Editor's note: In this opinion piece, PRF Correspondent Isobel Parkes, a PhD student at University College London, UK, takes a selective look at some of the key pain neuroimaging studies that have used machine learning (ML). She makes the case that the ML approach will continue to pay major dividends to the pain research field.
Machine Learning: The New Frontier in Pain Research?
You have probably heard that we are in the midst of an artificial intelligence and machine learning revolution.
The last few decades have seen new ways to produce, store, and analyze data, culminating in what is known as “big data” – datasets too large and complex to be dealt with by traditional data-processing methods. This has served as an impetus for the development of machine learning, an area within the field of artificial intelligence, which has advanced at an astounding rate, powered by deep learning algorithms that train multilayer neural networks.
The symbiotic relationship between big data and machine learning promises to revolutionize the pursuit of knowledge in science – including pain science – through novel and efficient ways of doing research. We, the researchers, now have the tools to interpret big data in meaningful ways.
Before discussing how these approaches are advancing pain research, two key definitions are in order. First, two of the fathers of the field of artificial intelligence (AI), Marvin Minsky and John McCarthy, defined AI in the 1950s as any task performed by a machine that would have previously been considered to require human intelligence.
Second, machine learning (ML) is a subfield of AI and responsible for the vast majority of AI advancements and applications you hear about. ML algorithms learn and improve through experience to find patterns in exceptionally large numbers of data.
Some highlights of progress using ML in the pain field
The study of pain is tremendously challenging because pain is a complex experience involving biological, psychological, and social contributors. This translates to extremely complex data. Pain is therefore uniquely suited to the type of interrogation that ML offers. Here I will focus on just a handful of the key brain imaging studies in the pain field that have used ML.
There has been much focus on the use of ML to generate objective biomarkers of pain, particularly in brain imaging studies. In research published in 2013, Tor Wager and colleagues used ML in a seminal study to define the neurologic pain signature (NPS), a pattern of fMRI activity across brain regions that was associated with heat-induced pain (see PRF related news article). It was among the first studies to show imaging as a promising biomarker of pain. Meanwhile, a 2015 study by Duff et al. used ML to identify fMRI signatures of brain activity to predict efficacy of candidate analgesic drugs (see PRF related news story), thus extending the value of this approach to the pain medicine development realm.
Multiple research groups have followed along this path, extending ML algorithms to define neurologic signatures in chronic pain disorders, including chronic low back pain, chronic pelvic pain, irritable bowel syndrome, fibromyalgia, and arthritis (see here for a review). An obvious clinical application of brain-based signatures is in the context of those with limited communication skills.
ML has also formed the basis of the emerging field of machine vision and emotion research, in which human facial expression has been used as a promising behavioral biomarker for emotion and pain in patients. This has been an important development for the accurate assessment of pain in infants, where self-report is not possible.
The first published ML-based infant pain evaluation study showing that facial expressions could be successfully used to classify pain states was the Classification of Pain Expressions (COPE) project. This work used three face classification techniques to distinguish painful expressions in newborns. Since then, several groups have also used ML methods to analyze infant facial expressions to recognize and assess pain, with notably high accuracy (Gholami et al., 2010Mansor et al., 2013Zamzmi et al., 2016).
The good times in machine learning continue
A paper published last year in Neurotherapeutics highlights the remarkable potential ML has in a clinical context. Psychological health plays a significant role in modulating chronic pain, with emotional disorders, including depression and anxiety, being common comorbidities of chronic pain. Tor Wager and colleagues therefore developed a platform to use on mobile devices, such as a cell phone, to track chronic back pain patients’ emotions by self-report. The mobile platform also asked patients to report the intensity of their pain and the corresponding body locations.
Bodily sensations can affect our emotions, and this is particularly apparent in chronic pain. Changes in bodily sensations can be associated with specific discrete emotions, and this can be represented by a bodily sensation map (BSM) where bodily correlates of emotions can be arranged topographically on the body. Pain-focused BSMs are used with chronic pain patients to indicate where the pain is in the body.
The authors of the Neurotherapeutics paper applied an ML approach to develop two predictive models based on the reported ratings of pain intensity, emotions, and corresponding BSMs. The ML models were able to predict pain levels two weeks after the patients’ reports. Future pain was best predicted by interactive effects of body maps of fatigue with negative affect, and also positive affect with past pain.
This approach is an important step toward a multidimensional biomarker/measure of pain, which begins to capture more of the complexity of the pain experience. For successful pain management in the future, personalized prevention and treatment are crucial because pain is a unique experience for each person. The study described above could contribute to the generation of a novel AI/ML framework in the clinic to provide an in-depth, sophisticated understanding of chronic pain in each individual patient.
The progress of ML approaches toward clinical application in the pain field continues apace. Earlier this year, an ML approach was used to develop the Tonic Pain Signature (ToPS). This fMRI-based signature reliably predicted sustained pain in healthy volunteers, and clinical pain in patients with subacute or chronic low back pain (see PRF related news story).
New questions
But it is not just that the application of ML to pain research will find new patterns in data that would otherwise be hidden from view. It is that these new patterns will allow us to generate questions about basic neurobiological mechanisms of pain that we can’t even conceive of right now.
It doesn't end there: New technology often facilitates scientific discoveries that turn our fundamental understandings on their heads; previous explanations may no longer hold merit when that technology is put to work. This may prove to be the case with the AI/ML revolution in the pain field.
The use of ML tools in pain research is still in a nascent stage, and yet we are already beginning to translate novel ML-based technology into the clinic. For the pain field, the AI/ML revolution is here to stay.
Isobel Parkes, PhD student, University College London, UK.
Image credit: vgraphix/123RF Stock Photo.