Access to the PPM Journal and newsletters is FREE for clinicians.

Machine Learning Approaches Could Lead to Objective Pain Measures

May 17, 2019
Researchers have developed an algorithm that may be able to predict pain intensity

With Vitaly Napadow, PhD

Researchers have combined machine learning with neuroimaging and autonomic measures to develop a tool that can classify high and low pain states and may soon be able to predict pain intensity in individuals with chronic pain.1 Machine learning is a type of artificial intelligence that builds a prediction model based on data.

The tool offers an important step toward developing objective measures of clinical pain intensity in the chronic pain population. The current lack of objective pain markers has limited optimal pain management for these patients and largely held pain research on diagnoses and treatments at a standstill.

(Source: 123RF)

Machine Learning and Pain States

Most research to date on rating pain has relied on healthy individuals.2,3,4 However, Vitaly Napadow, PhD, director of the Center for Integrative Pain NeuroImaging and an associate professor at the Martinos Center for Biomedical Imaging at Massachusetts General Hospital and Harvard Medical School, and her colleagues recently worked with 53 patients suffering from chronic lower back pain. Patients ranked their baseline level of pain on a scale of 0 to 100. The research team then worked with the patients to identify physical maneuvers that exacerbated their pain. Their aim was to achieve an increase in pain of at least 30% over baseline; in the end, the average increase in pain ratings was 74%.

“This is one of the first studies that has applied this approach in chronic pain patients rather than experimental pain in healthy individuals,” Dr. Napadow told PPM, noting that the distinction calls out the “many differences between chronic pain patients and healthy individuals.” Patients living with chronic pain appear to have altered central nervous system processing, which in turn, may lead to abnormal pain perception. In addition, these individuals may have aberrant activity/connectivity in their brain networks, along with amplification of sensory input to the brain.5-9

“What we’re trying to do is develop something that is specific to pain intensity, differentiates a higher pain state versus a lower pain state, and try to generalize across different pain patients and predict their pain intensity,” he said.

In addition to rating the subjects’ pain, Dr. Napadow’s team scanned the patients’ brains using magnetic resonance imaging (MRI) before and after performing the physical pain-exacerbating maneuvers. Neuroimaging data included resting-state functional MRI (fMRI) using two different approaches. The first approach involved regional cerebral blood flow (rCBF) to capture brain activity, while the second approach used functional MRI data to calculate functional connectivity between the region of the primary somatosensory cortex (S1) that codes for the back and the rest of the brain. High-frequency heart rate variability (HRVHF), an autonomic marker calculated from cardiac data, was also measured before and after the maneuvers.

“Each putative biomarker targeted a unique physiological dimension of central and autonomic processing supporting pain perception,” the investigators noted in their published study. Regional CBF captures changing activity across the brain, which may reflect pain exacerbation. S1CONN captures afferent nociceptive information from the low back. HFHRV captures autonomic nervous system changes associated with clinical pain perception. “Importantly, combining these multimodal putative biomarkers produced a synergistic effect for clinical pain prediction, both within and between patients with chronic lower back pain,” the investigators explained in their published study.1

A New Algorithm for Pain?

Dr. Napadow’s team developed a classification algorithm to distinguish between high and low pain states using imaging and autonomic data. Specifically, the researchers discovered that they could predict a patient’s subjective pain level based on objective neuroimaging and autonomic markers. They found that all three parameters (rCBF, S1CONN and HFVHRV) were significant components of the inter-patient classification between pain intensity states (the difference between low and high pain in a single patient vs high and low pain within a group of patients). However, the combination of all three variables resulted in the highest accuracy predictions of high-intensity pain ratings (92.45%).

Next, they tested their algorithm to see how well it could directly predict clinical pain intensity ratings across patients, using independent training and testing data sets. Using all three parameters (rCBF, S1CONN and HFHRV), the model was able to accurately predict pain intensity for both the independent training and testing data sets. However, in this case, only the S1CONN parameter significantly contributed to the prediction.

Future Outlook

While this particular study was limited to chronic low back pain, and being able to reproduce the pain using physical maneuvers, the algorithm may be applicable to other types of chronic pain. “Once we’ve defined these potential biomarkers, we could try to export it,” said Dr. Napadow. “We could try to see if this biomarker may also differentiate fibromyalgia patients from healthy subjects or predict the intensity of pain in these fibromyalgia patients, even if they’re not doing any physical maneuvers.” In fact, the team is already collecting data as part of a large study including individuals with fibromyalgia.

While neuroimaging biomarkers may not be practical for clinical use just yet, the research may, however, lead to the identification of non-neuroimaging predictive measures making the prediction of pain levels easier and faster within a clinical setting.

In addition, “if the model is generalized across different chronic pain populations and different contexts, this pain signature could have great promise for pain assessment in noncommunicative patients and identification of objective pain endophenotypes…aimed at [the] discovery of new approaches to combat chronic pain,” Dr. Napadow’s team concluded in their paper.1

Last updated on: June 21, 2019
Continue Reading:
Precision Medicine Tool Improves Predictability of Opioid Use Disorder
close X