Article In Brief
For the first time, researchers found that small signs of electrical activity detectible in the early stages of an acute brain injury can be targeted and amplified by machine learning to help patients “relearn” how to physically respond to specific motor commands.
A team of researchers has successfully tested a machine learning algorithm that can scan electroencephalogram (EEGs) for signs of basic verbal command recognition in otherwise unresponsive acute brain injury patients, potentially enabling a better way to identify patients more likely to achieve some degree of motor recovery and independence.
What's more, the investigators at Columbia University and New York University were able to link repeated activation of specific brain activity in unresponsive patients, early after injury, with better one-year functional outcomes, according to the findings reported in the June 27 issue of the New England Journal of Medicine.
Previous studies have shown that EEG or functional MRI (fMRI) can sometimes suggest that parts of the brain are activated in response to spoken commands in unresponsive patients, but whether the detected signal represents recognition or comprehension of commands has been unclear.
Most prior studies in this area have not focused on patients in the acute stages of brain injury, said study co-author Jan Claassen, MD, associate professor of neurology and medical director of Columbia's Neurological Intensive Care Unit.
“More patients who did have an EEG response to spoken commands later recovered than those who did not have this pattern,” Dr. Claassen said, adding that the new findings must be studied further in larger numbers of patients.
In the study, the researchers noted that a meta-analysis published in 2016 found that 14 percent of chronically unresponsive brain injury patients may suffer “cognitive–motor dissociation” between brain activation and behavior for months or even years after an injury.
“Such a dissociation between behavior and EEG activity responding to spoken motor commands is called cognitive–motor dissociation,” Dr. Claassen explained. “In our study, this was seen more often in patients with traumatic injuries or brain hemorrhages than in those with hypoxic–ischemic injury, but it was also discovered in patients with other acute brain injuries and, in some cases, those who had been lightly sedated.”
Dr. Claassen said that the frequency of cognitive–motor dissociation, and any prognostic associations, still require validation in larger, multicenter studies powered to detect differences in long-term outcomes.
“It is possible that these patients had overall greater functional integrity of the brain stem, thalamus, and cortex and of the connections among these structures, something that has been reported in earlier studies.”
He added that the next steps with the research will be to evaluate the machine-learning technique in other types of brain injuries, and in larger groups of patients, for longer periods of time, and to examine with greater granularity how the process and other factors influence quality of life in patients once they are discharged.
“It will also be important to more closely examine what is really happening in the brain with cognitive–motor dissociation using more sophisticated imaging techniques.”
Study Design, Findings
The prospective study was conducted in a consecutive series of 104 unresponsive brain injury patients at a single neuroscience intensive care unit for acute brain injuries. The injuries were the result of different causes. Patients were in a coma, vegetative state, or minimally conscious state, defined as “unresponsive, but with preserved visual fixation, visual pursuit, or localization to noxious stimuli” after an acute brain injury of any type and were either undergoing or expected to undergo continuous EEG monitoring within 12 hours.
A machine learning algorithm was used to analyze EEG for patterns of brain activity in response to repeated spoken commands. Machine learning applies a self-teaching artificial intelligence (AI) algorithm that can scrutinize data, identify patterns, and make decisions without human involvement.
Half of the 15 subjects, and 26 percent of 88 patients with no early signs of EEG activity improved enough to follow commands before being discharged, and after one year, 44 percent of those with signs of brain activation, and 12 of 84 patients without, recovered enough to function independently for at least eight hours in a day.
Machine learning was applied to EEG recordings to detect brain activation when patients were repeatedly commanded to open close their left or right hand.
Daily neurologic examinations assessed the ability or inability of patients to follow spoken commands—such as “stick out your tongue,” “show me two fingers with your right hand,” or “wiggle your toes”. Each EEG assessment was preceded by a clinical exam that included screening to categorize the clinical state of consciousness at that time.
Functional outcome at 12 months was determined using the Glasgow Outcome Scale–Extended with levels ranging from 1 to 8, with higher levels indicating better performance.
Each EEG was then used to train the machine learning algorithm, and performance was determined using a standardized learning curve scale, both before and after random commands to “keep opening” and “stop opening” each hand.
“Our study shows that cognitive–motor dissociation can be detected in the ICU early after brain injury in a number of unresponsive patients, and while functional MRI might detect more cases than EEG, fMRI is challenging in the critical care setting,” said Dr. Claassen.
“This is a very important study that helps us better understand a lot of ideas that have emerged in the last 15 years or so about dissociation of conscious awareness and apparent unresponsiveness in patients with chronic brain injuries,” said Nicholas D. Schiff, MD, professor of neurology and neuroscience at Weill Cornell Medical College in New York.
“This study opens up some big questions about brain function in unresponsive patients and how signs of brain activity can easily be missed in the first few days in the ICU. Larger studies are needed, but this shows us something incredibly important that could be occurring in some patients.”
Alejandro A. Rabinstein, MD, FAAN, professor of neurology at Mayo Clinic in Rochester, MN, told Neurology Today that the new findings add important information to current understanding of cognitive-motor dissociation in acute coma. However, he cautioned that they should be considered strictly investigational at this point.
“While the findings are interesting and show promise, this will not be ready for clinical use until further research is conducted. The performance of the machine learning algorithm must be replicated and, if possible, refined, before it can be translated into practical application.”
He said the study also suggests that traditional clinical examination at the bedside may miss recognition of some level of responsiveness that can be uncovered by EEG changes, using a protocol that would be feasible in the ICU during the acute phase.
“If this is validated, it could offer a new means of gauging the likelihood of emergence from coma,” he said, adding that neurologists should be aware of certain unique features of the study.
“Most previous studies have examined cognitive-motor dissociation in chronic disorders of consciousness and typically only in traumatic brain injury patients, whereas this study looks at this in different types of injury during the acute stage.”
“The findings of the study, which are in line with prior data that demonstrate cognitive-motor dissociation, support the notion that the inability to follow commands on a clinical exam is not indicative of the lack of brain activity in response to commands,” said Ariane Lewis, MD, associate professor of neurology and neurosurgery, and director of the neurocritical care division at NYU Langone Medical Center.
“However, it is clear that while some patients with purely electroencephalographic responses to commands recover, others do not, and some patients without brain activation on EEG do recover,” she noted.
“Further study is needed to optimize the utility of EEG for neurological prognostication in unresponsive patients, and to decrease the number of false positive and negative results by identifying which patients, and the timing of an EEG, is most informative.”
She said that EEG findings should be considered in conjunction with results from a clinical examination and neuroimaging when prognosticating the potential for recovery in clinically unresponsive patients.
Dr. Claassen is the principal investigator on studies from Charles A. Dana Foundation and Hames S. McDonnell Foundation; receives grant funding from the NIH, Bard Pharmaceuticals, and is a minority shareholder in iCE Neurosystems. Drs. Schiff, Rabinstein, and Lewis report no conflicts.