A new study uses machine learning to predict the development and spread of chronic pain based on a shared set of mental health, sociodemographic, and physical health factors.
At first glance, many kinds of chronic pain would seem to have little in common. What does head pain have to do with, say, knee pain? What possible connection could there be between stomach pain and, for instance, facial pain? A new study builds the case that these and other types of pain all share something in common: A variety of biopsychosocial factors that contribute to the pain experience.
Researchers led by Etienne Vachon-Presseau, McGill University, Montreal, Canada, used a machine learning algorithm, which learns from data the researchers feed into it, to create a model that would classify different chronic pain and pain-related medical conditions, and that would predict the development and spread of pain, based on a plethora of mental health, sociodemographic, and physical health factors.
The model performed well at classifying pain at eight different body sites, and at classifying pain-related medical conditions. It also fared well at predicting the spread of pain from one body site to others. Though the model was less successful at classifying migraine and non-migraine headache, the data still pointed to a shared set of biopsychosocial factors among those and roughly two dozen other pain-related medical conditions that might otherwise appear quite dissimilar.
The most important risk factors for co-occurring pain at multiple body sites turned out to be mood, sleep, neuroticism (feeling “fed up”), and body mass index, but many other biopsychosocial factors also played a role.
The new research “adds to the literature supporting the notion that sleep problems, cognitive problems, fatigue and mood problems are an inherent part of what we now refer to as nociplastic pain,” wrote Daniel Clauw, a pain researcher at the University of Michigan, US, in an email to Migraine Science Collaborative. (Nociplastic pain is pain that persists after an injury or disease has resolved.)
While Clauw was not surprised by the results of the research, writing that his own new study, which was recently accepted for publication, reports “nearly identical results” in adolescents, “[t]he fact that this work [from Vachon-Presseau and colleagues] was done in such large databases and then replicated is what is most noteworthy,” according to Clauw, who was not involved with the new work.
“The really strong message is that (in population terms) studying the biology of chronic pain needs to start with psychosocial factors (which are the most important predictors of onset),” said Gary Macfarlane in a tweet. Macfarlane is a leading pain epidemiologist but did not take part in the current study.
The research appeared in the July 2023 issue of Nature Medicine.
A withdrawal from the (Bio)bank
The availability of a unique data set made the research possible in the first place.
“This study, and a lot of what we do in the lab, gravitates around this wonderful opportunity, which is called the UK Biobank,” said Vachon-Presseau.
The UK Biobank is a biomedical database containing medical and genetic data from half a million volunteer participants in the UK. Vachon-Presseau, who is a brain imager by training, was originally interested in this resource because of the imaging data it has. But in 2019, the UK Biobank started asking participants questions about pain, and he and his colleagues, including current study first author Christophe Tanguay-Sabourin, also from McGill, decided to capitalize on that information to broaden their studies.
They were particularly interested in data that the UK Biobank has about pain conditions that often co-exist in the same individual, known as chronic overlapping pain conditions. These conditions include headache, fibromyalgia, temporomandibular disorder, and low back pain, among others.
That different pain conditions frequently overlap suggests that they share common risk factors. The new study asked what those factors are and whether they can be used to classify and predict chronic pain and pain-related medical conditions.
Pain often spreads
The researchers started by gaining a sense of what pain looked like in UK Biobank participants, considering chronic pain (pain for more than three months) at eight anatomical sites including the head, face, neck/shoulder, stomach/abdomen, back, hip, knee, or pain all over the body.
They saw that 44% of those with chronic pain reported pain at more than one site, with co-occurring pain more likely to be present at nearby anatomical sites rather than at body sites far away from each other. For instance, the odds of having both headache and facial pain were more than six times the odds of having both headache and knee pain.
The team further saw that all of the 25 pain-related medical conditions they looked at were characterized by pain at multiple body sites. That included migraine and non-migraine headache, which featured not only head pain but also pain at the neck and shoulder, the back, and other sites. Finally, the researchers also observed that the higher the number of pain sites, the longer the pain lasted, the more intense it was, and the more it had an effect on daily living activities (known as high-impact pain).
“How pain spreads over the body seems to be a much more important concept than we initially thought. Pain spreading ends up having the highest impact and it requires more resources and attention to treat. It would potentially be useful to screen for pain spreading early on,” said Tanguay-Sabourin.
Classifying and predicting pain
Next, the team used their machine learning algorithm to develop a risk score for pain, based on 99 biopsychosocial factors including mental health, sociodemographic, and physical health factors. The group asked if the risk score would accurately classify pain, as well as predict the development and spread of pain. They used the UK Biobank data to train the algorithm, and then a separate dataset of almost 50,000 participants in whom longitudinal data about pain were available to test how well the risk score performed.
Overall, the risk score accurately classified people with chronic pain from those without pain, at each of the eight body sites. It also accurately classified people with the 25 pain-related medical conditions from those without those conditions, but for some conditions it did better than others. For instance, the performance was best for classifying fibromyalgia, chronic fatigue syndrome, and chronic obstructive pulmonary disease. But the performance fared the worst for classifying migraine, and the fourth worst for non-migraine headache.
In terms of prediction, the researchers used their longitudinal dataset to see if the risk score predicted the spread of or recovery from chronic pain roughly nine years after baseline pain measurements. They found that baseline chronic pain at each pain site was associated with higher odds of experiencing pain at the same site, or at proximal sites, after the nine years. While how much pain a person had at baseline predicted the spread of pain to nearby sites, those with higher risk scores were more likely to have pain that spread to distal sites. The risk score also did well at predicting the worsening or improvement of pain-related medical conditions at nine years.
The risk score also succeeded in predicting high-impact pain. Here, the investigators looked at outcomes associated with that type of pain, such as the use of opioids and one’s overall health, finding that the risk score predicted the initiation of opioids and the development of disability.
Why are chronic pain conditions alike? Look to biopsychosocial factors.
The researchers had used their algorithm on a plethora of biopsychosocial factors to create the pain risk score. So which of those factors were the most important for accurately predicting pain spreading?
It turned out that mood (feeling tired and consulting a general practitioner for depression/anxiety), body mass index (above 30), neuroticism, and sleep (insomnia) were the best early predictors. Meanwhile, a person’s occupation was the least important factor. Socioeconomic factors and life stressors fell somewhere in the middle.
Further, the team tested a number of models for each pain site separately. The goal was to see if the biopsychosocial risk factors were specific to different pain conditions. They found that the factors that best predicted pain spreading were those that were most homogenous between the various pain conditions. And what factors were most similar? Once again, it was factors like tiredness, mood, and others that the researchers had identified in their earlier experiments as being the most important.
“These biopsychosocial factors are not specific to the underlying pathophysiology of the condition,” said Vachon-Presseau.
Keep it simple
The team also tested whether a simplified version of their risk score could predict the risk of pain spreading, again using the UK Biobank cohort. Rather than the 99 biopsychosocial factors that formed the basis of their more complicated model, the simplified version relied on just six factors: Sleep, neuroticism, two measures of mood, life stressors, and body mass index. This simplified model performed similarly to the more complicated one.
The group also tested the simplified version in two other cohorts, including the Northern Finland Birth Cohort and the Pre-symptomatic Evaluation of Novel or Experimental Treatments for Alzheimer’s Disease cohort. In these two cohorts, the simplified model performed comparably to the model based on the UK Biobank cohort.
In terms of future research, one question the study results raise is why the models were relatively lackluster for classifying migraine and non-migraine headache.
“There’s clearly a different etiology that creates headaches, and this is a very important question that we would like to dig deeper into and potentially address,” said Tanguay-Sabourin.
He added that it’s an open question whether headaches in people who have widespread pain conditions differ from headaches in those with primary headache disorders, such as migraine or cluster headache.
The researchers are also interested in learning if certain pathophysiological characteristics can classify the medical conditions, and to do so they are now analyzing various types and sources of data, ranging from genetic data and blood draws to bone density and more. So far, they have found that this approach does successfully classify some, but not all, of the medical conditions. But classifying the pain is a different story.
“These biological markers of disease don’t work as well to predict the pain; that’s really elusive from a biological perspective,” said Vachon-Presseau. “The biological measures can classify the medical conditions, but they will have a hard time capturing the whole experience of the patient.”
The investigators also hope to understand whether or not pain spreading is a single phenomenon that always looks the same regardless of the type of pain or pain-related medical condition, or instead may differ depending on underlying pathophysiology.
As far as improving the care of people with chronic pain, including headache, one interesting way that the new research could do so is simply by increasing the attention of pain and headache researchers to the issue of pain spreading. So far, investigators have been more focused on how acute pain becomes chronic.
“Of course, the acute to chronic pain transition is super important; if we can fix that, it’s going to be great. But it’s not the only story. Maybe for people who are already in chronic pain states, there’s still an evolution that is going on, and we need to be aware of and take care of it. It’s not true that once you’re chronic, your pain isn’t evolving,” Vachon-Presseau said.
Along similar lines, Tanguay-Sabourin told Migraine Science Collaborative, “in terms of treatment, even if you’re receiving treatment specifically to try to help your migraine, for example, if you don’t consider all of the other pain sites, maybe you’re limiting yourself in terms of how much you can help. Treatment could potentially be optimized by making sure you’re taking care of other co-occurring pain.”
Finally, could the researchers’ simplified risk score be used in the clinic to predict pain?
“At the moment it’s unclear, to be honest. The impact it will have on clinical practice in the long-term is very unknown at the moment,” according to Vachon-Presseau, who said he is working with colleagues who treat pain patients in Montreal to investigate the matter further.
But in the bigger picture, the study’s focus on the biopsychosocial contributors to chronic pain may be what matters most.
“I’m not sure the score itself will be used much,” Clauw told Migraine Science Collaborative, “but the concept that these [biopsychosocial] factors predict the subsequent development of pain is very important.”
A prognostic risk score for development and spread of chronic pain.
Tanguay-Sabourin et al.
Nat Med. 2023 Jul;29(7):1821-1831.
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
Sign Up For An MSC Newsletter
Share this article: