Cracking the Chronic Pain code: A scoping review of Artificial Intelligence in Chronic Pain research

OBJECTIVE

The aim of this review is to identify gaps and provide a direction for future research in the utilization of Artificial Intelligence (AI) in chronic pain (CP) management.

METHODS

A comprehensive literature search was conducted using various databases, including Ovid MEDLINE, Web of Science Core Collection, IEEE Xplore, and ACM Digital Library. The search was limited to studies on AI in CP research, focusing on diagnosis, prognosis, clinical decision support, self-management, and rehabilitation. The studies were evaluated based on predefined inclusion criteria, including the reporting quality of AI algorithms used.

RESULTS

After the screening process, 60 studies were reviewed, highlighting AI's effectiveness in diagnosing and classifying CP while revealing gaps in the attention given to treatment and rehabilitation. It was found that the most commonly used algorithms in CP research were support vector machines, logistic regression and random forest classifiers. The review also pointed out that attention to CP mechanisms is negligible despite being the most effective way to treat CP.

CONCLUSION

The review concludes that to achieve more effective outcomes in CP management, future research should prioritize identifying CP mechanisms, CP management, and rehabilitation while leveraging a wider range of algorithms and architectures.

SIGNIFICANCE

This review highlights the potential of AI in improving the management of CP, which is a significant personal and economic burden affecting more than 30% of the world's population. The identified gaps and future research directions provide valuable insights to researchers and practitioners in the field, with the potential to improve healthcare utilization.

Identification of Myofascial Trigger Point Using the Combination of Texture Analysis in B-Mode Ultrasound with Machine Learning Classifiers

Myofascial pain syndrome is a chronic pain disorder characterized by myofascial trigger points (MTrPs). Quantitative ultrasound (US) techniques can be used to discriminate MTrPs from healthy muscle. In this study, 90 B-mode US images of upper trapezius muscles were collected from 63 participants (left and/or right side(s)). Four texture feature approaches (individually and a combination of them) were employed that focused on identifying spots, and edges were used to explore the discrimination between the three groups: active MTrPs (n = 30), latent MTrPs (n = 30), and healthy muscle (n = 30). Machine learning (ML) and one-way analysis of variance were used to investigate the discrimination ability of the different approaches. Statistically significant results were seen in almost all examined features for each texture feature approach, but, in contrast, ML techniques struggled to produce robust discrimination. The ML techniques showed that two texture features (i.e., correlation and mean) within the combination of texture features were most important in classifying the three groups. This discrepancy between traditional statistical analysis and ML techniques prompts the need for further investigation of texture-based approaches in US for the discrimination of MTrPs.

Selective peripheral nerve recording using simulated human median nerve activity and convolutional neural networks

BACKGROUND

It is difficult to create intuitive methods of controlling prosthetic limbs, often resulting in abandonment. Peripheral nerve interfaces can be used to convert motor intent into commands to a prosthesis. The Extraneural Spatiotemporal Compound Action Potentials Extraction Network (ESCAPE-NET) is a convolutional neural network (CNN) that has previously been demonstrated to be effective at discriminating neural sources in rat sciatic nerves. ESCAPE-NET was designed to operate using data from multi-channel nerve cuff arrays, and use the resulting spatiotemporal signatures to classify individual naturally evoked compound action potentials (nCAPs) based on differing source fascicles. The applicability of this approach to larger and more complex nerves is not well understood. To support future translation to humans, the objective of this study was to characterize the performance of this approach in a computational model of the human median nerve.

METHODS

Using a cross-sectional immunohistochemistry image of a human median nerve, a finite-element model was generated and used to simulate extraneural recordings. ESCAPE-NET was used to classify nCAPs based on source location, for varying numbers of sources and noise levels. The performance of ESCAPE-NET was also compared to ResNet-50 and MobileNet-V2 in the context of classifying human nerve cuff data.

RESULTS

Classification accuracy was found to be inversely related to the number of nCAP sources in ESCAPE-NET (3-class: 97.8% ± 0.1%; 10-class: 89.3% ± 5.4% in low-noise conditions, 3-class: 70.3% ± 0.1%; 10-class: 52.5% ± 0.3% in high-noise conditions). ESCAPE-NET overall outperformed both MobileNet-V2 (3-class: 96.5% ± 1.1%; 10-class: 84.9% ± 1.7% in low-noise conditions, 3-class: 86.0% ± 0.6%; 10-class: 41.4% ± 0.9% in high-noise conditions) and ResNet-50 (3-class: 71.2% ± 18.6%; 10-class: 40.1% ± 22.5% in low-noise conditions, 3-class: 81.3% ± 4.4%; 10-class: 31.9% ± 4.4% in high-noise conditions).

CONCLUSION

All three networks were found to learn to differentiate nCAPs from different sources, as evidenced by performance levels well above chance in all cases. ESCAPE-NET was found to have the most robust performance, despite decreasing performance as the number of classes increased, and as noise was varied. These results provide valuable translational guidelines for designing neural interfaces for human use.

Enhancing the selective electrical activation of human vagal nerve fibers: a comparative computational modeling study with validation in a rat sciatic model

Objective.Vagus nerve stimulation (VNS) is an emerging treatment option for a myriad of medical disorders, where the method of delivering electrical pulses can vary depending on the clinical indication. In this study, we investigated the relative effectiveness of electrically activating the cervical vagus nerve among three different approaches: nerve cuff electrode stimulation (NCES), transcutaneous electrical nerve stimulation (TENS), and enhanced TENS (eTENS). The objectives were to characterize factors that influenced nerve activation and to compare the nerve recruitment properties as a function of nerve fiber diameter.Methods.The Finite Element Model, based on data from the Visible Human Project, was implemented in COMSOL. The three simulation types were compared under a range of vertical and horizontal displacements relative to the location of the vagus nerve. Monopolar anodic stimulation was examined, along with latency and activation of different fiber sizes. Nerve activation was determined via the activating function and McIntyre-Richardson-Grill models, and activation thresholds were validated in anin-vivorodent model.Results.While NCES produced the lowest activation thresholds, eTENS generally performed superior to TENS under the range of conditions and fiber diameters, producing activation thresholds up to three times lower than TENS. eTENS also preserved its enhancement when surface electrodes were displaced away from the nerve. Anodic stimulation revealed an inhibitory region that removed eTENS benefits. eTENS also outperformed TENS by up to four times when targeting smaller diameter nerve fibers, scaling similar to a cuff electrode. In latency and activation of smaller diameter nerve fibers, eTENS results resembled those of NCES more than a TENS electrode. Activation threshold ratios were consistent inin-vivovalidation.Significance.Our findings expand upon previously identified mechanisms for eTENS and further demonstrate how eTENS emulates a nerve cuff electrode to achieve lower activation thresholds. This work further characterizes considerations required for VNS under the three stimulation methods.

Myofascial Trigger Point Identification in B-Mode Ultrasound: Texture Analysis Versus a Convolutional Neural Network Approach

OBJECTIVE

Myofascial pain syndrome (MPS) is one of the most common causes of chronic pain and affects a large portion of patients seen in specialty pain centers as well as primary care clinics. Diagnosis of MPS relies heavily on a clinician's ability to identify the presence of a myofascial trigger point (MTrP). Ultrasound can help, but requires the user to be experienced in ultrasound. Thus, this study investigates the use of texture features and deep learning strategies for the automatic identification of muscle with MTrPs (i.e., active and latent MTrPs) from normal (i.e., no MTrP) muscle.

METHODS

Participants (n = 201) were recruited from Toronto Rehabilitation Institute, and ultrasound videos of their trapezius muscles were acquired. This new data set consists of 1344 images (248 active, 120 latent, 976 normal) collected from these videos. For texture analysis, several features were investigated with varying parameters (i.e., region of interest size, feature type and pixel pair relationships). Convolutional neural networks (CNN) were also applied to observe the performance of deep learning approaches. Performance was evaluated based on the classification accuracy, micro F1-score, sensitivity, specificity, positive predictive value and negative predictive value.

RESULTS

The best CNN approach was able to differentiate between muscles with and without MTrPs better than the best texture feature approach, with F1-scores of 0.7299 and 0.7135, respectively.

CONCLUSION

The results of this study reveal the challenges associated with MTrP identification and the potential and shortcomings of CNN and radiomics approaches in detail.

The Use of Virtual Reality in the Rehabilitation of Chronic Nonspecific Neck Pain: A Systematic Review and Meta-Analysis

OBJECTIVES

There has been a major interest in using virtual reality (VR) as a pain-management tool. This systematic review evaluated the literature on the use of VR in the treatment of chronic nonspecific neck pain (CNNP).

METHODS

Electronic database searches were conducted in Cochrane, Medline, PubMed, Web of Science, Embase, and Scopus between inception and November 22, 2022. Search terms used were synonyms of "chronic neck pain" and "virtual reality." Inclusion criteria were as follows: chronic neck pain patients or pain lasting longer than 3 months; nonspecific neck pain; adult population; VR intervention; and functional and/or psychological outcomes. Study characteristics, quality, participant demographics, and results were independently extracted by 2 reviewers.

RESULTS

VR interventions demonstrated significant improvement in patients experiencing CNNP. Scores in the visual analogue scale, the Neck Disability Index, and range of motion were significantly improved compared with baseline but not better than gold standard kinematic treatments.

DISCUSSION

Our results suggest that VR is a promising tool for chronic pain management; however, there is a lack of VR intervention design consistency, objective outcome measures, follow-up reporting, and large sample sizes. Future research should focus on designing VR interventions to serve specific, individualized movement goals as well as combining quantifiable outcomes with existing self-report measures.

Denoising and decoding spontaneous vagus nerve recordings with machine learning

Neural interfaces that electrically stimulate the peripheral nervous system have been shown to successfully improve symptom management for several conditions, such as epilepsy and depression. A crucial part for closing the loop and improving the efficacy of implantable neuromodulation devices is the efficient extraction of meaningful information from nerve recordings, which can have a low Signal-to-Noise ratio (SNR) and non-stationary noise. In recent years, machine learning (ML) models have shown outstanding performance in regression and classification problems, but it is often unclear how to translate and assess these for novel tasks in biomedical engineering. This paper aims to adapt existing ML algorithms to carry out unsupervised denoising of neural recordings instead. This is achieved by applying bandpass filtering and two novel ML algorithms to in-vivo spontaneous, low-SNR vagus nerve recordings. The performance of each approach is compared using the task of extracting respiratory afferent activity and validated using cross-correlation, MSE, and accuracy in terms of extracting the true respiratory rate. A variational autoencoder (VAE) model in particular produces results that show better correlation with respiratory activity compared to bandpass filtering, highlighting that these models have the potential to preserve relevant features in complex neural recordings.

A comparison of extraneural approaches for selective recording in the peripheral nervous system

The peripheral nervous system is a key target for the development of neural interfaces. However, recording from the peripheral nerves can be challenging especially when chronic implantation is desired. Nerve cuffs are frequently employed using either two or three contacts to provide a single recording channel. Advancements in manufacturing technology have enabled multi-contact cuffs, enabling measurement of both temporal (i.e., velocity) and spatial information (i.e., spatial location). Selective techniques have been developed with different time resolutions but it is unclear how the number of contacts and their spatial configuration affect their performance. Thus, this paper investigates two extraneural recording techniques (LDA and spatiotemporal signatures) and compares them using recordings made from the sciatic nerve of rats using high density (HD, 56 contact), reduced-HD (16 contacts), and low density (LD, 16 contact) datasets. Performance of the two techniques was evaluated using classification accuracy and F₁-score. Both techniques show an expected improvement in classification accuracy with the spatiotemporal signature approach showing a 21.6 (LD to HD) - 24.6% (reduced HD to HD) increase and the LDA approach showing a 2.9 (reduced HD to HD) - 41.3% (LD to HD) increase and had comparable results in both the LD and HD datasets.

Tutorial: A guide to techniques for analysing recordings from the peripheral nervous system

The nervous system, through a combination of conscious and automatic processes, enables the regulation of the body and its interactions with the environment. The peripheral nervous system is an excellent target for technologies that seek to modulate, restore or enhance these abilities as it carries sensory and motor information that most directly relates to a target organ or function. However, many applications require a combination of both an effective peripheral nerve interface and effective signal processing techniques to provide selective and stable recordings. While there are many reviews on the design of peripheral nerve interfaces, reviews of data analysis techniques and translational considerations are limited. Thus, this tutorial aims to support new and existing researchers in the understanding of the general guiding principles, and introduces a taxonomy for electrode configurations, techniques and translational models to consider.

Ultrasound Image Quality Evaluation using a Structural Similarity Based Autoencoder

Ultrasound (US) imaging is a widely used clinical technique that requires extensive training to use correctly. Good quality US images are essential for effective interpretation of the results, however numerous sources of error can impair quality. Currently, image quality assessment is performed by an experienced sonographer through visual inspection, however this is usually unachievable by inexperienced users. An autoencoder (AE) is a machine learning technique that has been shown to be effective at anomaly detection and could be used for fast and effective image quality assessment. In this study, we explored the use of an AE to distinguish between good and poor-quality US images (caused by artifacts and noise) by using the reconstruction error to train and test a random forest classifier (RFC) for classification. Good and poor-quality ultrasound images were obtained from forty-nine healthy subjects and were used to train an AE using two different loss functions, with one based on the structural similarity index measure (SSIM) and the other on the mean squared error (MSE). The resulting reconstruction errors of each image were then used to classify the images into two groups based on quality by training and testing an RFC. Using the SSIM based AE, the classifier showed an average accuracy of 71%±4.0% when classifying images based on user errors and an accuracy of 91%±1.0% when sorting images based on noise. The respective accuracies obtained from the AE using the MSE function were 76%±2.0% and 83%±2.0%. The results of this study demonstrate that an AE has the potential to differentiate good quality US images from those with poor quality, which could be used to help less experienced researchers and clinicians obtain a more objective measure of image quality when using US.

A randomized double blinded placebo controlled study to evaluate motor unit abnormalities after experimentally induced sensitization using capsaicin

Central sensitization is a condition that represents a cascade of neurological adaptations, resulting in an amplification of nociceptive responses from noxious and non-noxious stimuli. However, whether this abnormality translates into motor output and more specifically, ventral horn abnormalities, needs to be further explored. Twenty healthy participants aged 20-70 were randomly allocated to topical capsaicin or a placebo topical cream which was applied onto their left upper back to induce a transient state of sensitization. Visual analogue scale (VAS) ratings of pain intensity and brush allodynia score (BAS) were used to determine the presence of pain and secondary allodynia. Surface electromyography (sEMG) and intramuscular electromyography (iEMG) were used to record motor unit activity from the upper trapezius and infraspinatus muscles before and twenty minutes after application of capsaicin/placebo. Motor unit recruitment and variability were analyzed in the sEMG and iEMG, respectively. An independent t-test and Kruskal-Wallis H test were performed on the data. The sEMG results demonstrated a shift in the motor unit recruitment pattern in the upper trapezius muscle, while the iEMG showed a change in motor unit variability after application of capsaicin. These results suggest that capsaicin-induced central sensitization may cause changes in ventral horn excitability outside of the targeted spinal cord segment, affecting efferent pathway outputs. This preclinical evidence may provide some explanation for the influence of central sensitization on changes in movement patterns that occur in patients who have pain encouraging of further clinical investigation.Clinical Trials registration number: NCT04361149; date of registration: 24-Apr-2020.

Compensation Strategies for Bioelectric Signal Changes in Chronic Selective Nerve Cuff Recordings: A Simulation Study

Peripheral nerve interfaces (PNIs) allow us to extract motor, sensory, and autonomic information from the nervous system and use it as control signals in neuroprosthetic and neuromodulation applications. Recent efforts have aimed to improve the recording selectivity of PNIs, including by using spatiotemporal patterns from multi-contact nerve cuff electrodes as input to a convolutional neural network (CNN). Before such a methodology can be translated to humans, its performance in chronic implantation scenarios must be evaluated. In this simulation study, approaches were evaluated for maintaining selective recording performance in the presence of two chronic implantation challenges: the growth of encapsulation tissue and rotation of the nerve cuff electrode. Performance over time was examined in three conditions: training the CNN at baseline only, supervised re-training with explicitly labeled data at periodic intervals, and a semi-supervised self-learning approach. This study demonstrated that a selective recording algorithm trained at baseline will likely fail over time due to changes in signal characteristics resulting from the chronic challenges. Results further showed that periodically recalibrating the selective recording algorithm could maintain its performance over time, and that a self-learning approach has the potential to reduce the frequency of recalibration.

Effect of Neighbourhood Size in Entropy Mapping of Ultrasound Images

Image filtering is a technique that can create additional visual representations of the original image. Entropy filtering is a specific application that can be used to highlight randomness of pixel grayscale intensities within an image. These image map created from filtering are based on the number of surrounding neighbourhood of pixels considered. However, there is no standard procedure for determining the correct "neighbourhood size" to use. We investigated the effects of neighbourhood size on the entropy calculation and provide a standardized approach for determining an appropriate neighbourhood size in entropy filtering in a musculoskeletal application. Ten healthy subjects showing no symptoms related to neuromuscular disease were recruited and ultrasound images of their trapezius muscle were acquired. The muscle regions in the images were manually isolated and regions of interest with varying neighbourhood sizes (increasing by 2 pixels) from 3x3 to 61X61 pixels were extracted. The entropy, relative signal entropy over noise entropy, statistical effect size as well as the percentage change of the effect size and instantaneous slope of the effect size was examined. The analysis showed that a neighbourhood size within the range of 21-25 pixels provides the maximum amount of information gained and coincides with a percentage change of the effect size of less than 5% and instantaneous slopes < 0.05.

Impact of Encapsulation Tissue Growth on Selective Recording in Nerve Cuff Electrodes: A Simulation Study

Peripheral nerve interfaces (PNIs) allow us to extract motor, sensory and autonomic information from the nervous system and use it as control signals in neuroprosthetic and neuromodulation systems. Recent efforts have aimed to improve the recording selectivity of PNIs, including by using spatiotemporal patterns from multi-contact nerve cuff electrodes as input to a convolutional neural network (CNN). Before such a methodology can be translated to humans, its performance in chronic implantation scenarios must be evaluated. We investigated the performance of a CNN-based selective recording approach in the presence of encapsulation tissue, a common immune response to the implantation of a neural interface. This factor was simulated using anatomically accurate computational models of a rat sciatic nerve and nerve cuff electrode. Performance over time was examined in three conditions: training the CNN at baseline only, supervised retraining with explicitly labeled data at periodic intervals, and a semi-supervised self-learning approach. The periodic recalibration approach demonstrated the best results, with an average F1-score of 0.96 ± 0.04, 0.89 ± 0.08, and 0.80 ± 0.08 for SNRs of -5 dB, -10 dB, and -15 dB, respectively, across all time points. Thus, the periodic recalibration approach may be an effective solution to compensate for changes in signal recordings seen over time as a result of encapsulation tissue. The self-learning approach, in which a network is retrained periodically using predicted labels, generally showed degradation in classification performance over time, even as the frequency of training was increased, attributed to an eventual accumulation of mislabeled training data.

Influence of Anatomical Detail and Tissue Conductivity Variations in Simulations of Multi-Contact Nerve Cuff Recordings

Accurate simulations of peripheral nerve recordings are needed to develop improved neuroprostheses. Previous models of peripheral nerves contained simplifications whose effects have not been investigated. We created a novel detailed finite element (FE) model of a peripheral nerve, and used it to carry out a sensitivity analysis of several model parameters. To construct the model, in vivo recordings were obtained in a rat sciatic nerve using an 8-channel nerve cuff electrode, after which the nerve was imaged using magnetic resonance imaging (MRI). The FE model was constructed based on the MRI data, and included progressive branching of the fascicles. Neural pathways were defined in the model for the tibial, peroneal and sural fascicles. The locations of these pathways were selected so as to maximize the correlations between the simulated and in vivo recordings. The sensitivity analysis showed that varying the conductivities of neural tissues had little influence on the ability of the model to reproduce the recording patterns obtained experimentally. On the other hand, the increased anatomical detail did substantially alter the recording patterns observed, demonstrating that incorporating fascicular branching is an important consideration in models of nerve cuff recordings. The model used in this study constitutes an improved simulation tool and can be used in the design of neural interfaces.

Effects of the choice of reference on the selectivity of a multi-contact nerve cuff electrode

Tripolar referencing is typically used in nerve cuff electrode recordings due to its ability to maximize the signal-to-noise ratio of contacts at the centre, but this may not be the optimal choice for a multi-contact nerve cuff consisting of contacts in off-centre rings. We conducted a simulation study to compare the effects of 3 different reference types on the recording selectivity of a multi-contact nerve cuff: the tripolar reference (TPR), common average reference (CAR), and multiple tripolar references based on consecutive groups of 3 rings (cTPR). For this purpose, we introduce a novel measure called the contact information metric (CIM). Selectivity was tested in 2 noise settings, one in which white Gaussian noise was added inside the nerve cuff electrode and the other in which electromyogram (EMG) noise was added outside the nerve cuff electrode. The mean CIMs values calculated for the best 8 contacts were 3.42±6.25, 2.70±3.37, and 3.65±1.90 for the TPR, the CAR and the cTPR, respectively, in the case of EMG noise added outside the nerve cuff electrode. This study shows that the use of cTPR reference is the optimal choice for selectivity when using a multi-contact nerve cuff electrode which contains off-centre rings.