Research by José del R. Millán, PhD, and colleagues, which was published online September 10 in Nature.com’s Scientific Reports, may lead to a new generation of self-learning, easy-to-use neuroprosthetic devices. Current neuroprostheses, also called brain-machine interfaces (BMIs), emit error signals when a movement fails. Researchers used those error signals to develop technology that makes the BMI capable of learning the correct movements, rather than allowing the prosthesis to repeat the error.
Image courtesy of the study authors and Nature.com’s Scientific Reports.
Most BMIs operate by interpreting variations in the brain’s electrical activity through an EEG. To be effective, the method requires significant training on the patient’s side; patients must communicate the information (e.g., extend left arm) by modulating their brain activity. Yet despite the lengthy training that is required, patients are unable to perform some complex movements.
When an error is made, the brain emits an electrical signal signifying the failure of the action. This signal is called error-related potential (ErrP), “potential elicited when actions do not match users’ expectations,” according the study.
“With ErrPs, it is the machine itself that will learn to make the right movements,” said Millán, a researcher at the École Polytechnique Fédérale de Lausanne (EPFL), Switzerland, and holder of the Defitech Chair in BMI. He called the innovation a “paradigm shift” that the authors expect “to become a key component of any neuroprosthesis.”
The detection of the error signal allowed Millán’s team to create neuroprostheses that are capable of learning the proper movements based on ErrP. For example, if a neuroprosthesis user should fail to grasp a glass of water, the neuroprosthesis will understand that the action was unsuccessful, and the next movements will change accordingly until the desired result is achieved. The machine knows that the goal is reached when the actions performed no longer generate an ErrP, relieving the user from learning to control the prosthesis. This approach could be the source of a new generation of intelligent prostheses, able to learn a wide range of movements. With the support of the ErrP decoder, it may be possible to learn and master a multitude of motor movements, even the most complex ones.
The 12 subjects in the experiment were first asked to train their prosthesis to detect ErrP. Equipped with an electrode headset, they observed the machine, programmed to fail in 20 percent of cases, performing 350 separate movements. The setting of the ErrP detector lasted 25 minutes, on average. Once this step was completed, the subjects performed three experiments to evaluate the effectiveness of this new approach. In the last one, they were asked to identify a specific target using a robotic arm placed two meters away. In all three experiments, the neuroprosthesis demonstrated learning capabilities by continuously adapting its actions and enhancing its precision.
“According to our expectations, this new approach will become a key element of the next generation brain-machine interfaces that mimic the natural motor control. The prosthesis can function even if it does not have clear information about the target,” said Millán.
Editor’s note: This story was adapted from materials provided by the EPFL.
To read more about Millán’s research, visit “Intelligent Neuroprostheses Mimic Natural Motor Control.”
