Browse the corpus

Pain is an unpleasant sensory and emotional experience associated with actual or potential tissue damage. It is subjective, shaped by biological, psychological, and social factors, and thus cannot be read off from only tissue damage. Pain is distinguished from nociception (the mechanism via which noxious stimuli are encoded) because nociception can occur without pain and pain without nociception (). So, a signal that tracks noxious stimuli is not always a measure of the pain a person might experience. Pain is a major cause of disability. If left untreated, it can lead to negative personal and social outcomes, such as depression and work absenteeism. It also imposes unnecessary pressure on families and caregivers but also urges individuals to seek continuous medical help from many physicians, especially when appropriately pain controlled is not achieved (, ). Pain is usually detected, assessed, and monitored by communication with patients using history taking and different pain assessment scales. However, some patients, such as infants, Alzheimer's patients, intubated and mechanically ventilated patients, and unconscious patients, cannot self-report pain through oral or written communication. Therefore, it is challenging to assess pain in this group of patients and make a further diagnosis. Critically, the absence of self-report in such groups removes the reference standard for training and validation of automated systems. Defining a trustworthy ground truth — a patient's self-reported pain — is a central challenge for pain evaluation based on AI. Hence, other methods should be considered for pain assessment. In such cases, technological support can be helpful. However, despite advancements in medical technology, there is still a significant need for accurate pain detection and management. Recent progress in brain imaging techniques has resulted in neurocomputational models that utilize neurological signals to classify different pain states. These models analyze changes in the structure and function of the brains of chronic pain patients (, ). Electroencephalogram (EEG) data-based models have associated specific neuronal activities in the motor cortex with various pain states. EEG is a non-invasive, cost-effective, and widely available method for predicting brain function abnormalities and extracting markers of chronic pain ().

Chronic pain is conventionally determined as pain persisting for more than three months. It is not an entirely sensory signal but a biopsychosocial condition. Nociceptive (biological) input interacts with affective, cognitive, and social processes. Artificial-intelligence (AI) systems have almost entirely worked with the biological dimension — electroencephalographic (, , , ), facial (–), neuroimaging (, , ), and physiological correlates of pain (–). This is because such signals can be measured and labelled. The psychosocial dimension, which determines how pain is experienced and reported, is largely unaddressed by current models. Therefore, multimodal systems combining several biological domains based on patient-reported and behavioral data are needed. Chronic pain disrupts the adaptive integration of sensory and contextual processes and is associated with significant structural and functional changes in the brain. EEG has been shown to effectively track these changes. However, existing EEG studies on chronic pain-related cortical activity have yielded inconsistent results (, ). Some studies have reported increased theta and gamma oscillations in chronic pain patients, while others have observed reductions in alpha and beta power. Further research is needed to establish an efficient brain-based marker for chronic pain (, ).

Machine learning can effectively improve the pain-related neuroimaging analysis. It can uncover patterns in clinical and experimental data, providing valuable information for acquiring new knowledge and understanding the complexity of pain. Convolutional Neural Network (CNN), known for its success in computer vision and speech recognition, has shown potential for analyzing EEG signals due to its ability to handle large datasets and exploit hierarchical structures in natural signals (). Several studies have explored the application of CNN in EEG analysis for various purposes, achieving impressive results (). However, there is limited work on using CNN for EEG-based chronic pain detection, which is a challenging but important task in pain diagnosis and treatment. The goal of this scoping review was to synthesize and analyze the available evidence related to the application of AI in pain research and make relevant clinical and scientific recommendations.

The main families of AI used in the studies reviewed are defined here: Classical machine learning (e.g, random forests and logistic regression) analyzes predefined, hand-engineered features and is more effective on structured, tabular information (laboratory values or questionnaires). On the other hand, deep learning learns features from raw data via multi-layer neural networks. Its subtypes can be used with different signal types: convolutional neural networks (CNNs) can read images and image-like inputs (e.g., facial photographs, MRI, and spectrograms); recurrent networks with long short-term memory (LSTM) and bidirectional LSTM (BiLSTM) are used for time series such as EEG or physiological traces; graph neural networks (GNNs) can analyze data with an explicit relational structure, for instance, facial-landmark graphs; and transformer-based large language models (LLMs) can read text, which is a basis for conversational and patient-education programs. Multimodal/fusion models mix these streams. Since pain is multidimensional, integration of complementary signals is better matched to the construct in comparison to a single-signal approach.

The initial search identified 1120 articles. After removing duplicates, 291 records were screened for eligibility. After careful review, 47 articles were considered suitable for inclusion (, –, –) (). The characteristics of included studies are described in the (). describes in detail the AI method used and its purpose. AI algorithms and models are described in . Potential applications of artificial intelligence in pain management are summarized in , and Parameters and data sources used for AI-based pain detection and assessment are summarized in .

Sixteen studies included in this review explored AI-driven approaches for better pain detection. Among them, six focused on enhancing pain detection through facial expression analysis and one on head movement. In 2019, Menchetti developed an Automated Facial Expression Recognition (AFER) system for detecting pain-related Action Units (AUs) in non-verbal patients. Compared to the Emotient system, AFER demonstrated superior accuracy and F1-scores, successfully identifying 8 out of 9 pain-associated AUs (). Bargshady (2020) introduced an enhanced deep neural network that couples a convolutional neural network with a bidirectional long short-term memory (BiLSTM) network (the EJH-CNN-BiLSTM model) for multi-class pain intensity detection from facial expression images, achieving an average accuracy of 91.2% on the training set, 90% on the testing set, and 98.4% AUC (). Later, Patania et al. (2022) proposed a graph neural network (GNN) that utilized the graph structure of facial fiducial points. On the BioVid Heat Pain database (the model was given to distinguish no-pain vs. highest-intensity pain; the fairness was tested by leave-one-subject-out cross-validation), it surpassed standard models, reaching 73.2% accuracy. This figure reflects genuine (stimulus-evoked pain) rather than acted expression and was considered a state-of-the-art result for that benchmark at the time (). In the same year, Fontaine et al. (2022) trained a DL system (ResNet-18) on 2,810 images from 1,189 patients, achieving high sensitivity (89.7%) for detecting severe pain (≥4/10) and outperforming nurses, who identified pain with a lower sensitivity (44.9%) (). Fabi et al. (2022) investigated racial bias in facial action-unit (FAU) detection. A model that was predominantly trained on European faces analyzed synthetic European and African faces in the study. This bias was not shown to be uniform. Although European faces were favored, several action units were readily detected on African faces too. This shows that the visual detection of individual muscle activation may vary with skin tone or race-dependent features. The authors concluded that greater exposure to own-race faces and perceptual detection ability do not explain the racial bias in pain recognition (). Recently, in a feasibility study, Cascella et al. (2024) evaluated YOLOv8 object detector on a custom pain/no-pain dataset using facial expression images ( n = 500). The precision showed 0.868, however, a small balanced subset may not reflect clinical diversity (). Notably, complementing facial cues, Semwal and Londhe (2021) previously proposed an automated pain detection method using head movement dynamics, leveraging a CNN-BiLSTM network to learn spatial and temporal dependencies. Evaluated on the UNBC-McMaster dataset, this approach outperformed facial expression-based methods, suggesting that head pose analysis may be a more reliable indicator of pain and a promising alternative for automated pain assessment ().

Three studies explored AI-based methods in orofacial pain. In 2019, Hu et al. developed CLARAi, a mobile neuroimaging-based augmented reality and AI framework for real-time pain detection and localization. Functional near-infrared spectroscopy was used to measure cortical activity in 21 patients experiencing tooth pain, with a three-layer neural network achieving 80.37% accuracy for pain classification and a six-layer convolutional neural network reaching 74.23% accuracy for pain localization (). Later, Kreiner and Viloria (2022) introduced a multilayer perceptron neural network for diagnosing orofacial pain and temporomandibular disorders. Their model outperformed 12 general dental clinicians ( p = .0072), particularly in detecting non-dental and referred pain (). In the same year, Jerritta et al. leveraged machine learning for pain detection based on facial geometry, focusing on mouth area features. Using the Biovid Heat Pain database, KNN and Regression Tree classifiers achieved 95%–98% accuracy, with KNN proving superior. Notably, geometric and ratio-based features were the most effective for pain assessment ().

Four studies focused on automated pain detection in particular groups or sensors. Lopez-Martinez et al. (2019) utilized functional near-infrared spectroscopy (fNIRS) signals from 43 healthy participants to detect pain using Bayesian hierarchical modeling with wavelet features. Their model identified pain presence with 69% accuracy, based on cortical hemoglobin concentration changes (). Detection of procedural pain in neonates with intensity discrimination was analyzed by Giordano et al. (2024). The study analyzed an AI facial-analysis system (FaceReader 9, “Baby Face” model). The results correlated strongly with experts' VAS (Visual analogue scale) (). Post-surgical pain assessment was conducted by Sada et al. (2024). The AI-based PainChek Infant tool enabled parents to make informed decisions about whether to apply intervention or not. This led to reduced pain in 94.5% of assessments (). Snene et al. (2024) tested the feasibility of continuous AI tracking of pain intensity (Empatica E4 wristbands) in patients with dementia. LightGBM algorithm was found to have the highest accuracy (0.74) ().

Two studies aimed to optimize the data inputs for more accurate pain detection. Prajod et al. (2022) used Explainable AI to analyze differences between pain detection models trained on experimental (BioVid) and clinical (UNBC) pain datasets. While both convolutional neural networks achieved similar within-dataset accuracy, the clinical pain model demonstrated superior generalizability, recognizing pain with 85% accuracy compared to 66% for the experimental model (). In the same year, Jerritta (2022 b) developed a machine learning-based pain detection method utilizing EMG and HRV signals from the Biovid Heat Physiological Signal Database. After preprocessing with Butterworth and notch filters, Daubechies wavelets were applied for feature extraction. As a result, KNN outperformed Regression Trees in detecting pain levels, with EMG signals emerging as the most effective for pain recognition. Based on the presented findings, advanced AI-driven methods have the potential to improve the accuracy and reliability of automated pain detection, particularly across different age and racial groups ().

Summarizing all the studies, two main observations can be highlighted. Firstly, facial expression is considered the most mature modality. Deep models that are trained on facial action units either matched or exceeded clinical expectations, with Fontaine et al. reaching 89.7% sensitivity for severe pain in comparison to 44.9% for bedside nurses () and the network of Kreiner and Viloria, which outperformed twelve dentists (). The second observation is that the accuracy is significantly dependent on the data source and not only on the algorithm alone. For example, Prajod et al. demonstrated that models trained on clinical pain from the UNBC-McMaster archive generalise better than those trained on experimentally induced pain from BioVid (85% vs. 66%) (). This may explain why accuracy in the section may fluctuate from roughly 73% on BioVid () to the mid-90s on UNBC-derived data (). A recurring caution is fairness, as Fabi et al. showed that the detectability of pain-related facial cues may vary by itself with skin tone (), thus, a model validated on one population cannot be referred to another.

Ten studies have proposed AI-driven methodologies to improve pain prediction. Four studies were conducted across different modalities. In 2015, Gram et al. applied ML to EEG data recorded during cold pain stimulation to predict individual morphine analgesic response, achieving 72% accuracy using SVM models trained on spectral features (). In 2020, Wang et al. proposed an autoencoder model based on convolutional neural networks (EEGNet) for feature extraction of pain-related EEG signals, which had a superior accuracy of 74.6 ± 11.2% in pain level prediction compared to conventional methods (). The following year, Hsiao et al. utilized multimodal neurophysiological and psychometric signatures to predict individual heat pain sensitivity, achieving up to 80% prediction accuracy using SVM models trained on resting-state EEG features (). Also in 2021, Han et al. achieved a notable increase in precision, attaining 92.54% accuracy and a correlation of 0.84 between real and predicted pain levels by using pre-surgery EEG recordings and an LDA model to predict postoperative pain intensity ().

Four studies were conducted in condition-specific states. Soin et al. (2022) used machine learning algorithms to predict diagnoses related to spinal pain by analyzing 85 specific data points. The decision tree software achieved a 72% accuracy rate in predicting the diagnosis based on the reported symptoms. The study suggests that AI-driven methods have the potential to enhance pain prediction in clinical settings, particularly in analyzing low-level visual information (). Huang et al. (2025) also studied low back pain and found that Multivariate Logistic Regression (MLR) within a Multidimensional Model (MDM) framework achieved AUC = 0.813 for future recurrence risk forecasting and outperformed the standard STarT BACK Tool (). Muhaimil et al. (2024) specifically used T2-weighted MRI of the lumbar spine to predict low back pain, where the best deep learning was GoogLeNet with 0.85 accuracy (). Shahid et al. (2024) used age and sex together with serum biochemistry, including uric acid, C-reactive protein, creatinine, and other blood and liver-function parameters, to predict joint pain by comparing several models. The best-performing model at 97.44% accuracy was a random forest, which combined bootstrap sampling, random feature selection at each split, and tree pruning. Uric acid was identified as the strongest single predictor ().

Two studies reviewed AI for prediction after treatment. One study, Areias et al. (2024), used Light Gradient Boosting Machine (LightGBM) and Recurrent Neural Network (RNN) to predict treatment response using time-series data such as range of motion (ROM), exercise performance, psychological scores, and social deprivation indices. Both models achieved similar performance with an AUC of 0.70 (). Another study by Cheema et al. (2025) studied chronic low back pain to predict which patients are associated with better or worse pain-related rehabilitation outcomes. The study applied Bootstrap Random Forest (RF) supervised machine learning, where the Random Forest model achieved an overall 75.6% predictive accuracy ().

Based on prediction studies, two distinctive targets can be traced. First is the forecasting of intrinsic pain sensitivity or analgesic response, which analyses the neurophysiology. For instance, Gram et al. forecasted morphine response from pre-stimulus EEG (about 72%) (), Hsiao et al. predicted heat pain sensitivity (80%) (), and Han et al. reached 92.5% for acute post-operative pain (). The second target implies prediction of clinical outcomes like chronicity, recovery, and joint pain based on routinely collected information (, –). More modest models are clinically applicable and externally oriented — Areias et al. reported an AUC near 0.70 for a musculoskeletal recovery model (). In contrast, the prediction value is explicit, where it outperforms the existing tool, as in Huang et al., whose multidimensional model was more effective than the STarT Back screening tool (0.813 vs. 0.555) (). Nevertheless, prediction is considered the least represented domain due to the gap between controlled and externally validated settings.

Eighteen studies have introduced AI-driven methodologies to improve pain classification. Eight studies advanced pain classification using facial expression analysis with ML and DL models. Morabit et al. developed a data-efficient image transformers and distillation (DeiT) model for binary pain recognition, achieving 84.15% accuracy on the UNBC-McMaster dataset and 72.11% on the BioVid dataset, outperforming the pre-trained GoogleNet model (). Similarly, Ghosh et al. introduced a sentiment analysis system to classify pain intensity levels, reaching 82.54% and 83.71% accuracy for 2- and 3-class UNBC-McMaster databases and 75.67% on the 2D Face-set with Pain-expression dataset (). Bargshady et al. introduced an ensemble deep learning model (EDLM) for pain level classification, integrating a three-stream hybrid deep neural network. This model achieved 89% accuracy and an AUC of 0.93 on the Multimodal Intensity Pain (MIntPain) database, surpassing other deep learning methods for multi-level pain detection (). Meanwhile, Gomutbutra et al. applied five ML models to classify pain in chronic pain patients using FAUs, with SVM incorporating FAUs and gender achieving the highest accuracy of 58% in the Asian elderly population (). Kristian et al. developed a multimodal ML-based system for infant pain classification using facial expressions and voice features, with ensemble models outperforming single-modality approaches based on F1 scores (). Xu et al. improved pediatric pain classification by applying transfer learning to map automatically detected FAUs from iMotions software to manually coded AUs, increasing the AUC from 0.67 to 0.72 (). Kelati et al. explored facial electromyograms (fEMG) for pain level classification, with KNN achieving 99.4% accuracy, indicating that EMG signals from the corrugator and zygomaticus muscles reliably differentiate pain from non-pain conditions (). Lastly, de Araujo et al. (2024) used ChatGPT-4 and built-in Python/Streamlit to classify odontogenic and temporomandibular pain. AI achieved 86% accuracy ().

Eight studies advanced pain classification focusing on specific conditions. In 2021, Pikulkaew et al. applied DL techniques to detect shoulder pain using sequential facial images, achieving an impressive accuracy of 98.75% for no-pain and 95.05% for pain classification, compared to doctor diagnoses. However, the classification accuracy for initial pain levels dropped to 75.49%, primarily due to an imbalanced dataset. Despite this, the technique performed well in scenarios where continuous monitoring of patients for over 24 h is not feasible (). Similarly, Piri et al. (2022) applied AI-based classification methods to PET/CT scan data, distinguishing regions of interest based on fluorodeoxyglucose and sodium fluoride uptake. The results indicated heightened tracer uptake in the facet joints and intervertebral discs of individuals with low back pain, suggesting inflammation and degenerative changes as potential sources of nociceptive pain (). Kimura et al. (2021) developed a wearable EEG system to assess pain during movement, addressing the limitations of resting-state evaluations. Using a single-electrode device and machine learning, the system successfully classified hip pain with 79.6% accuracy (). Chen et al. (2022) utilized deep convolutional neural networks (CNN) to classify pain states from EEG signals in 10 chronic back pain patients, achieving an AUC of 0.83 ± 0.09 during movement stimulation and 0.81 ± 0.15 during video stimulation, surpassing existing methods in accuracy (). He et al. (2025) focused on a deep learning Convolutional Neural Network (CNN) model in patients with lumbar disc herniation (LDH). This model achieved 93.4% accuracy, which can automatically detect LDH (). Mao et al. (2025) aimed to use a YOLOv10 object-detection–style deep network for the detection of temporomandibular joint degenerative disease, with 98% accuracy reported for the classification of radiographic signs in an image (). The same condition detection was studied by Choi et al. (2025), where ResNet-18/50/101, VGG-16/19, and GoogLeNet outperformed eight orofacial pain specialists on a 200-image reader set (). A Convolutional Neural Network using the YOLO (You Only Look Once) architecture (CNN-YOLO) was studied by Liawrungrueang et al. (2024) for the detection of lumbar intervertebral disc degeneration. The model achieved 97% accuracy in training for implicit grading of discogenic back-pain severity ().

Two studies advanced pain classification across different contexts. In 2021, Hsiao et al. utilized multimodal neurophysiological and psychometric signatures to classify individual heat pain sensitivity, achieving up to 93.8% classification accuracy using support vector machine (SVM) models trained on resting-state EEG features (). Later, Wu et al. (2022) presented a deep learning-based approach for objective pain intensity assessment in non-verbal patients, utilizing multi-channel EEG signals to learn invariant spatial-spectral-temporal representations. By employing an Attentive-Recurrent-Convolutional Neural Network (ARCNN), the model achieved 92% accuracy in pain classification in a subject-independent scenario (). All of the above studies provide evidence that AI-based methods could significantly refine pain classification workflows, particularly through neuroimaging, with the potential for real-time applications.

The classification studies all show the same dividing line. Behavioral classifiers of pain expression based on facial, EEG, or EMG signals show wide variation in accuracy (58%–99%). Kelati et al. reached 99.4% on EMG (), while Gomutbutra et al. achieved 58% on a clinical facial action unit sample (). This tracks the dataset quality and class balance. In contrast, structural classifiers that may label the anatomical lesion from MRI, CT, PET, or CBCT are more consistent, reaching clustering at 93%–98% (–). Such an observation may, however, be misleading for pain assessment because degenerative changes are commonly asymptomatic. So, grading disc degeneration or temporomandibular pathology is a classifying structure, not the pain experience itself.

Three articles reported AI accuracy for remote patient monitoring, patient education, and counseling. Ah-Yan et al. (2024) studied pain management support ChatGPT (GPT-3.5) and Copilot (GPT-4) for the generation of safe, valid, and useful advice for low-back pain. On a 4-point Likert scale, AI provided 3.33 for validity, 3.19 for safety, and 3.60 for usefulness. This demonstrated good performance (). Granviken et al. (2025) described how AI-driven personalized decision support can be used to manage musculoskeletal pain using multilevel mixed logistic models. The results were insignificant [measured by Odds ratio (OR)] (). Zapata-Caballero et al. (2025) evaluated patient education by ChatGPT 3.5 by asking common patient questions. It achieved the highest total scores in lifestyle and misconceptions (14/15), lowest in support groups 9/15 and further reading 10/15 (). One study evaluated the reliability of generative AI outputs. Temel et al. (2025) assessed whether ChatGPT-4o can consistently and accurately draw L4/L5/S1 lumbar radicular pain patterns. The results indicated poor intra-model consistency ().

AI has been used for pain management in a wide variety of diseases and conditions, such as acute postoperative pain (, ), chronic pain (, ), hip pain (), low back pain (including lumbar radiculopathy, post-laminectomy syndrome, sacroiliitis, lumbar disc herniation, lumbar intervertebral disc degeneration, and degenerative lumbar spine diseases) (, –, , , , , ), shoulder pain (, , , , ), temporomandibular disorders (TMD) and orofacial pain (, , , ), odontogenic pain (), hypersensitive pain (), joint pain (), chronic pelvic pain (), pain associated with dementia and neonatal infant procedures (, , ). Additionally, AI was used to assess pain sensitivity, anesthesia depth, and post-operative pain.

The studies were clustered into the following main thematic groups: Automatic identification of the presence of pain or its intensity: This is the biggest group that utilized EEG-based neural signatures and oscillations, facial expressions and action units (AUs), physiological signals from EMG and ECG, video-based data, and data based on multimodal behavior. This group of AI algorithms allows unbiased real-time pain recognition with high accuracy. Diagnosis of structural causes of pain: This group identifies the pathology of pain based on MRI, PET, and CT images. Such systems allow for increasing diagnostic accuracy, speed, and clinical decision-making. Prediction of future pain: This group supports treatment planning and stratification of risks in predicting chronic pain, responder vs. non-responder classification, and response to therapies. Management and monitoring: This group aims to improve clinical procedures such as analgesia, remote patient monitoring, physiotherapy, and neuroimaging of pain localization. Conversational AI: This group expands AI as a tool for patient counseling, differential screening, and education in pain medicine.

Across the included studies tabulated, AI demonstrated a potential to measure pain objectively and non-invasively. For instance, EEG-based algorithms can detect pain via brain signatures with high sensitivity in comparison to traditional methods. Machine learning and deep learning can consistently provide high accuracy in classification and prediction. There is 80% accuracy for pain sensitivity prediction, and 90% accuracy in post-operative and facial expression-based detection. Likewise, imaging-based and multimodal systems have shown more than 90% diagnostic accuracy. AI can also be efficient in terms of computational performance. For example, autoencoders can decrease training time, kernel-based methods can minimize memory storage, and cross-validation can improve reliability and reduce overfitting. Additionally, lightweight models like linear discriminant analysis are suitable for clinical environments by being fast, portable, and easily deployable. Another benefit of AI is that it can extract data from raw signals that may be difficult to capture with traditional methods. CNNs, transformers, and hybrid architectures can learn to detect pain from EEG, video, and physiological data. The performance may be further improved by combining neuroimaging, facial expressions, and physiological signals. In cases where self-reporting is unavailable (infants, unconsciousness, and cognitive impairment), AI can still evaluate pain based on crying, facial expressions, or physiological monitoring. In addition, AI helps clinical decision-making by providing real-time monitoring and guidance. Overall, AI's main strength is that it can reduce bias and standardize the pain assessment. It may remove observer and racial biases to allow fairer assessment.

Despite the benefits, the included studies consistently report the challenges that can constrain AI's clinical implementation. The first challenge is the lack of generalizability due to small sample sizes and limited population diversity. Many models were trained on controlled and homogeneous data. Another challenge is the issue of scalability. Deep learning approaches' effectiveness could have been decreased due to small datasets, which can lead to overfitting. Next, multimodal systems, cross-validation, and training duration require resources that might be costly and unfeasible for clinical settings. There can also be specific challenges: in the case of facial expression analysis, there can be an overlap with other emotional states. Misclassification rates, even low, can significantly affect the decision-making in medication dosing and critical care, making AI not yet suitable in high-risk settings. Head-movement-based models can be erroneous in the case of minimal head movements in patients. Finally, it should be noted that many models are experimental and newly proposed.

This review highlighted the recent results from experiments and observational studies of artificial intelligence (AI) applications in pain medicine. Overall, the results demonstrated consistent performance of AI—specifically deep learning architectures—on pain detection, prediction, and classification (, , ). Deep learning achieved high accuracy in controlled settings for objective pain evaluation (, , ). Nevertheless, the evidence revealed important limitations and challenges. Beyond a study-by-study account, the included studies can be mapped onto the five functional roles described in section . The most populated role is the automatic identification of the presence or intensity of pain drawn from EEG, physiological signals, and facial expressions (, –, , ). This may suggest that accurate pain detection may require integrated behavioral, multidimensional, and neurophysiological representation. The next role is diagnosis of the structural cause of pain. It is mostly based on imaging models (, , , ). The third role is relatively rare, which is pain forecasting for stratification and treatment planning (, –). Management and monitoring are the fourth role that covers remote tools and systems for decision support (, ). Finally, the fifth role is conversational AI used for education and counselling (, ). Such AI is commonly built on large language models.

This mapping of five roles uncovers six critical knowledge gaps that hinder clinical implementation. First is the problem of the ground truth against which an AI system trains and judges. The label is a self-report scale, like VAS or NRS, if a patient can communicate. As mentioned in the introduction, the populations of sedated, intubated, neonates, or those with dementia may not self-report, and studies use a surrogate. Two surrogates are mostly used: expert observation and the noxious stimulus. Expert observation utilizes validated behavioral scales. It can be labels assigned by the clinician or by an observational VAS scale (–). The noxious stimulus can be used itself in the calibrated heat of the BioVid database, for example. In this case, a controlled stimulus is applied to measure the body's involuntary reaction. These two surrogates possess distinctive cons. The model may inherit subjectivity and inter-rater reliability from a clinician, thus imitating rather than improving the judgement. Stimulus-linked models may conflate nociceptive input with perceived pain, which means they may not always capture pain correctly. This creates a paradox where the ground truth is the weakest in the populations that require automated assessment the most. Thus, reported accuracies in this population should be interpreted together with the reference standard used quality.