Integrating AI and Machine Learning Into Pain Research and Therapy

Abstract

Artificial intelligence (AI), along with its core technology, machine learning (ML), has become deeply integrated in almost every area of science and aspect of everyday life. Consequently, this development has reached pain research and therapy as well. Over the past two decades, AI and ML methods have become more widespread in pain-related studies, though there were initial challenges in publishing such work in pain-focused journals. A decade ago, pain journals were sometimes reluctant to accept papers that used machine learning as the main approach to data analysis, considering it a topic more appropriate for computer science or bioinformatics journals. However, this scenario has undergone a significant transformation, and machine learning has become an integral part of the methodological arsenal used in pain research (Lotsch and Ultsch 2017; Lötsch et al. 2022).

Reflecting this broader trend, the European Journal of Pain has witnessed a gradual increase in publications applying AI and machine learning to pain research and therapy (Figure 1). According to a PubMed search for the string ((“machine-learning”) OR (“machinelearning”) OR (“machinelearned”) OR (“machine learning”) OR (machine-learned) OR (“machine learned”) OR (“artificial intelligence”) OR (“explainable AI”) OR (“explainable artificial”) OR (XAI) OR (“knowledge discovery”) OR (“deep learning”) OR (“supervised learning”) OR (“unsupervised learning”) OR (“random forest”) OR (“support vector”) OR (“SHAP”) OR (“SHapley Additive exPlanations”) OR (“LIME”) OR (“Local Interpretable Model-Agnostic Explanations”)) AND (“Eur J Pain”[Journal]), 16 such articles, after manual curating the hits, have now been published in the journal. This accumulation of studies justifies assembling them into a virtual issue. This dedicated collection will provide readers interested in machine learning with easier access to relevant work published in the EJP and serve as a convenient resource for exploring this rapidly developing field within pain research.

Machine learning encompasses various methodological approaches that can be roughly categorised as supervised, unsupervised, or reinforcement learning. Supervised learning involves training algorithms to map input data (features) to predefined output labels (MacQueen 1967). It is often used for diagnostic classification tasks, where known cases (e.g., patients versus healthy individuals) are used for training and subsequent prediction. Of note, most supervised algorithms can be used for both, classification and regression. By contrast, unsupervised learning is designed to discover hidden structures or patterns within data without prior labeling (Steinhaus 1956). It employs techniques such as clustering or dimensionality reduction to explore complex datasets. Reinforcement learning is a third major paradigm in which algorithms learn optimal actions through interactions with an environment. They are guided by rewards or penalties and are increasingly used in areas such as treatment strategy modelling (Sutton and Barto 2018). Additional approaches, including natural language processing and ontology-based knowledge discovery, further expand the scope of machine learning in biomedical research.

A range of the machine learning techniques outlined above have been utilised in publications in this journal. To accompany this virtual issue, we reviewed relevant publications and summarised the applied techniques in Table 1. This overview provides insight into the diversity of machine learning approaches used in pain research from the European Journal of Pain's perspective. It also highlights the most frequently applied methods. Although the number of papers is limited, the range of methodologies aligns well with the findings of a recent, broader review of machine learning studies related to pain (Lötsch et al. 2022). Additionally, this summary may help identify specialised methodological applications that are not always explicitly detailed in abstracts. These applications may require consulting the full texts, for which Table 1 may offer some key methodological details. Finally, it should be noted that the purpose of this special issue is to collect papers on the use of machine learning in pain research or clinical contexts. In contrast, the use of LLMs for article preparation is already addressed in the updated instructions for authors of this journal https://www.wiley.com/en-us/publish/book/ai-guidelines.

The publications in this virtual issue highlight the variety of machine learning applications in pain research and their growing clinical relevance. Supervised learning, particularly classification and regression, is commonly employed to predict chronic pain recurrence, identify pain conditions using neuroimaging and clinical data, and forecast opioid analgesic response based on EEG biomarkers. Unsupervised methods, such as clustering, help to subgroup patients by pain phenotypes and cognitive biases. Explainable AI enhances the interpretability of these methods. Although reinforcement learning is less common, it holds potential for optimising personalised treatments. Deep learning and computer vision methods automate pain detection through facial expression analysis, and natural language processing supports conversational agents for patient self-management. Machine learning also supports the analysis of attentional and interpretive biases linked to the development of chronic pain, while uncovering biological mechanisms through genetic and sensory testing data. Collectively, these studies highlight the growing role of machine learning (ML) in enhancing prediction, diagnosis, and treatment strategies within pain research.