Editor’s Note: The 2016 annual meeting of the Society for Neuroscience took place November 12-16, 2016, in San Diego, US. Part 3, below, of our coverage of the meeting is a summary of a study that combines brain imaging and machine learning in the search for a pain biomarker. (See Part 1 and Part 2).
In an early step toward finding a brain imaging biomarker for pain, researchers have detected brain activity patterns in individual patients with chronic low back pain (cLBP) that differ depending on the pain level that the patients experience. The team from Massachusetts General Hospital, Boston, US, used machine learning algorithms to analyze brain imaging data captured with arterial spin labeling (ASL). They presented the work in a poster at the 2016 annual meeting of the Society for Neuroscience.
Pain is a subjective experience, and the only way to gauge that experience is through patient self-report. Unlike many other diseases and medical conditions, there is no validated biomarker of pain—no blood test or brain scan. Some researchers hope that patterns of brain activity might one day serve this elusive role.
In a press conference held at the meeting with directors of the National Institutes of Health, Nora Volkow, director of the National Institute on Drug Abuse (NIDA), and David Shurtleff, deputy director of the National Center for Complementary and Integrative Health (NCCIH), Bethesda, US, both bemoaned the lack of objective biomarkers as an impediment to advancing the understanding and treatment of chronic pain. Shurtleff specifically touted the new work from senior author Bruce Rosen. “This research is a preliminary step in the right direction for developing a tool for objectively measuring a pain state. This could give us the insight we need for understanding brain mechanisms associated with various pain conditions,” Shurtleff later told PRF in an email.
A machine learning approach
In many previous imaging studies, researchers have used strategies to evoke pain in healthy people or chronic pain patients, but the goal is to find the brain activity signatures of sustained clinical pain.
“It’s hard to turn clinical pain on and off,” said another senior author, Vitaly Napadow. “So we’ve used a strategy to exacerbate it.”
So the researchers looked at brain images from individual patients with cLBP when their pain was at a low, baseline state and compared them to images after each patient performed physical maneuvers intended to temporarily increase pain. The maneuvers included movements such as sit-ups or back-arching motions. Pain ratings increased in all 39 patients, by an average of 80 percent.
The approach “is really powerful, because it allows us to do paired analysis in the same patients when they are in high and low pain states,” Napadow said.
The team, including presenting author Jeungchan Lee, used ASL functional magnetic resonance imaging (fMRI) to track the volume of regional cerebral blood flow in the brain. Similar to conventional fMRI that measures blood oxygen level-dependent (BOLD) signals, ASL rests on the premise that sites of higher activity in the brain demand increased regional blood flow. But unlike BOLD fMRI, which measures blood oxygenation, ASL fMRI uses a non-invasive radio frequency signal to magnetically “tag” the water molecules in blood as they flow through the arteries that feed the brain. By constantly comparing images of tagged and untagged blood, the researchers were able to use ASL to quantify regional blood flow. Compared to BOLD signals, ASL is “more specific, but less sensitive,” Napadow said. Whereas BOLD fMRI is very good at detecting discrete events, ASL is better for slowly evolving or stable states, he said.
To analyze the imaging data, the researchers used a conventional method of statistical analysis called the paired-t test and two different machine learning algorithms, called the k-nearest neighbor and the support vector machine (SVM). Predictably, in the high pain condition compared with low pain, conventional analysis showed that brain activity was increased in regions associated with pain, including the putamen, which is part of the basal ganglia, and the thalamus.
But when Lee used machine learning algorithms, which allowed him to compare low and high pain states in individual patients, these algorithms identified other specific brain areas that were activated by exacerbated pain. The results are preliminary, however, and Napadow says they will continue to analyze the data before publishing the pain “signature.” The finding suggests that machine learning analysis may better discriminate between high and low pain states, compared to conventional imaging analysis.
The primary somatosensory cortex (S1) is a brain area that contains a homunculus—that is, a somatotopic map of the body’s surface. When pain is evoked in a particular area of the body, activity in the cortex representing that area typically increases, whereas the rest of S1 will show less activity. Interestingly, conventional analysis of the imaging data showed, as expected, that somatosensory cortical activity did decrease in areas representing non-back regions, but no increase was detected in cortex representing the painful back regions. The SVM machine learning algorithm, in contrast, revealed increased activation in S1 that mapped to the painful back region.
Toward a biomarker
Napadow stressed that this work is in its early stages, and that it represents just one contribution among many that will be needed to find a validated biomarker of clinical pain—but to what end? Would a biomarker be used for diagnostic purposes, to “prove” that someone does or does not suffer from chronic pain? Might it be used to deny legal, medical, or financial benefits to patients? How would it be used as a tool to aid in treating patients? These are open questions, and many researchers—Napadow among them—have concerns about how a potential biomarker might be misused.
One thing is certain, Napadow says: researchers are unlikely to find anything nearly as specific and sensitive for pain as a genetic test, where DNA evidence can confirm or deny a match between, for example, a sample of blood found at a crime scene and one from a suspect, with 99.99 percent accuracy. “Even with that level of specificity, that sort of DNA testing still gets questioned in legal settings. And we are nowhere near that when it comes to an imaging biomarker for clinical pain.”
While patient self-report of pain is a subjective measure, Napadow said, it should not be discounted. “Imaging is just part of the story,” he said. Self-report can be used in conjunction with imaging. For example, researchers might use imaging to probe responses to therapies, tracking patient-reported experiences along with imaging data. Another proposed use of signatures of chronic pain is for teaching patients self-regulation of pain. For example, several groups are now using real-time fMRI-guided neurofeedback to teach chronic pain patients to control their brain activity (for review, see Chapin et al., 2012).
Stephani Sutherland, PhD, is a neuroscientist, yogi, and freelance writer in Southern California.
Image credit: Society for Neuroscience