In this project, I aim to utilize dimensionality-reduction techniques, such as Principal Component Analysis, for the purpose of reducing complex hand kinematics collected during therapy tasks and assessments and providing therapists with novel ways to evaluate hand function during rehabilitation.
Hand Complexity
Human hands are complex! They have 27 degrees of freedom if we include the wrist, 27 bones, 34 muscles, and over 100 ligaments and tendons, which makes it for a difficult task for engineers to replicate (e.g., prostheses) and for therapists to track (e.g., during therapy).
When it comes to hand therapy, whether it is for an individual who recently had stroke or for someone who had a hand surgery, the task of the therapist is to assess the function and assist in improving it - both holistically and atomistically. This is because while the hand contains fingers, each of which can act independently and are important individual units, ultimately, they act as a whole, forming gestures or grasping objects. As a result, when providing hand therapy, medical professionals must consider all degrees of freedom simultaneously - a cumbersome task!
Dimensionality Reduction to the Rescue
Dimensionality reduction (DR) techniques have been around for many decades and served the machine learning community of engineers and mathematicians well and true. Recently, they have been employed in the biomechanics communities to help understand the underlying mechanisms of control of complex biological systems... including hands.
In my past work, I have already utilized several DR techniques in assessing how well they can compress complex high-dimensional hand kinematics (joint angles) data into low-dimensional (2D) space. More on it can be read here and here. However, in that work, DR was utilized for the purpose of simplifying complex prosthetic hand control.
In my current work, I am trying to understand whether we can utilize these DR techniques to assess hand function during rehabilitation tasks in a lower-dimensional (2D once again) space. The idea is that instead of keeping track of all 27 degrees of freedom, the therapist can evaluate hand function simply by looking at a 2D plot. Wouldn't that be neat?
Endless Possibilities
From there, the possibilities are practically endless. If a given DR technique can in fact compress high-dimensional joint angles data into a lower-dimensional reliably and repeatedly within the same user and across different users, the applications for such a simple low-tech tool are many:
Assess hand improvement over a single (or many) therapy sessions for a single individual
Point to the specific part of the task or a particular hand feature that is "deviated" from the desired norm
Compare the same task performance between individuals
Early identification of problems with hand function (such as in rheumatoid arthritis)
Provide biofeedback to therapy participants about their progress during rehabilitation in a format that is easier to understand and visualize
Where Are We At?
We are currently collected pilot data with individuals without upper-limb disabilities using a sensorized glove. This is to understand which tasks are best for evaluating hand function.
Next, we will be collecting hand kinematics data in stroke individuals. Parallel to that, we are assessing several DR techniques in their applicability for the task of repeatability and stability across outputs.
In the future, there is room for utilizing simple camera-based hand-tracking solutions to perform data collection in rehabilitation settings in lieu of the complex and expensive sensorized data gloves. The current issues with such technologies include object occlusion and model mismatch that oftentimes lead to poor tracking quality.
コメント