Abstract:
Movement quality assessment has been integral to appraise physical therapy in upper limb rehabilitation after stroke. The outlook is attractive because it provides a non-invasive insight to movement quality for the purpose of benchmarking the level of impairment before the rehabilitation begins and measuring the extent of recovery afterwards. However, a clinically acceptable measurement technique has been limited to laboratory setup with complex subject preparation. The attachment of numerous retro-reflective markers is time consuming and requires careful palpation on bare skin to record movements. The recorded marker positions need to be manually labeled in post-processing to interpret the movements according to accepted clinical model. This laboratory setup would require a highly skilled bio-mechanics' scientist and therapist to locate the anatomical landmarks and to analyze the kinematics data. Furthermore, the cost and complexity of the setup limits the amount of assessment replication. Recently, off-the-shelf marker-less motion capture becomes available and provides plausible joint estimation in form of skeletal data in real-time. This feature is particularly attractive especially for upper limb assessment due to multiple joints required to be monitored over time. Moreover, it trades off cost, complexity and accuracy of the assessment. Multiple researches have reported acceptable accuracy in joint angles estimation in various task similar to the existing stroke assessment task. These reports had inspired this research towards automating the kinematic disability assessment for stroke patients. Firstly, an extensive literature review was conducted in the beginning of this research to discover the gaps in current kinematic assessment. It revealed the importance of determining the extent of compensation in stroke patients which was overlooked when assessing end-point movement. Particularly, reaching and drawing assessment tasks which were ubiquitous in evaluating robotic rehabilitation outcomes lack the compensatory movement measure. Kinematic result of patient's hand movement can be misleading if the extent of joint coordination and torso compensation are not taken into account. While previous studies in marker based setting have provided a number of parameters to determine the extent of compensation during assessment, such similar work in marker-less setting was non-existent. Therefore, this research investigated the use of marker-less motion capture to determine the extent of compensation in existing clinically-accepted assessment tasks to provide further insight to the outcomes of the assessment. The excessive movement of the torso when performing assessment tasks that require arm-forearm coordination is typical in stroke patients. Therefore, a three dimensional measurement model which explains the adaptation of this compensatory strategy is essential to determine the extent of motor recovery. A Torso Principal Component Analysis (PCA) Frame model was developed utilizing Kinect's joint prediction as a proposal to assess torso orientation over time. By re-orientating and aligning the axes to the clinically accepted torso orientation model, all the independent torso angles can be decomposed and reported as parameters to represent compensatory behavior. The Torso PCA Frame model was first evaluated by parametrizing its distribution in the assessment session as attributes to predict normal and compensatory behavior in artificial stroke movement setting. Healthy participants were fitted with elbow brace to limit arm-forearm coordination which may artificially induce compensatory torso movements to complete the task.They performed gross movements typical in stroke assessment and their torso distribution over the session were recorded. Results show that the accuracy of the Torso PCA Frame model was at 98.7% and were suitable as parameters to delineate compensation in that setting. To perform comparison with clinically accepted data, the Torso PCA Frame model was then evaluated by comparing the torso angle with marker-based clinical model and Kinect's intrinsic chest orientation to assess the torso movement. Healthy participants were recruited to perform circle tracing (CT) and point-to-point (PTP) planar tasks in simultaneous setting of marker-based and marker-less system. Results showed that the torso angles computed using Torso PCA Frame model were insignificantly different to clinical measures in PTP task (0.103±0.881° in forward bending, 1.631±1.456° in lateral flexion and −3.488±2.765° in axial rotation) but forward bending was significantly different in CT task (3.700±0.473°). Extended evaluation also shows that the mean of axial rotation angles were significantly similar across both tasks (F2,18 =1.800, p=.194 in PTP task and F2,18 =1.876, p=.182 ) in marker-based and marker-less setting. Torso PCA frame model was evaluated afterwards against healthy participants which were fitted with elbow brace and strapped across the chest to emulate the limited coordination of stroke patients. Five participants were randomly chosen to emulate this behavior and the results showed that the forward bending angles were significantly different between normal and artificial stroke participants in PTP task (−7.532±4.171° ,p=.001) but not in CT task (−.261±4.172° , p=.899). To investigate the usability of Torso PCA Frame model to detect torso compensation, the Torso PCA model was evaluated in stroke participants to assess their movements performing circle and point-to-point tracing. Results showed that forward bending and lateral flexion were significantly different between normal and stroke patients in both tasks (−12.130± 4.211° , p<.0005 and11.008±9.468° , p=.024 respectively), while only forward bending was significantly different in CT task (−4.770±4.221° , p=.028). These results were different from the outcome in artificial stroke movement setting. Nevertheless, the artificial stroke movement setting was accurate to identify excessive forward bending in PTP task and was eminent to show that the model was able to enunciate typical compensatory behavior in stroke patients. To enunciate genuine motor recovery on the current end-point movement quality assessment, measure of compensatory movement is essential. By aggregating existing kinematic parameters derived from the result of the review with the proposed torso compensatory assessment, a new kinematic assessment scheme which represent the four underlying problems of stroke patient's movement was proposed. Along with the independent torso angles computed, the kinematic parameters outline not only the extent of motor recovery but also the extent of compensatory movements. To evaluate the effectiveness of the new kinematic assessment scheme, the pool of kinematic parameters derived from the result of the review along with the torso angles obtained from the Torso PCA Frame model were used to form kinematic attributes in supervised learning algorithm. Results showed that the accuracy of the aggregated model was at 93.33% and were suitable as parameters to delineate normal participant from stroke patients. In summary, the development of Torso PCA frame model can be valuable for automated setting as the setup was simple and the results obtained were comparable to clinically computed angles within the planar assessment setting ubiquitously implemented in automated rehabilitation. This research has collaborated with Laura Ferguson Rehabilitation Center, Auckland for the assessment of stroke patients. The outcome of this research includes an extensive review of kinematic assessment in upper limb movement after stroke which was published in peer-reviewed SCI journal. The proposed torso model which is used as the basis to measure compensatory movement was also presented at an international conference.