across the clusters. Additionally, more people in the two clusters with highest overall UL performance have concordant stroke compared to the three clusters with lower UL performance. These findings are consistent with prior work indicating that people with concordant stroke (dominant limb = paretic limb) tend to have differences in the patterns of UL use (56, 71) and experience better recovery (19). One can envision that these categories could be used in future trials to analyze smaller subsets of individuals based on their UL category and to better understand how UL performance variables quantify change during rehabilitation therapy.

From a clinical perspective, the categories that emerged offer the future opportunity to transition measurement of UL performance in daily life for persons receiving UL rehabilitation away from the current confines of rehabilitation research labs, and into standard of care (4, 72). The results of this analysis are a first step in simplifying measurement of UL performance in daily life by exploring the underlying structure in the set of observed variables (73). A future option could be to offer a user-friendly, software package to rehabilitation clinicians that would calculate the 5 UL performance variables included in this analysis from data extracted from bilateral wrist-worn accelerometers. Based on a person's values across the variables, a category of UL performance could be determined and used to communicate current UL performance and used to set goals for future UL performance. Based on the aspects of movement (duration, magnitude, variability, symmetry) selected to form the categories, it is possible that these categories could be highly relevant for many clinical conditions affecting UL performance in daily life, not just those with stroke. Just as with mobility, there are plenty of biological and psychological reasons why people could have limited UL performance in daily life (58, 74, 75). Thus, the names selected for each category might be applicable to other clinical populations that have similar or different UL impairments and capacity limitations, beyond the typical asymmetrical deficit which is a major aspect of stroke UL movement (3).

Publicly available datasets were analyzed in this study. This data can be found here: https://simtk.org/projects/referentaccdata.

The studies involving human participants were reviewed and approved by Human Research Protection Office, Washington University. The patients/participants provided their written informed consent to participate in this study.

CL: financial. JB, JK, KL, MB, and CL: manuscript preparation. JB, JK, KL, and CL: data analysis. JB and MB: data collection. JB and CL: study design. All authors contributed to the article and approved the submitted version.

This work was funded by the US National Institutes of Health R01HD068290, T32HD007434, and TL1TR002344.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.