A new research project will develop an intelligent spine interface, with the long-term aim of helping spinal injury patients regain limb function and bladder control.
The project, a collaboration between engineers and neuroscientists at Brown University, Intel, Rhode Island Hospital, and Micro-Leads Medical, has received $6.3 million in funding from DARPA.
As part of the study, patients with spinal injuries will have electrodes embedded in their spines, above and below the injury. An AI system running a biologically-inspired neural network will “listen” and learn about what the signals mean, with the aim of reconnecting the two parts of the spine electronically.
The project will record and analyse motor and sensory signals in the spine of patients with spinal injuries (Image: Intel)
The project will build on work already ongoing in the field of brain-machine interfaces to control external effectors. This includes the BrainGate program, which successfully interfaced with the brain to control a computer cursor and even a robotic limb, and other international research projects on brain-spine interfaces and spine stimulation.
David Borton, an assistant professor at Brown’s School of Engineering and researcher at the University’s Carney Institute for Brain Science, will lead the project.
“What’s new about this project is we actually want to start a conversation with the spinal cord,” Borton said. “We want to be able to not only stimulate it or talk to it, but also be able to listen to it and learn to extract signals that are useful from the spinal cord itself, and use those to drive spinal cord stimulation.”
The researchers will record signals from the area of the spine above the patient’s injury, then use machine learning to decode these signals, which are currently not fully understood, and work out how best to use them. The idea is then to apply these signals to the lower part of the spine with the hope of stimulating the correct response.
Electrical System
Brown and Intel are working with Rhode Island Hospital, building on the Hospital’s work in monitoring the brains of epilepsy patients. Surgeons at Rhode Island Hospital will implant a pair of electrode arrays either side of the patient’s injury, which is particularly difficult as the types of injuries patients have will all be different. The Hospital has built a new space especially for this program which includes the required rehabilitation equipment.
An example of an electrode array like the ones from Micro-Leads Medical that will be used in the project (Image: Brown University)
The physical implants will use a high-resolution spinal cord stimulation technology developed by Micro-Leads, called HD64. The first phase of the project will use 24-contact electrode arrays, moving to 64-contact arrays in the second phase. The contact sizes are in the order of 1 millimetre squared, and since a neuron is around 20 microns, each electrode will record or stimulate hundreds of thousands of neurons at a time. The signals to be recorded are electrical signals; as neurons communicate with each other, there is an electrical voltage change, and the electrode senses and records the change in electric field.
“That’s the exciting part of what we’re going to find out. Typically, there are different frequency bands in the signal that can represent different underlying neuronal processes. So that can be a clue for us as to what is actually going on,” said Hanlin Tang, principal engineer at Intel’s AI Products Group, himself a former neuroscientist and the Intel lead on the project. “But it is a lot of work on the machine learning side, to be able to interpret these signals well enough to know what to stimulate on the other side of the gap.”
Intel’s team will use its hardware and machine learning expertise to help build an AI system that interprets the signals.
“The key challenge here is that listening into the spine is not high fidelity,” Tang said. “It’s like trying to relay a message, but you can’t really hear one side and you can only mention a few words on the other side. Using machine learning, you might be able to use some prior knowledge to try to fill in the gaps and be a good interface to bridge this type of injury.”
The AI will also tackle mapping between the two electrode arrays, from one side of the injury site to the other, a crucial task.
Electrode arrays will be embedded in the patient’s spine, which can be used to record the signals sent from the brain (Image: Intel)
Borton explained that the nervous system is very plastic and can learn over time — “neurons that fire together, wire together” — meaning that recording from one part of the spine and stimulating another should allow the nervous system to learn what that particular signal means.
“We are not making an exact one-to-one mapping,” Borton said. “The interface we plan to develop will record from many hundreds of thousands of neurons and signals all superimposed on each other. And we’ll be stimulating a very sparse subset of point contacts, which will impact the activity of the thousands of different neurons, nonspecifically. The nervous system will hopefully learn to interpret that, as long as we get a good starting point.”
Neural Network
The Intel AI team will work with Thomas Serre, an associate professor of cognitive, linguistic and psychological sciences at Brown, who has expertise in developing biologically-inspired artificial neural networks. Serre’s recent work on neural networks based on how the visual cortex handles visual processing has shown that biologically-inspired architectures produce models which can be trained on less data and be more efficient.
Neural networks for the intelligent spine interface will be based on medical science’s understanding of the anatomical and functional architecture of the lower limbs, which can be modelled, to a certain degree, Borton said.
Training data is a key requirement for any neural network, but the intelligent spine project will have access to much less training data than a typical AI system, which is one of the challenges.
Will the AI require training for each individual patient?
“That’s one of the things we are hoping to find out,” Borton said. “The answer is, very likely, yes. Another open question is, if we do train it on one participant, how much retraining is needed and how deep, how many layers down do you actually have to retrain this model? That could be very interesting. It might even tell us something about what’s conserved across different lesions of the spinal cord over time, as we collect data from many more patients, that could lead to new diagnostic discoveries.”
Hardware and software
The Brown team will work with researchers from Intel, which will provide hardware, software and research support for the project.
Intel’s Hanlin Tang described how the first year of the project will be spent on neural network development. In the second year, the algorithms will be applied and Intel will begin to optimise them for the machine learning accelerators the company has in development, specifically, the Intel Nervana neural network processor line for training and inference. The software stack will be nGraph, a cross-platform software developed by Intel.
“What’s really exciting about this is the workloads aren’t entirely known. It’s a bit different to working with an enterprise customer where they hand you five workloads to optimise,” Tang said.
One of the biggest hardware and software challenges will be achieving real-time operation to restore locomotion and bladder control for patients.
“We need real time interpretation of all the channels and different frequency bands, then translating it, and learning how to stimulate the other side and bridge the gap,” he said.
The eventual aim is to use this research to develop the technology to a point where a small, implantable device helps patients with movement and bladder control during rehabilitation and beyond, and hopefully have a real impact on the lives of the many, many people living with spinal cord injuries.
Source: https://www.eetimes.com/document.asp?_mc=RSS%5FEET%5FEDT&doc_id=1335174&page_number=2
ytd2525 
Leave a comment