Artificial intelligence (AI), defined as intelligence operated by a machine, has been introduced to share/replace the work requiring human intelligence. Concerning the medical field, human intelligence is necessary for the selection of the most possible diagnosis, the provision of the most appropriate assessment, and the determination of the most relevant therapeutic strategy. Herewith, it is common for two clinicians to give two distinct treatment decisions for the very same clinical scenario. In such cases, in general, the suggestions of the senior colleague are taken into consideration. This refl ects the fact that the medical profession needs time to develop based on its complexity. Therefore, the medical service providers always ask whether AI can learn from the most experienced clinical practitioners for minimizing the training cost.

A main subset of AI is machine learning, denoting a model that can automatically learn and improve from given data. Deep learning, a subdivision of machine learning, incorporates deep neural networks for model training. Deep neural networks simulate relays of human neurons which have inputs from multiple sources and yield a fewer number of outputs. Several layers of neurons are interposed between the input and output, modifying the weight of each input and thus giving a summarized value to the output (Figure 1). Deep learning algorithms allow automatic extraction of the embedded features without exactly telling the computer what features need to be retrieved. Therefore, deep learning is appropriate for implementing complex operations.

How can AI be applied in pain medicine? A simple literature search of PubMed by using the combination of “artifi cial intelligence” and “pain” yields the following examples. Pedoia et al. [1] collected T2 weighted magnetic resonance images (MRIs) of knees from 4,384 participants. They proved the feasibility of voxel-based relaxometry in a combination of densely connected neural network to differentiate patients with and without osteoarthritis. Fraiwan et al. [2] collected whole spine radiographs from 338 participants with scoliosis (n = 188), spondylolisthesis (n = 79), and normal spine (n = 71). By using the deep transfer learning model, the maximum accuracy of three-class classification could reach 98.02%. Ito et al. [3] collected 217 MRIs of the temporomandibular joints from 20 participants. Their proposed convolutional neural network was capable of segmenting the articular disc with acceptable performance metrics. Kim et al. [4] included 180,271 lumbar radiographs from 34,661 patients with recent lumbar MRIs. By using a deep learning-based algorithm, the area under the curve of the receiver operating characteristic for predicting lumbar herniated nucleus pulposus was up to 0.73. Maraş et al. [5] collected lateral cervical radiographs from 416 patients and applied the transfer learning method to discriminate participants with normal spine from those with pathologies (e.g., loss of cervical lordosis, narrowing of the disc space, or degenerative vertebral changes). They found that a pre-trained VGG-16 network had better performance than other models concerning accuracy (93.9%), sensitivity (95.8%), specificity (92.0%), and precision (92.0%) of classification. Wu et al. [6] enrolled 746 video clips from 64 critically ill patients. The use of VGG-16 network could dichotomize facial expression of pain with an accuracy ranging from 0.81 to 0.88.

Through a simply survey of recent literature, we are aware that, in the fi eld of pain medicine, AI can be applied to classification of pathological images and facial expressions as well as segmentation of target structures. Until now, there are not many studies investigating the usefulness of deep learning algorithm for the prediction of treatment responses in various painful syndromes. As patients’ data have mostly been digitalized, we believe that more extensive application of AI on outcome prediction and diagnosis/grading of painful pathologies would emerge in the near future.

Figure 1. Illustration of a Deep Learning Neural Network

Each author certifi es that he or she has no commercial associations.