Sensor-based categorization of upper limb performance in daily life of persons with and without neurological upper limb deficits

Abstract

Background: The use of wearable sensor technology (e.g., accelerometers) for tracking human physical activity have allowed for measurement of actual activity performance of the upper limb (UL) in daily life. Data extracted from accelerometers can be used to quantify multiple variables measuring different aspects of UL performance in one or both limbs. A limitation is that several variables are needed to understand the complexity of UL performance in daily life.

Purpose: To identify categories of UL performance in daily life in adults with and without neurological UL deficits.

Methods: This study analyzed data extracted from bimanual, wrist-worn triaxial accelerometers from adults from three previous cohorts (N=211), two samples of persons with stroke and one sample from neurologically intact adult controls. Data used in these analyses were UL performance variables calculated from accelerometer data, associated clinical measures, and participant characteristics. A total of twelve cluster solutions (3-, 4- or 5-clusters based with 12, 9, 7, or 5 input variables) were calculated to systematically evaluate the most parsimonious solution. Quality metrics and principal component analysis of each solution were calculated to arrive at a locally-optimal solution with respect to number of input variables and number of clusters.

Results: Across different numbers of input variables, two principal components consistently explained the most variance. Across the models with differing numbers of UL input performance variables, a 5-cluster solution explained the most overall total variance (79%) and had the best model-fit.

Conclusion: The present study identified 5 categories of UL performance formed from 5 UL performance variables in cohorts with and without neurological UL deficits. Further validation of both the number of UL performance variables and categories will be required on a larger, more heterogeneous sample. Following validation, these categories may be used as outcomes in UL stroke research and implemented into rehabilitation clinical practice.

Keywords: Accelerometry; Cluster Analysis; Outcome Assessment; Rehabilitation; Stroke; Upper Exremity.

Figure 1

(A) Scree plot representing how the within-cluster variance changes as increasing numbers of clusters are formed with 5 UL performance variables. (B) Line plot representing how the total explained variance changes with increasing numbers of clusters on dataset including 5 UL performance variables. The dashed lines represent the total variance explained for a 3- (blue), 4- (red), or 5- (green) cluster solution. (C) Visual representation of the 5-clusters with 5 UL performance variables across dimension 1 (x-axis) and dimension 2 (y-axis). The cluster number is presented in the location of the centroid of each cluster. The shape of the point within the cluster represents the if a participant was from a stroke (triangle) or control (+ sign) cohort.

Figure 2

Scatterplot matrix of the 5 input variables as a function of the 5 different clusters. The diagonal shows density plots (i.e., the univariate distribution) of each input variable as a function of the different clusters. The lower left panels show the bivariate distributions for each pair of variables with the point shapes and gray scales corresponding to the different clusters (see legend). The upper right panels show the Spearman rank order correlations for each pair of variables (on the whole, ignoring clusters). ^***p < 0.001.

Figure 3

Bar plot of the counts of participants from each of the 3 cohorts that separated into the 5-clusters. The two clusters with the lowerst overall UL performance are comprised of persons from the stroke cohorts only. The cluster with moderate UL performance contains primarily persons with stroke and a few neurologically intact adult controls. The two clusters with the highest overall UL performance include primarily neurologically intact adult controls, as well as persons with stroke.

Figure 4

Coxcomb charts of the five clusters, illustrating the contribution of the UL performance variables on a standardized scale. The first column plots group data, while the 2nd and 3rd columns plot individual participant examples. (A) Minimal Activity/Rare Integration cluster; (A1) group chart of people within this cluster; (A2) is a person from stroke cohort 1, ARAT = 4; and (A3) is a person from stroke cohort 2, ARAT = 10. (B) Minimal Activity/Limited Integration cluster; (B1) group chart of people within this cluster; (B2) a person from stroke cohort 2, ARAT = 10; and (B3) a person from stroke cohort 1, ARAT = 6. (C) Moderate Activity/Moderate Integration cluster; (C1) group chart of people within this cluster; (C2) a person from stroke cohort 1, ARAT = 36; and (C3) a person from the adult controls. (D) Moderate Activity/Full Integration cluster; (D1) group plot for this cluster; (D2) a person from stroke cohort 2, ARAT = 42; and (D3) a person from the adult controls. (E) High Activity/Full Integration cluster; (E1) group chart of people within this cluster; (E2) a person from stroke cohort 1, ARAT = 55; and (E3) a person from the adult controls.