Leveraging Machine Learning to Develop Objective Biomarkers for Pain

Dr Mackey discussed potential applications of machine learning technology in chronic pain management and research.

The following article is part of conference coverage from the PAINWeek 2018 conference in Las Vegas, Nevada. Clinical Pain Advisor’s staff will be reporting breaking news associated with research conducted by leading experts in pain medicine. Check back for the latest news from PAINWeek 2018.

LAS VEGAS — Sean Mackey, MD, PhD, chief of the pain medicine division at Stanford University in California, discussed potential applications of machine learning technology in chronic pain management and research in a presentation during the 2018 PAINWeek conference, held September 4-8.1

Pain intensity is currently assessed based on patient self-report, which is still considered the gold standard in the field. However, such reports may be limiting as they are subjective in nature and have limited use in young children, the elderly, and patients in intensive care. Efforts to leverage electroencephalogram, skin conductance, and heart rate measurements as a way to evaluate pain have all failed, begging the question of whether brain imaging may provide a more accurate assessment of pain intensity. Such objective pain biomarkers may be used at a number of levels during clinical assessment and treatment of patients, serving as markers for diagnosis; prognosis; treatment response, efficacy, and toxicity; and the risk for developing a disease, but also as an outcome measure in clinical trials and pain studies. In short, such pain biomarkers would enable the rise of a precision medicine for pain.

In a study published in 2011 in PLoS One, Dr Mackey’s group sought to determine whether certain brain patterns could be associated with painful stimuli using functional magnetic resonance imaging (fMRI) and support vector machine (SVM) learning.2 The brains of 24 individuals were scanned using fMRI during the presentation of both painful and non-painful stimuli. A linear SVM was trained on 8 individuals to discriminate between painful and non-painful stimuli on whole-brain activity patterns. This trained SVM model was then tested on another 16 individuals and showed 81% accuracy in discriminating noxious vs non-painful stimuli (P <10-7). Activity in pain-processing regions (ie, primary and secondary somatosensory cortices and insular, cingulate, and primary motor cortices) was found to reflect painful stimuli using the SVM model and thus may serve as a “signature for pain detection.” In particular, activity in the posterior insula was found to be critical in the differential detection of painful stimuli. A later study published in the New England Journal of Medicine by a group at the University of Colorado in Boulder identified a neurologic signature for noxious vs non-painful thermal stimuli, also using fMRI, that had high sensitivity and specificity and was greatly attenuated by remifentanil.3

In another study by the Mackey group published in 2014, researchers were able to classify chronic low back pain with 76% accuracy based on the comparison of gray matter density determined by MRI and SVM in 47 patients with chronic low back pain and 47 healthy participants. Increases in gray matter density in the middle frontal gyrus, primary somatosensory cortex, and insula and reduced gray matter density in the supplementary motor area and inferior frontal gyrus were found to be associated with the presence of chronic low back pain.4

Dr Mackey expressed some ethical concerns regarding using neuroimaging for the detection of pain. This approach should not replace currently used self-reports to assess pain intensity, he stated, but rather, should enhance those measures. He also noted that this technology is not yet ready for use in clinical practice, cautioned against its medical-legal use, and expressed concerns related to its inappropriate use.

Related Articles

In a 2017 consensus statement published in Nature Reviews Neurology, a presidential task force of the International Association for the Study of Pain, constituted to “examine the capabilities of brain imaging in the diagnosis of chronic pain, and the ethical and legal implications of its use in this way,” recognized the potential of this approach while acknowledging the need for further development.5 “[Brain imaging] has the potential to increase our understanding of the neural underpinnings of chronic pain, inform the development of therapeutic agents, and predict treatment outcomes for use in personalized pain management,” wrote the statement authors, who formulated a set of necessary standards of evidence to be met before using the technology in clinical practice or in the legal arena.

“Brain-based biomarkers should be used as an adjunct to, rather than a replacement for subjective reports of the pain experience,” concluded Dr Mackey. “Current brain-based measures should be used only to understand brain mechanisms underlying pain, factors that lead to persistence of pain, and targets in the brain for safe and effective pain management,” he added.

Follow @ClinicalPainAdv


  1. Mackey S. Brain-based biomarkers for pain: scientific, legal, and ethical issues. Presented at: PAINWeek 2018, September 4-8, 2018, Las Vegas, Nevada. Presentation SIS-01.
  2. Brown JE, Chatterjee N, Younger J, Mackey S. Towards a physiology-based measure of pain: patterns of human brain activity distinguish painful from non-painful thermal stimulation. PLoS ONE. 2011;6(9):e24124.
  3. Wager TD, Atlas LY, Lindquist MA, Roy M, Woo CW, Kross E. An fMRI-based neurologic signature of physical pain. N Engl J Med. 2013;368(15):1388-1397.
  4. Ung H, Brown JE, Johnson KA, Younger J, Hush J, Mackey S. Multivariate classification of structural MRI data detects chronic low back pain. Cereb Cortex. 2014;24(4):1037-1044.
  5. Davis KD, Flor H, Greely HT, et al. Brain imaging tests for chronic pain: medical, legal and ethical issues and recommendations. Nat Rev Neurol. 2017;13(10):624-638.

For more coverage of PAINWeek 2018, click here.