The KINARM robot produces granular datasets of performance metrics associated with sensory, motor, visuospatial, and executive function. The primary objective of this study was to reduce multivariate data produced by the KINARM and generate interpretable and intuitively understandable components across 6 behavioral tasks to improve characterization of participant performance. KINARM data dimensionality was substantially reduced, while still retaining a large proportion of variance. Therefore, complex participant data produced by the KINARM robot can be reduced into a small number of components that characterize participant performance.
Interpretability and classification of components
We initially used PCA to reduce our multivariate dataset, and we then investigated if this statistical technique identified interpretable and classifiable components of participant performance. Within a task, components explained intuitive performance metrics. For example, in APM (a task that assesses limb proprioception), the 3 components identified were distinct measures of function (position accuracy, movement variability and contraction/expansion). Furthermore, KINARM parameters were separated reproducibly across multiple tasks. For example, OHA was divided into 5 components: impulsivity, accuracy, hand bias, speed and area, and miss bias. Three of these components (accuracy, hand bias, and speed and area) are comparable to the PCA results observed from the OH task, thereby providing evidence that these tasks assess similar underlying behaviors. However, despite assessing similar underlying behaviors, the variance explained by each component differs. For example, speed and area accounts for more variability than accuracy, whereas in OHA accuracy accounts for more variability than speed and area. The addition of two unique components (i.e., miss bias and impulsivity) during the OHA may quantitatively represent the increase in cognitive load during this task. These two components that are absent from OH, and which account for 28% of overall variability, likely reflect the different impacts of impulsive movements in OH and OHA. In OH, quick and effectively impulsive movements are rewarded; quick movements with little inhibition will result in many targets being hit. In contrast, in OHA there is a penalty for impulsive movements that are not processed thoroughly prior to execution (i.e., deciding whether or not an object is a target or distractor prior to executing a movement towards it). Furthermore, behaviorally during the OH task, the subject can hit a total of 300 targets during the trial. In contrast, there are only 200 targets during OHA, plus 100 distractors. Therefore, hand bias and speed and area account for less variability in OHA, as now the subjects must inhibit their automatic responses (i.e., impulsivity), which also results in an increased need for improved accuracy. This contrast in cognitive load between tasks may anecdotally account for the difference in variance explained by the components mentioned above across OH and OHA. This result of similar inter-task grouping of parameters across related KINARM tasks was also observed during VGR and RVGR. Taken together, our results demonstrate the potential utility of PCA when applied to granular datasets, as well as the increased interpretability of KINARM parameters when categorized into behaviorally meaningful components that characterize performance.
Furthermore, parameter associations, positive or negative, strong or weak, within their respective components were logically related to participant performance. For example, the PCA of the BonB task identified the component speed and success. Targets completed, mean ball speed, mean right- and left-hand speed, and hand speed difference all had substantial positive loadings, whereas mean movement time had a substantial negative loading. Therefore, as participants increased their right- and left-hand speed, and thus ball speed, participants decrease the time from a target being displayed on screen to the target being reached by the participant, which results in a greater number of targets being completed throughout the task. In addition, the parameter norm absolute hand speed difference loaded substantially on bar angle but had a very low loading on speed and area. Behaviorally, this result is intuitive. If a participant reacts to a target quickly, the bar angle may become skewed if not carefully balanced, which is reflected by the substantial positive loadings of the other angle related parameters onto this component. These direction-based associations, low or high, were observed across most KINARM tasks and their respective parameters. These components will need to be further validated across KINARM platforms and patient populations. Overall, our analysis produced a consistent inter-task classification of behavioural variables (e.g., VGR and RVGR, OH and OHA), which demonstrates that these tasks assess similar underlying behaviours and further supports the validity of using this statistical technique on KINARM performance metrics. Our current analysis has substantially reduced these granular datasets into biologically plausible, interpretable, and behaviorally meaningful components.
Cross-loading of parameters with multiple components
PCA indicated that some of the KINARM parameters cross-loaded with multiple components. For example, the APM task identified 3 components (position accuracy, movement variability, and contraction/expansion), which is fairly consistent with previous research that described these 3 variables as observed patterns of impairment after stroke [2]. Despite mostly a strong separation of these three components, shift x loaded onto both the position accuracy and movement variability components, with an inverse sign from negative to positive respectively. Shift indicates a systematic bias to move either left/right or front/back in the workspace. Therefore, in the position accuracy component, shift X (left/right) indicates, with a negative association, that shift X biases to the left as absolute error in the X and XY plane increases. In contrast, component 2, movement variability, indicates that as X shifts in the positive direction (i.e., hand moves to the right), overall variability increases.
Furthermore, certain parameters did not intuitively cross-load onto multiple components. For example, the OH task identified 3 components (hand bias, speed and area, and accuracy). Unexpectedly, the parameter total hits, which should be intuitively related to hits with left and right, did not cross-load onto the accuracy and hand bias components. However, this result demonstrates that total hits is highly related to accuracy, which is separable from hand bias. These components are not highly interrelated, as indicated by the use of orthogonal rotation, and measure separable aspects of participant performance. Therefore, hand bias measures the bias towards the participants limb preference, and use in space, which is separable from overall accuracy and the total targets being hit (e.g., right-handed participants may have biased use of their right hand, but this is separable from the number of targets that were accurately hit using either hand).
Interestingly, only three parameters did not substantially load onto any component: miss bias of OH and bar length variability and hand path bias from the BonB task. The reason for these low loadings is unclear. However, miss bias loaded highly onto a separate component for the OHA task, which may indicate that any bias of misses toward one side of the work space, or the other, is related to the increase in cognitive load during OHA. Furthermore, the low loadings for bar length variability and hand path bias may be the consequence of only analyzing level one of BonB, as this task increases with difficulty in the subsequent levels [19]. However, previous research has identified that level one of BonB identified the most performance impairments (i.e., highest number of parameters failed) in stroke participants, relative to control subjects [19]. Furthermore, the number of parameters failed tended to decrease with increasing task difficulty in successive levels, potentially due to the increased variability observed among controls, which ultimately influenced cut-off criteria used to quantify impairment among stroke participants [19]. Therefore, level one of BonB may be the most sensitive level used to detect impairment and is the most relevant for the current analysis. Future analysis may need to apply PCA to each level separately to investigate if the components are similar across all three levels. However, despite identifying some parameters that cross-load with multiple components and only a couple that did not substantially load onto any component, the majority of KINARM performance metrics did not cross-load with one another. Most components demonstrated high loadings with their respective components, which indicates a strong separation of components that quantify participant performance.
Limitations and future directions
We excluded left-hand dominant participants, as many participants were right-handed, which means our findings may not generalize to left-handed participants. Furthermore, we could not conduct PCA across all task parameters, as participants did not complete all KINARM tasks. Therefore, we were unable to conduct PCA on broadly pooled performance scores across tasks, which confined our results to descriptions of each task only. However, the PCA has yielded task specific separations of components to further characterize sensorimotor function among healthy subjects, which may serve as a normative data set for future clinical comparisons of subject performance. In addition, two to three participants per task, except for RVGR, were missing participant data for education obtained. However, these missing data should not substantially impair our range of education. In addition, some tasks do not have Z-scores for all parameters being recorded (these metrics could not be standardized), and thus these parameters were excluded from this analysis. We conducted PCA only on the End-Point robot data, and therefore, we will need to conduct PCA on healthy participant data generated using the KINARM Exoskeleton. It is unclear how antigravity support provided by the Exoskeleton will affect participant performance and the generated components. Therefore, PCA of Exoskeleton data will be imperative to future investigations to implement PCA across KINARM platforms. Also, the current analysis was conducted in healthy participants, and future PCA will need to examine participant performance using a clinical sample, such as stroke, to validate these components and characterize performance in clinical populations to further demonstrate the utility of our current findings. Furthermore, the current analysis does not address individual or group differences in our healthy participant sample, which may complicate future comparisons between healthy subjects and clinical populations. However, using PCA across 6 KINARM tasks, which assesses a broad range of upper limb sensorimotor function, has led to a substantial reduction in the dimensionality of our data, and produced interpretable components of performance.
Therefore, PCA of KINARM data has the potential to become a valuable clinical tool. Applying PCA and identifying main sources of variability in a clinical examination can help make data from research tools, such as the KINARM, more concise and easily interpretable. Furthermore, identifying sources of strong variability may improve the detection of fluctuations in performance, which could increase the clinical relevance of robotic assessment as an evaluation tool. Therefore, analyzing data produced by a clinical population offers the potential to increase the clinical utility of the KINARM robot by maximizing interpretability of participant performance, while also minimizing information loss, to increase the characterization of performance among various populations. It is not clear if patient performance will result in similar components of performance patterns as healthy participants. Therefore, future applications of this analysis may offer potential insight into specific patterns of sensorimotor impairment among patient populations.