Date Presented 03/27/20
Current assessments of activity poststroke occur in clinical settings. Researchers have attempted to use wearable activity monitors to assess activity in natural environments, but they are unable to distinguish between activities. We have developed an algorithm within a depth sensor that is able to distinguish between activities that occur within a kitchen setting. The purpose of this study was to test the system with individuals poststroke in an unstructured home environment.
Primary Author and Speaker: Rachel Proffitt
Contributing Authors: Mengxuan Ma, Marjorie Skubic
PURPOSE: Current assessments of activity and upper extremity function post-stroke are usually completed in clinical settings or rely on client report. Technologies, such as wearable activity monitors (e.g., FitBit-like devices), have been touted as options for assessing activity in natural environments. However, they fall short in being able to distinguish between various daily activities (Bailey, Klaesner, & Lange, 2015). We have developed an algorithm within an ambient depth sensor that is able to distinguish between 8 different activites that occur within a kitchen setting (e.g., sweeping, walking, reaching forward to open the fridge door) (Ma, Meyer, Lin, Proffitt & Skubic, 2018). The purpose of this study was to test an initial algorithm for recognizing and classifying daily activities of individuals with and without stroke in an unstructured home environment using a depth sensor.
DESIGN AND METHOD: This study was a developmental, iterative design, completed in two phases. In Phase 1, the “Daily Activity Recognition and Assessment System” was trained in a simulated kitchen with 11 individuals (age 18-40) without any disabilities. Each participant completed 3 sets of prescribed activities. The VicoVR depth sensor was placed in an unobtrusive location and recorded depth images (30 frames per second) and skeletal joint data. Over the course of the 11 participants, the DARAS algorithm was trained and refined. In Phase 2, the trained algorithm was tested on 5 participants data collected from natural home kitchen environments. Three individuals without disabilities completed meal preparation tasks and all data were recorded by the VicoVR sensor. Two individuals post-stroke completed normal daily activity in their kitchen and all data were recorded by the Foresite sensor (similar depth sensor). All data (depth images and skeletal data) from Phase I were used to trained the Convolutional-deconvolutional neural network (CNN). The trained model was tested on the data from Phase 2. Per-frame accuracies and per-action accuracies were calculated for the 5 individuals from Phase 2.
RESULTS: For the mock kitchen in Phase 1, the per-frame accuracy was 85.1% and the per-action accuracy was 92.1%. In Phase 2, the per-frame accuracies ranged from 52.0% to 89.7% and the per-action accuracies ranged from 54.3% to 88.5%.
CONCLUSION: The per-action and per-frame accuracies indicate that the initial training of the algorithm was acceptable (above 85%). However, the algorithm and model requires testing and training with a larger, more diverse sample. Additionally, we have only tested on two individuals post-stroke. The next step in this research is to test on a larger sample in different kitchens. Once the algorithm reaches 85% confidence, we will be able to calculate kinematic variables that assess movement quality during the various recognized actions.
IMPACT STATEMENT: This study is the first of its kind to develop an unobtrusive assessment of daily activity. It will add a new, objective measure of occupational performance and participation to the occupational therapy measurement toolbox.
References
Bailey, R. R., Klaesner, J. W., & Lang, C. E. (2015). Quantifying real-world upper-limb activity in nondisabled adults and adults with chronic stroke. Neurorehabilitation and Neural Repair, 29(10), 969-978.
Ma, M., Meyer, B. J., Lin, L., Proffitt, R., & Skubic, M. (2018). VicoVR-based wireless daily activity recognition and assessment system for stroke rehabilitation. Proceedings of the 2018 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 1117-1121.
