Brain-powered Machine Learning


“Where are you right now?

What’s the score of the game?

How many fingers am I holding up?”

In movies, a rattled athlete or soldier is asked a series of questions, such as these, to diagnose a concussion.

Although these questions may be cinematic clichés, they still make up a portion of the cognitive and neurological tests used to spot concussions, also called mild traumatic brain injuries.

Oftentimes, though, these tests are still highly subjective since responders can influence results.



Recent reports have shown that the number of concussions continues to grow in both amateur and professional athletics. And football isn’t the only sport affected, either.

Thinking more could be done, neuroscience engineer Dr. Bill Rose and former researcher Dr. Corey Hart, who is currently applying neurally inspired cognitive computing solutions to problems in the energy sector, went to work.

In research today, doctors use functional magnetic resonance imaging (fMRI) scans to research activity in the brain. An fMRI scan uses MRI technology to measure brain activity over time by looking at blood flow. 

For the last two years, Dr. Rose and the team have been studying how to apply machine learning algorithms to brain images from these fMRI scans.

Machine learning is a type of artificial intelligence that focuses on computer programs teaching themselves to learn and adapt over time when exposed to new data.




Example of Functional Magnetic Resonance Imaging (fMRI) scan during working memory tasks.

Dr. Rose describes the challenge of objectively detecting concussions using biomedical data as the “holy grail” of mild traumatic brain injury (mTBI) research.

“Diagnosis and treatment of mild traumatic brain injuries represents a long-standing challenge for physicians in combat situations, both professional and amateur sports, and hospital emergency rooms,” said Dr. Rose. “Our research goal was to develop an objective methodology for detecting the existence of these injuries mapped to regions of interest in the brain.”

As Dr. Rose points out, Leidos has a long history of using advanced algorithms to quickly seek anomalies in geospatial intelligence, like aerial photographs. The team applied a novel way of looking for changes in data across time.

“Instead of focusing the algorithm to detect an anomaly from hundreds of aerial images, we’re feeding the algorithm many brain images, such as from fMRI or EEG scans, to detect anomalies as compared to normal, non-injured brains,” said Dr. Rose. “As the algorithm ingests more data, it gets smarter, and is able to tell the injured brains from the non-injured.”



Comparing two baseline “normal” patients (left) with two mild traumatic brain injury (mTBI) patients (right). This figure showcases brain regions of interest as well as neural network connectivity of the algorithm.

Dr. Rose and team collaborated with the University of Maryland’s Neuroimaging Center and School of Medicine to share their expertise and fMRI data.

A second experiment was carried out using EEG data from Ursinus College, which has a policy of testing their athletes pre- and post-season for concussive injury.

Using anonymized EEG data, Dr. Rose’s team compared two experimental designs, in an effort to pinpoint the best approach. In the end, a spiking neural network came out on top, with an accuracy rate above 90 percent.

In the past, a human would look at this data, but this technology can mimic what a human would do to detect a concussion more objectively. An important distinction here, though, is that the machine is simply flagging abnormalities for doctors to further investigate, not diagnosing.

Today, the team is in the process of writing a results article for peer-reviewed journals in the biomedical engineering field and has two patent filings for these technologies.



As machine learning research and technology become more sophisticated, Luiz Pessoa, director of the University of Maryland’s Neuroimaging Center is optimistic that medical breakthroughs are within reach.

“Advances in MRI hardware and pulse sequences have pushed this [machine learning] technique much further,” Pessoa said. “Research projects have gone from studying hundreds of patients to thousands of patients. It’s a really interesting time right now.” 

Ultimately, Dr. Rose and his team are looking to devise a portable solution, capable of running on a mobile phone and paired to a portable EEG head set, enabling rapid, objective and accurate detection of concussions on sports fields or the battlefield.