Scientists Are Using AI, Simulation Tech to Fix Our Brains
The human brain contains millions of nerve cells that communicate using electrical and chemical signals.
But one abnormal signal can disrupt other cells inside the brain, causing a person to react with a series of physical and mental impulses known as a seizure. That’s what happens inside the brains of the estimated 65 million people1 across the globe who suffer from epilepsy. In a non-epileptic brain, there is a balance of cells that excite signals and cells that stop signals. But in someone with epilepsy, there is an imbalance of signal activity, which can cause different types of seizures.
No two cases of epilepsy are the same, which makes it challenging to treat epilepsy with surgery — an option many of the 30-40 percent of patients2 with drug-resistant epilepsy will at some point consider. The most common is resective surgery3, which involves removing a portion of the brain to stop the seizures. While patients can be seizure-free after surgery, there are a number of risks, from memory loss to loss of motor skills.
But simulation technology — amplified by artificial intelligence — could be the key to one day treating this life-altering condition. The marriage of AI and innovative supercomputing models have brought scientists closer to effective simulation of the brain, and these tools are showing early signs of success in treating epilepsy.
Trading CPUs for GPUs to Map the Human Brain
The previous gold standard for simulation involved the processing power of central processing units (CPUs) strung together in supercomputers. But supercomputers are reaching their physical limitations trying to replicate the brain’s 100 billion neurons4 simultaneously firing and transmitting signals to thousands of destinations.
“The process has hit a wall, and it has become much harder to build faster computers without employing radically different architectures,” said Thomas Nowotny. He and research partner James Knight, Ph.D., at the University of Sussex began testing a new approach, using graphic processing units (GPUs) with new simulation software—and the result was breakthrough for the simulation speed and efficiency.
In the University of Sussex research, one GPU the size of a brick achieved processing speeds up to 10 percent faster than currently possible using their supercomputer. This translated to energy savings as previously, the high-speed, hot-running chips that powered the supercomputers they used generated a significant amount of heat and required energy-consuming cooling systems. And once interconnected, GPUs could create a model 50 times more powerful and approach the complexity of a monkey’s visual system, Nowotny and Knight said.
GPUs, which have traditionally been used in gaming applications, are fast and efficient simulators. Past simulations have not been able to precisely target specific areas in the brain linked to seizures. But with more cores than CPUs, GPUs have more power to pinpoint specific regions in the brain that cause epileptic seizures — and could one day guide doctors around which specific areas to remove.
Using Machine Learning to Mimic the Brain
Eight hundred miles away in Provence, France, brain mapping and simulation combined with machine learning are also helping researchers understand the brain of someone with epilepsy. Viktor Jirsa is a leader of the Human Brain Project’s (HBP) theoretical neuroscience team, based at Aix-Marseille University.
Jirsa’s group of neurosurgeons, neuroscientists, physicists, engineers and mathematicians uses MRI images and mathematics to create a three-dimensional model of an epileptic patient’s brain. The scientists then engage computing centers across Europe, using both HBP systems and technology from the newer GPU-based systems at Sussex — both designed to meet the complex requirements of neuroscience research. The supercomputer then maps the connections between different brain nodes and create what Jirsa calls an “avatar,” or a computerized physical representation of the pathways, leveraging visualization tools and methods that can handle large datasets.
Meanwhile, surgeons open up a patient’s skull and attach electrodes for an intracranial electroencephalogram (EEG). Jirsa’s team then links remotely to classical supercomputing centers HBP runs in Switzerland and Germany5. The Aix-Marseille team and other HBP projects across Europe can also access an innovative supercomputing platform called the SpiNNaker neuromorphic system, an array of cores assembled at the University of Manchester.
SpiNNaker contains 57,600 processors, each of which have 18 ARM 32 bit cores and 128 MB of mobile DDR SDRAM, totalling 1,036,800 cores and over 7 TB of RAM, according to the university6.
The Aix-Marseille team employs supercomputers to merge their avatar with the EEG measurements, then relies on AI algorithms to expand the electrical map throughout the brain model.
The supercomputer then runs simulations, tracing a seizure’s electrical disturbances back from various brain regions toward a point of origin. This is the precise target that is handed back to the surgeons for precision cutting — it can be a region a few millimeters across up to a quarter of a lobe that needs to be excised for relief of seizures.
Paving the Way for Future Treatment
SpiNNaker could have life-altering impacts for thousands of patients suffering from the most debilitating forms of epilepsy. Jirsa says the existing surgical techniques are imprecise and have been stuck in the same mode for decades. These methods are only successful in permanently limiting seizures 50 to 60 percent of the time7. While the HBP trials so far have been retrospective only, the simulation targeted the most relevant brain regions 79 percent of the time, a large improvement, Jirsa said.
Soon, Jirsa says, a new prospective clinical trial begins in Europe with a cohort of 400 patients, with each selection randomized, blinded and matched to a control group. The HBP team will model selected patients’ brains, run the simulations and recommend surgical targets to physicians, who will take the data into consideration before surgery. Jirsa says this “Virtual Epileptic Patient” is a large cohort for a surgery trial.
“It should give us enough results to be statistically significant,” Jirsa adds. “It’s very exciting.” If the trial proves successful, researchers will continue to refine the process and teach more neurosurgery teams how to replicate it.
The HBP is planning to roll out a new software interface for collaborating researchers that will better coordinate their remote use of the various hardware computing centers across Europe. The team is also working to find new shareable data-handling techniques for the information needed to conduct brain simulations, as well as more effective virtualization techniques that allow scientists to visualize in 3D the elusive interactions of the nodes and synapses of the brain.
“There is a very big responsibility, and at the same time it’s exciting that we have this chance to improve clinical practice and ultimately patients’ lives,” Jirsa said.
There’s still a long way to go before scientists can determine the impact of these simulations. But as doctors continue to treat patients with epilepsy, emerging technology and data processing methods could bring them to the greatest scientific discovery of all: a cure.
This content is produced by WIRED Brand Lab in collaboration with Western Digital Corporation.