Almost Half of All Prosthetic Users Give Up on Their Medical Devices. Machine Learning Can Change That.
The concept of prosthetic limbs has existed for thousands of years, with the earliest known prosthesis – a big toe – found in Egypt and dated to almost 1000 BCE.
More recent history has seen the rapid development of different prostheses, including robot-powered limbs outfitted with dozens of sensors that gather real-time data about positioning, angle, and force exerted.
But there’s a big problem with many prosthetic limbs, one that has likely existed ever since that first big toe was designed for an Egyptian noblewoman: People don’t like them.
Indeed, even in the age of technologically advanced prosthetic limbs, the rejection rate – or the rate at which patients stop using their prosthesis – stands at nearly 50 percent.
But that rejection rate drops by nearly half when patients use body-powered or electric prostheses. And many researchers now see machine learning (ML) as the perfect complement to these non-passive prostheses.
The promise of machine learning for better prosthetics
Prosthetics certainly are more precise and much stronger than they were a few decades ago and light-years more advanced. All those real-time sensors embedded in modern-day prostheses capture a massive amount of data.
Until now, however, they haven’t been able to make a huge difference in the quality of life for prosthetic wearers, with most amputees only having a very basic level of control. Many prosthetics are not intuitive to use and can be heavy to lug around.
Similarly, most conventional prosthetic devices have limited degree-of-freedom (DOF) movement and range. That means most of these limbs perform only one or two functions, and often not very well.
“The many muscles and tendons that would have controlled the fingers are gone, and with them the ability to sense exactly how the user wants to flex or extend their artificial digits,” writes AI and science reporter Devin Coldewey in TechCrunch. “If all the user can do is signal a generic ‘grip’ or ‘release,’ that loses a huge amount of what a hand is actually good for.”
That’s likely to change, however, as next-generation artificial limbs with multiple DOF are paired with ML models that automatically detect and learn the finer points of gripping and other detailed movements.
These models learn by observing muscle signals and other stimuli as a prosthesis user attempts various movements. Armed with this information, prosthetics – and the AI models that power them – learn to perform certain mundane yet vital actions automatically, similar to muscle memory in an actual hand or leg, ensuring more effective and far less one-dimensional grips and movements.
As researchers from the University of Alberta explain, these techniques can help share “the burden of control” between the user and their intelligent device, with prosthetic devices essentially “filling in the gaps” for the user.
Improving prosthetics through machine learning
Researchers have made recent notable advances to improve prosthetics using reinforcement learning and other ML techniques, at least in controlled experiments.
In 2017, Vasan et al. (the U of A researchers mentioned above) demonstrated an actor-critic reinforcement learning method for prosthetic limb training. The technique allowed users to train a 3-DOF prosthetic arm through coordinated movements and grips using their non-amputated arm.
Vasan et al.’s experiment showed that ML-powered prosthetics can be trained to perform “simultaneous gestures and movements” based only on data collected by the robot arm and above-elbow myoelectric signals from the user.
“These preliminary results,” the authors write, “also suggest that our approach may extend in a straightforward way to next-generation prostheses with precise finger and wrist control, such that these devices may someday allow users to perform fluid and intuitive movements like playing the piano, catching a ball, and comfortably shaking hands.”
Two years later, in 2019, University of North Carolina, North Carolina State University, and Arizona State University researchers developed a technique underpinned by reinforcement learning to tune robotic prosthetic knees.
The system collects data from the device while comparing the patient’s gait to a typical healthy gait profile in real-time, adjusting 12 control parameters throughout the gait cycle nearly instantaneously as the data is analyzed.
The technology allows patients to walk on a level surface within 10 minutes, a sea change from the half-day of human-led training typically required for robotic knee brace modifications.
More recent studies have gleaned even more promising results. Swami et al.’s 2021 paper shows how a 2-DOF prosthetic arm was trained using data collected from the activities of 10 individuals performing daily living (ADL) tasks using a wrist brace.
A neural network first classified each movement, with random forest regression computing movement velocity. The study’s classification achieved an F-1 score of 99 percent, while the regression scored +0.98 under the Pearson correlation measurement (with +1 being a total positive correlation).
CapeStart: Your medical machine learning partner
CapeStart has years of experience helping medical researchers apply off-the-shelf and custom machine learning algorithms and models to systematic reviews and real-world clinical applications. Contact us today to schedule a brief discovery call with one of our experts and learn what we can do for your next project.