Towards validation of clinical measures to discriminate between nociceptive, neuropathic and nociplastic pain: cluster analysis of a cohort with chronic musculoskeletal pain ============================================================================================================================================================================= * Paul W Hodges * Raimundo Sanchez * Shane Pritchard * Adam Turnbull * Andrew Hahn * Jon Ford ## Abstract The International Association for the Study of Pain defines three pain types presumed to involve different mechanisms - nociceptive, neuropathic and nociplastic. Basis on the hypothesis that these pain types should guide matching of patients with treatments, work has been undertaken to identify features to discriminate between them for clinical use. This study aimed to evaluate the validity of these features to discriminate between pain types. Subjective and physical features were evaluated in a cohort of 350 individuals with chronic musculoskeletal pain attending a chronic pain management program. Analysis tested the hypothesis that, *if* the features nominated for each pain type represent 3 different groups, *then* (i) cluster analysis should identify 3 main clusters of patients, (ii) these clusters should align with the pain type allocated by an experienced clinician, (iii) patients within a cluster should have high expression of the candidate features proposed to assist identification of that pain type. Supervised machine learning interrogated features with the greatest and least importance for discrimination; and probabilistic analysis probed the potential for coexistence of multiple pain types. Results confirmed that data could be best explained by 3 clusters, clusters were characterized by *a priori* specified features, and agreed with the designation of the experienced clinical with 82% accuracy. Supervised analysis highlighted features that contributed most and least to the classification to pain type and probabilistic analysis reinforced the presence of mixed pain types. These findings support the foundation for further refinement of a clinical tool to discriminate between pain types. ## Introduction Chronic pain contributes to 71% of disability worldwide1. Back pain alone costs the US economy >$100 billion annually2, the most of any condition. Thousands of randomized controlled trials show modest effects3. One size does not fit all, yet >90% of back pain is labelled “non-specific” of unknown cause4 without evidence-based rationale to guide treatment. Trial-and-error application of treatment is wasteful, inefficient, and rejected by those with chronic pain5. High variability in individual responses could be improved by predictive biomarkers that guide treatment allocation. Unfortunately, attempts to target treatments have had limited success6. Identification of the predominant mechanism that explains and individual’s pain could help guide treatment7. Pain involves an array of inputs and outputs, and diverse biological and neural mechanisms, influenced by factors including emotions and cognitions8. Although activation of nociception by actual or threatened tissue damage is one input, many other inputs and mechanisms interplay. These differ between individuals9. The International Association for the Study of Pain (IASP) defines 3 pain types that are hypothesized to relate to different underlying mechanisms7, which should require different treatments. These are Nociceptive pain – arising from actual or threatened damage to non-neural tissue, due to activation of nociceptors; Neuropathic pain – caused by lesion or disease of the somatosensory nervous system; and Nociplastic pain –arising from altered nociception despite no clear evidence of actual or threatened tissue damage causing activation of nociceptors or evidence for disease or lesion of the somatosensory system10. Biomarker signatures for pain mechanisms have potential to predict response to targeted treatments for chronic pain. There is no “gold standard” to discriminate between mechanisms. Instead, methods have been proposed using multi-modal data (e.g., quantitative sensory testing (QST), pain qualities, questionnaires) to identify features suggesting predominance of one mechanism. Progress has been compromised by diversity of opinion regarding *how* pain types might be identified11,12 or *if* this is possible13. Recent systematic reviews7,14 and expert consensus15 have identified and refined candidate measures that could aid in the discrimination between the pain types, and other work has defined clinical criteria for identification of a single type (e.g., nociplastic16,17; neuropathic18). In the absence of a gold standard tool to identify mechanism, validation has been difficult. One alternative to evaluate the validity of the discrimination between the IASP pain types is to collect data for a cohort of individuals with diverse presentation of chronic pain, and test the hypothesis that, *if* the features nominated for each pain type represent 3 different groups, *then* (i) cluster analysis should identify 3 main clusters of patients, (ii) these clusters should align with the pain type allocated by an experienced clinician, (iii) patients within a cluster should have high expression of the candidate features that are proposed for that pain type. This study aimed to test these hypotheses using a cohort of individuals with chronic musculoskeletal pain attending a pain management clinic, and then to (i) interrogate which features had the greatest and least importance for the discrimination; and (ii) to probe the potential for coexistence of multiple pain types. ## Methods ### Study design This study evaluated a cohort of individuals with diverse presentations of chronic pain using the highest ranked features allocated by expert consensus15 to aid discrimination between the IASP pain types. For this study all features were assessed with simple subjective and physical examinations that were feasible in a clinical setting without specialised equipment. All items were assessed for all participants (with some limited exceptions). The natural tendency for the data to cluster in to three groups was evaluated first. If three clusters were present, we then interrogated the relationship with the predominant pain type allocated based on the scoring and review by an independent experienced clinician. Because the data assessed as “typical features” of each pain type *and* the classification of the participant’s predominant pain type use the same data, it might be argued the analysis step is biased to confirm the hypothesis. This outcome cannot be assumed because the clustering algorithms (unsupervised machine learning models) do not use the dependent variable (pain type) for modelling or prediction, that is, neither the score-derived nor examiner-derived pain type was used in generation of clusters. Second, we employed supervised machine learning to understand the relative importance of test items to the pain-type clusters. Third, we considered probabilistic models to provide insight into the coexistence of multiple pain types, which remains hidden in deterministic analysis. ### Participants Participants were 350 individuals with chronic pain who presented for attendance at a chronic pain management program. They were included if they had a primary complaint of chronic pain (pain of >3 months duration), were referred to attend a multidisciplinary pain management program, were aged between 18 and 75 years, and had sufficient fluency in English to complete the assessment. Participants were excluded if they had red flags that would be unsuitable for pain management (including active cancer, signs of cauda equina syndrome based on bladder or bowel disturbance, risk of spinal fracture, signs of potential infection, foot drop that my cause tripping, spondyloarthropathy). The Medical Research Ethics committee of LaTrobe University provided ethical approval for the collection and analysis of the data. Participants were asked to provide informed consent for analysis of their de-identified data when attending their initial assessment. Characteristics of the participant group were recorded using the electronic Persistent Pain Outcomes Collaboration (ePPOC) standardised assessment data set{Tardif, 2017 #8118}. ### Measures Features were assessed as part of the clinical examination. Items were selected from the highest ranked items from the recent expect consensus study15, and refined using the clinical criteria reported to identify nociplastic16,17 and neuropathic18 pain. The full list of features considered for each pain type are presented in Supplementary Data 1-3. From the candidate items, a list of subjective and physical examination items was generated that was possible to apply in a clinical setting as subjective or physical tests. Table 1 presents the final list of features. The 31 items included 11 for nociceptive pain, 8 for neuropathic pain and 12 for nociplastic pain. For nociplastic pain, physical item 3 was only completed if item 2 was negative. All items were scored on a scale of 0-100, anchored with “not present” at 0 and strongly present at “100”. Although all items were used as independent items in the analysis, we also calculated the average score for the questions for each pain type separately. The dominant pain type was allocated based on that with the highest average score (Score-derived pain type). In addition, an experienced clinician independently reviewed the patient’s chart, including the response to the pain type questions and assigned a likely predominant pain type (Examiner-derived pain type). View this table: [Table 1](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/T1) Table 1 Variables considered in the study ## Data analysis ### Data preparation Quality of data was assessed in several ways depending on data type. Independent variables: The dataset contains responses on a scale from 0 to 100. Missing data were analysed through a graphical distribution analysis of the percentage of missing data. We assumed that missing metrics were ignored by the examiner because they were not necessary to allocate a pain type to the participant (e.g., *nociplastic physical 3* was not performed if features 1 or 2 were positive). For this reason, missing values were imputed with zeros to avoid losing sample size. Dependent variables: Dominant pain mechanisms were analysed using a distribution plot to consider the generalization ability of the model, and to consider class balance or imbalance for selection of the clustering algorithm selection and configuration. This was undertaken separately to for pain types derived from pain scores, and that diagnosed by an examiner. Unambiguous diagnosis was present in 80% of cases. ### Pre-clustering analysis using descriptive statistics and correlations A pre-clustering analysis was performed to ensure the independent variables contained sufficient information to distinguish between the three types of pain, by way of validating the potential to consider a clustering approach. First, a correlation analysis of the independent variables was performed using Pearson’s correlation, visualised through a correlation plot, and arranged by nociceptive, nociplastic, and neuropathic criteria. To support the clustering approach, a high correlation within categories and low correlation between categories would be expected. Second, a heatmap was created, ordering data by pain type obtained from scores, visualising the “intensity” of responses to corresponding criteria to identify patterns. To support the clustering approach, patients within a cluster should have high expression of relevant features. Third, before clustering we also computed the Hopkins statistic19, to measure the natural cluster tendency in the data. Values close to 1 indicate that the data are highly clustered, and close to 0.5 indicates that the data are randomly distributed. Values below 0.75 suggest that the clusters are sparse and with fuzzy borders between them. For this analysis, data were reduced using principal components analysis for visualization purposes as it allows projecting n-dimensional data into 2-dimensions. The clustering tendency was computed with the original features. The clustering tendency was presented using a scatter plot with the first and second principal components on the x- and y-axes, respectively. A 2D density plot was included to visualise the concentration of pain types. ### Consistency of clusters with pain type classification To determine whether analysis would identify three cluster, and if so, whether these were aligned to the three pain types we analysed the data ins several ways. First, the number of clusters was evaluated, considering clusters between two and ten. Three common clustering methods were assessed: k-means, Gaussian mixture models (GMM), and Hierarchical clustering20. The silhouette coefficient was calculated for each scenario. This coefficient provides an indication of how well each data point was clustered, with positive values indicating better classification21. We determined the optimal number of clusters as that which maximized the silhouette coefficient. To support our hypothesis that the clinical features relate to three pain types, this optimal number of clusters would be three. Second, using the three clusters, the consistency of each clustering methods was evaluated by comparing each point’s clustering against the pain type derived from score-derived and examiner-derived pain types. Each of the six resulting comparisons was evaluated using the True Positive Rate (TPR), which accounts for the percentage of cases that were correctly classified according to the average score (score-derived classification) and the experienced examiner (examiner-derived classification). This value is from 0 to 100%. Third, for hierarchical clustering (which provided the second highest TPR, see results, but enables investigation of the internal structure), a dendrogram was generated to graphically analyse the hierarchical structure, showing the model’s clusters and allocation to predominant pain type based on highest scores. Within this approach the Euclidean distance and averaged linkage grouping criteria were used to allocate participants to each cluster. ### Prediction of pain type based on pain features using supervised machine learning and evaluation of relative contribution of test items to discrimination If the unsupervised analysis confirmed the presence of three clusters, we planned to used supervised learning to explore the possibility of extracting relevant information to understand the relative value of individual test items to the classification. Due to small sample size, we applied a logistic regression model, which has fewer parameters than other models, and allows the prediction of pain type based on the independent variables and interpret the importance of the features to make this prediction. This approach was applied to classification of the predominant pain type deduced from the highest scores and that allocated by the independent examiner. We used 80% of the data for model training. The coefficients of both models were evaluated, analysing the sign of these coefficients to determine each variable’s contribution to the prediction. The accuracy of both models was also evaluated using the TPR, ensuring consistency with the previous clustering comparison. ### Probabilistic predictions of pain mechanisms to investigate coexistence of pain types Finally, a probabilistic analysis was conducted to explore the uncertainty in model predictions, both supervised and unsupervised. GMM and logistic regression were used to output probabilistic classifications, and a probabilistic variation of k-means called c-means was also used to generate cluster membership likelihoods. The likelihood of the dominant pain type for each model is presented as boxplot. Values that are concentrated around 100% show the model is more certain about the prediction, whereas sparse values show higher uncertainty. Each comparison is presented with the Fuzzy Partition Coefficient (FPC), which evaluate the degree of uncertainty in the predictions, with values closer to 100% indicating well-defined and distinct clusters (i.e., if all pain types were mutually exclusive, the FPC would be 100% for all pain types). ## Results ### Participants Data were available for 350 participants presenting for pain management with a primary complaint of chronic pain. Participants (62.7% female; 37.3% male) had an average (standard deviation) age, height and weight of 46(12) years, 167(11) cm and 84.2(20.9) kg, respectively. Regarding the duration of pain, 1.7% of individuals had experienced pain for less than 3 months, 22.3% for 3 to 12 months, 33.9% for 12 months to 2 years, 29.8% for 2 to 5 years, and 12.4% for more than 5 years. When asked to describe their pain, 0.4% reported that it was rarely present, 4.7% said it was occasionally present, 8.5% described it as often present, 75.3% stated it was always present with varying intensity, and 11.1% reported it was always present with consistent intensity. The main sites of pain reported were the low back (28.3%), shoulder (22.2%), and neck (19.1%), with other less common sites including the mid back (5.7%), knee (4.8%), foot (3.9%), hip (3.5%), wrist (2.6%), hand (2.6%), and other areas (7.4%). The average total Pain Disability Questionnaire Score was 104 (27). From the possible data 265 data points contain at least one variable with missing values. ### Data preparation Figure 1 shows the distribution of missing values which indicates that most metrics have <15% of missing data. As expected, metrics such as “*nociplastic subjective 3*” (See Table 1 for item definitions) and “*nociplastic physical 3*” have the highest percentage of missing values, exceeding 30%. Pain type derived by the examiner was unavailable for 20% of participants. In this case the expert examiner had identified the pain type as ambiguous. ![Figure 1](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/08/14/2024.08.13.24311924/F1.medium.gif) [Figure 1](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/F1) Figure 1 Percentage of missing values for the different variables of the study. Blue bars represent metrics with less than 15% missing data, whereas red bars represent metrics with more than 15% missing data. The available sample is unbalanced for the different pain types (Figure 2). According to the pain type (both score-derived or examiner-derived), most cases were classified as nociplastic, followed by nociceptive and neuropathic with the least cases. In the case of the examiner’s unambiguous classification, all nociceptive participants are available, but availability of classification of the other pain types was reduced by diagnostic uncertainty. ![Figure 2](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/08/14/2024.08.13.24311924/F2.medium.gif) [Figure 2](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/F2) Figure 2 Distribution of participants across three different pain types. The y-axis represents the patient count for each pain mechanism, and the red outline box indicates the examiner’s available classification for each pain type. In the cases where the expert did not allocate a classification it was indicated as ambiguous, this uncertainty was higher in participants with higher neuropathic scores. This might bias the results using examiner-derived pain types, particularly for underrepresented groups such as those with neuropathic pain. ### Pre-clustering analysis using descriptive statistics and correlations As we required to proceed to the clustering step, the correlation among variables was higher and positive within the criteria for each pain type, than between pain types (Figure 3; high correlation expected within each red box). This finding demonstrates strong internal consistency. The highest correlation (R=0.82) was between *nociceptive subjective items 2* and *3*, which both relate to the relation of pain to aggravating factors. Subjective and physical features of nociceptive that address similar constructs are highly correlated. Negative relations can be seen between features of different pain types, such as nociceptive and nociplastic, where negative values present in almost all comparisons. *Nociceptive subjective 5* and *neuropathic subjective 6* have the strongest negative correlation, which is expected as these items relate to low or high scores, respectively, on the same neuropathic questionnaire. This neuropathic questionnaire was modestly correlated with nociplastic features, suggesting its items reflect both pain types. ![Figure 3](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/08/14/2024.08.13.24311924/F3.medium.gif) [Figure 3](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/F3) Figure 3 Correlation plot between the study independent variables. Each cell represents the Pearson correlation coefficient between pairs of variables, with the colour indicating the strength and direction of the correlation. Red boxes group specific questions for each pain mechanism. The pattern of expression of each feature across participants is shown as a heat map in Figure 4. As expected, the density of blue, determined by the intensity of expression of each feature is high with the red boxes that are bounded on the x-axis arranged by pain type classified according to the scoring, and on the y-axis by the features for each pain type. In some cases (e.g., *nociceptive physical 6*, or *nociplastic comorbidities 1*) there is stable expression across all pain types, which suggests limited value in discrimination between the pain types. Note *nociplastic physical 3* is commonly scored as 0 because it is only tested if *nociplastic physical 2* is negative. ![Figure 4](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/08/14/2024.08.13.24311924/F4.medium.gif) [Figure 4](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/F4) Figure 4 Heatmap of all the independent variables. Each cell represents the value of the participants answers to a given question, with the participants arranged by dominant pain type allocated by the largest score. Red boxes group specific questions for each pain mechanism with participants with corresponding dominant pain type. When analysing the clustering tendency of the data, the Hopkins statistic was 60%. This indicates some degree of natural clustering is present in the data, but we should not expect perfectly defined clusters. The data distribution presented in Figure 5 shows that the density curves for pain types overlap between each other, suggesting that many points may be classified with more than one pain type. Consideration of the weightings for each of the feature in the principal components demonstrates that principal component 2 distinguishes neuropathic from the other 2 and, as expected, is weighted by neuropathic features. Further, nociceptive and nociplastic are distinguished but principal component 2, which is weighted by features of these pain types. ![Figure 5](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/08/14/2024.08.13.24311924/F5.medium.gif) [Figure 5](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/F5) Figure 5 Cluster tendency of the independent variables. The upper panel demonstrates the Hopkins Statistic. The x- and y -axis represent the 1st and 2nd principal component of the data. For each pain type, a density map is depicted to illustrate the differentiation of pain type’s distribution. The bottom panel shows the features that contribute to the 2 principal components. Note that *nociceptive subjective 5* is a low score on the neuropathic questionnaire. ### Assessment of consistency of classification with predominant pain type When testing the clustering algorithms, the optimal number of clusters was analysed as shown in Figure 6. The three clustering models minimize the share of negative silhouette at three clusters. At this point the overall silhouette is maximized for hierarchical approach, and for GMM and k-means is only beaten by the silhouette of using two clusters. Combining both metrics would suggest using three clusters. ![Figure 6](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/08/14/2024.08.13.24311924/F6.medium.gif) [Figure 6](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/F6) Figure 6 Cluster Silhouette Analysis for different number of clusters. Silhouette indicates how well each data point was clustered, so it should be maximized. Three clustering methods were tested, varying the number of clusters between 2 and 10. Low values for the share of negative silhouette also serves as an overall metric of proper clustering. When predicting pain type using clustering methods, results differ depending on which version of the ground truth (score-derived or examiner-derived pain type) was used, as shown in Figure 7. For interpretation of this analysis, the pain type that was most likely to be represented by each cluster was identified by inspection of the heat map which showed the features that were commonly expressed in each cluster. As shown for the hierarchical cluster method in Figure 8, the largest cluster was dominated by high expression of features of nociplastic pain, and so on. When prediction accuracy was analysed using both versions, the best model was the GMM. The examiner’s diagnosis achieved a TPR of 82%, and the score derived pain type achieved a TPR of 89%. ![Figure 7](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/08/14/2024.08.13.24311924/F7.medium.gif) [Figure 7](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/F7) Figure 7 Clustering prediction accuracy. Each of the 3 clustering algorithms were tested against the dominant pain type as computed with the score, and against the pain defined by the examiner. Each comparison shows the percentage of predictions from one pain type that were classified as any pain type. The diagonal represents the correct predictions, and the sum of the diagonal is the True Positive Rate (TPR). ![Figure 8](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/08/14/2024.08.13.24311924/F8.medium.gif) [Figure 8](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/F8) Figure 8 Heatmap of all the independent variables for the clustered data. Each cell represents the value of the participants answers to a given question, with the participants arranged by the three clusters. Red boxes are bounded by the allocation of the participants to a cluster in the x-axis, and the pain type specific questions on the y axis. The hierarchical method achieved close to 80% TPR, and we can observe the structure of this model’s predictions in a tree structure (i.e., dendrogram) shown in Figure 9. It shows that the nociplastic predictions are the most consistent. ![Figure 9](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/08/14/2024.08.13.24311924/F9.medium.gif) [Figure 9](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/F9) Figure 9 Cluster dendrogram. The tree structure represents the closeness between data points, where they merge iteratively until only one cluster is formed. The distance to the centre represents the difference between the merging branches. The colour of the branches represents the predominant pain type as defined by the model, and the colour of the points at the end of each branch represents predominant pain type defined by the pain score. Note the high degree of consistency ### Prediction of pain type based on pain features using supervised machine learning The sign of coefficients from application of supervised machine learning using logistic regression models, were in the expected direction with positive values for features of corresponding pain type (Figure 10), and negative values for other pain types. The coefficient magnitudes were greater for the pain derived score model than for the examiner. This was expected as the score was directly derived from these measures. No features had significant positive magnitudes for more than one pain type. Several features show non-significant values for the three predicted pain types, which means they contributed little to the prediction of pain type (Table 2). Features such as *nociplastic comorbidities 2*, *neuropathic subjective 6* and *nociceptive physical 2, physical 6* and *subjective 3* contributed little to the prediction and were redundant. ![Figure 10](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/08/14/2024.08.13.24311924/F10.medium.gif) [Figure 10](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/F10) Figure 10 Coefficient heatmaps from logistic regression analysis for score-derived and examiner-derived pain types. The heatmap displays coefficients for each feature based on the observed pain type. Blue indicates a positive relationship, whereas red indicates a negative relationship with the pain mechanism categories. The outlined boxes on both heatmaps highlight the grouped coefficients for each pain mechanism. View this table: [Table 2](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/T2) Table 2 Most and least important features to support classification of pain type After performing predictions, the supervised classification models were evaluated with the TPR, and can be compared with previous results from cluster analysis. Figure 11 shows that the model for the score derived pain type achieves a TPR of 94%, whereas the model for the examiners pain type achieves a TPR of 87%. ![Figure 11](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/08/14/2024.08.13.24311924/F11.medium.gif) [Figure 11](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/F11) Figure 11 Supervised classification accuracy. Each of the pain type as computed with the score, and against the pain defined by the examiner were tested against a multiclass logistic regression model. Each comparison shows the percentage of predictions from one pain type that were classified as any pain type. The diagonal represents the correct predictions, and the sum of the diagonal is the True Positive Rate (TPR). ### Probabilistic predictions of pain mechanisms to evaluate coexistence of pain types Both supervised and unsupervised machine learning techniques were used to generate probabilistic outcomes. This was used to addressing the complexity of pain type classification by giving insight into the coexistence of multiple pain types (that remains hidden in deterministic analysis). Figure 12 shows the comparison of 4 different probabilistic models, where the dominant cluster likelihood is represented as boxplots. The GMM have high concentration of values close to 100%, which is consistent with the FPC of 95%, the highest of the sample. In contrast, the supervised models (score-derived or examiner-derived pain types) have FPC in the range of 72-80%, and in both cases the neuropathic pain type presents the lowest likelihood values among all pain types. Supervised clustering highlights probability for coexistence of pain types, and greatest uncertainty for the neuropathic type. ![Figure 12](http://medrxiv.org/http://medrxiv.stage.highwire.org/content/medrxiv/early/2024/08/14/2024.08.13.24311924/F12.medium.gif) [Figure 12](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/F12) Figure 12 Probabilistic classification analysis. Each colour represents the dominant pain type, and the values represent the model likelihood for each participant. The top panels show the results of probabilistic clustering (unsupervised), and the lower panels represent the probabilistic predictions of the supervised models already analysed. Each model is accompanied by the Fuzzy Partition Coefficient (FPC) which measures the degree of uncertainty in the model predictions. ## Discussion The results of this study support the validity of the clinical features proposed to discriminate between the three pain types defined by IASP. The data were best explained by 3 clusters, these were characterised by features proposed to be relevant for the three pain types, and clusters aligned strongly with the pain type identified by the experienced clinician with a classification accuracy of up to 82%. Logistic regression demonstrated items that contribute substantially to the classification and those providing no additional value. Probabilistic analysis demonstrated uncertainty of classification which suggests co-existence of mechanisms in some individuals. These findings provide a foundation to further develop and refine a tool to discriminate between pain types. ### Validity of methods to discriminate between pain types In the absence of a gold standard method to assess pain type, the mathematical approach used in this study provided an innovative way to investigate the validity of features to discriminate between them. Although this analysis cannot directly assess whether a participant’s pain was explained by the allocated type, three features support this interpretation – data were naturally explained by three clusters, each cluster was characterised by high expression of features considered relevant for a specific pain type, and the classification had high agreement with an experienced clinician. The Gaussian mixture model had the greatest accuracy (True Positive Rate) with 82% agreement. The 18% inaccuracy in allocation might be accounted for by mixed pain types (see below) or because all criteria contributed the same weight to the determination of cluster, yet the examiner might have placed more weight on specific features when making the clinical decision. Such weighting is a feature of current clinical criteria for nociplastic 16 and neuropathic 16 pain – e.g., subjective evidence of a pain distribution consistent with neuropathology indicates *possible* neuropathic pain, whereas physical evidence carries greater weight and indicates *probable* neuropathic pain. Future work should consider allocation of weights to items. The lower TPR provided by other clustering methods is accounted for by their properties. Unlike Gaussian mixture models which is a “soft” clustering technique enabling allocation of individuals to multiple clusters and clusters of different shapes22, K-Means23 and Hierarchical24 clustering are a “hard” clustering methods where individuals can only be allocated to a single cluster25 (ignoring the potential for mixed pain types) with spherical shape, and Hierarchical clustering can have difficulty with clusters of different size26. The existence of three clusters defined by features considered to be relevant to discriminate between pain types provides some validation that pain presentations can be allocated to these three pain types. It is important to note that this does not exclude the possibility that further pain types might be identified if additional features were included. The pain types represent those defined by the IASP10, and identified in systematic review of the literature7. The features used to support the classification are derived from a list of items identified through a rigorous process of systematic review14, expert consensus15 and consideration of the clinical criteria that have been published for neuropathic18 and nociplastic16 pain. ### Relative importance of features Supervised machine using logistic regression models highlighted positive relationships for all features suggested to be relevant for each pain types, but with variation in the magnitude of the coefficient. Consideration of the coefficients from the examiner-derived supervised analysis enable identification of the features that contributed most to the classification and those that contributed little additional information (Table 2). This analysis could be helpful to refine the number of required features that must be assessed but should be considered preliminary considering the nature of the data we collected using informal tests and the subjective rating scale. It is useful to consider the important of items in the context of the published clinical criteria for nociplastic and neuropathic pain. For nociplastic pain the most important criteria were the presence of regional distribution, physical findings of mechanical/heat/cold sensitivity, and painful after sensations. This aligns with the IASP clinical criteria which consider regional pain to be an “obligatory” criterion. Evoked hypersensitivity phenomena (including painful aftersensations) are necessary for nociplastic pain to be considered “possible” and additional evidence of sensitivity must also be found for it to be considered “probable” (the highest level)16. No additional value was provided by subjective evidence of sensitivity. This was not surprising because these features were strongly correlated with the physical measures. Allodynia is another example of and “evoked hypersensitivity phenomenon” for nociplastic pain, but physical and subjective evidence of sensitivity to light touch/allodynia did not make an additional contribution to classification. This is explained by high correlation with other features of this type and the higher prevalence of aftersensations in those with nociplastic pain. Although comorbidities are a commonly discussed as features of nociplastic pain27,28 and are prominent features in the Central Sensitization Index29, the comorbidities (sensitivity to sound etc, sleep disturbance, fatigue, cognitive problems) did not provide additional information to the classification. This is explained by observation of the heat maps which show they can be present in any pain type. It has been argued that comorbidities could be removed from diagnosis of nociplastic pain as they do not directly relate to nociceptive function30. For neuropathic pain, classification could be achieved if patients reported their pain in a neuroanatomically plausible distribution, hypoaesthesia (same distribution), and pain could be evoked by provocation of the nerve. No additional benefit was provided by report of dysesthesia (in the distribution), report of evidence of lesion or disease to the somatosensory nervous system, the neuropathic questionnaire above cut-off pain, or pain provoked by movements that load or compress neural tissue. When considering the clinical criteria for neuropathic pain18, the present study did not include formal testing of lesion or disease of somatosensory system which is required for a diagnosis of “definite” neuropathic pain. We relied on patient report. Pain and sensory symptoms within a neuroanatomically plausible distribution are consistent with criteria for probable neuropathic pain when determined by physical examination and the high correlation between subjective and physical assessment explains why physical measures provided little additional contribution. Pain induced by nerve provocation might be considered an example of physical evidence of pain distribution (feature for probable diagnosis). It is perhaps not surprising that response to the Neuropathic Pain Questionnaire did not add to the classification as these questionnaires predate the change in definition of neuropathic pain from “dysfunction” to “lesion or disease” of the somatosensory nervous system31, and the introduction of “nociplastic” pain12. Questionnaires such as this include questions related to sensitivity and would be expected to have limited capacity to discriminate between neuropathic and nociplastic pain7. There are no clinical criterial for nociceptive pain. The most important features for classification of this pain type include multiple features that would *exclude* the probability of other pain types. Features that should *not be* present include regional distribution (which could suggest nociplastic), and hypoaesthesia (which could suggest neuropathic pain), and features that should *be* present include no generalised hypersensitivity, which again implies a focus on exclusion of nociplastic pain. Together this implies convincing evidence of nociceptive pain involves a process of exclusion. This situation might be rectified by development of clinical criteria for nociceptive pain. The only other feature contributing strongly to classification is subjective and physical evidence of pain provoked by specific postures/movements in a proportional manner, which imply pain in response to tissue loading. This is consistent with the use of “movement evoked pain” or “mechanical pain” as alternative terms for nociceptive pain7. Features that were unhelpful include signs of inflammation and absence of autonomic signs which were both present across all pain types. ### Distribution of pain types and coexistence of mixed pain types The distribution of different pain types presented in this study should not be considered generalisable. The group were drawn from patients attending a chronic pain management program and might be expected to include a greater proportion of individuals with nociplastic pain, which we found. There will not yet be convincing evidence of distribution of pain types in existing literature for several reasons. First, the nociplastic pain type was only recently introduced and although sharing features with “chronic primary pain”32 or “central sensitization pain”33 it is not synonymous with those classifications. Second, methods do not yet exist to discriminate between pain types—hence the present work. The probabilistic analysis enabled consideration of mixed pain types. The Fuzzy Partition Coefficient (FPC) measures the degree of uncertainty in the model predictions such that a value of 100% would indicate that all patients allocated to a paint type could only be allocated to that pain type. Our data show that this was not achieved for any pain type. Although this concurs with report of mixed pain types7, it is not completely consistent with the clinical criteria for nociplastic pain which include the obligatory criterion that “there is no evidence that nociceptive/neuropathic pain (a) is present or (b) if present, is entirely responsible for the pain”16. The nuance in this definition that overlaps with our data is the implication that another mechanism can co-exist if it is not “entirely responsible” for the presentation. How that is judged is unclear. A premise of the clinical criteria for nociplastic pain is that a patient can have nociplastic pain or not16. This highlights a major issue of the current clinical criteria for nociplastic and neuropathic pain – these criteria aim to enable the diagnosis of a single pain type, without indication of whether feature of the other pain type are present or how they can be excluded. This is important for several reasons. First, it is plausible that predominant pain mechanism might change over time – e.g., a patient’s presentation might begin with characteristics of nociceptive pain, and transition to nociplastic features34. Second, as classification of pain types is proposed to assist in guiding matching patients to treatments7, whether a patient has a mixed pain type and which type predominates is likely to impact success of the matched treatment. Of note, our probabilistic analysis suggested greatest uncertainty for neuropathic pain. This could reflect the frequent presence of “sensitivity” in neuropathic pain7,35,36. A major motivator for the development of “nociplastic” pain as the third pain type was the change in the definition of neuropathic pain12. When the criteria for neuropathic pain changed from “dysfunction” to “disease or lesion” of nervous system, there was a need to account for patients who presented with sensitivity and symptoms unexplained by ongoing nociceptive input, but without nervous system damage or lesion12. Our overlap in neuropathic and nociplastic could reflect some lack of clarity of this distinction ### Development of a tool to discriminate between pain types There is a need to develop a tool to discriminate between pain types to not only provide insight into the plausible mechanisms underlying a patient’s pain, but to guide matching of the right treatment to the patient. This study applied a preliminary version of features to support discrimination, based on outcomes of expert consensus regarding the most likely features to support this decision7. This study provides some initial validation of the features, and some initial consideration of which features are most important to inform to inform the classification. It is important to note that features of a tool to discriminate between pain types will differ from those for clinical criterial to diagnose a specific pain type because the purpose of these tools differs. Clinical criteria for a pain type specify the features that must be present, but this does not exclude features that are present in all pain types, whereas a tool for discrimination should include features that are present in either one or two, but not all three pain types. A tool to discriminate between pain types fills an important gap as clinical criteria, require exclusion of the other pain types (explicitly for nociplastic, and implicitly for neuropathic), and this discrimination tool operationalises the identification of the other pain type to inform that decision. Further work is required to refine the pain type discrimination tool. First, this study used clinical judgement to assess each feature – formal operationalisation of assessment of these items would support interpretation and consistent implementation. Second, calculation of weights of each feature and refinement of the content will be necessary. Third, psychometric properties of the tool need to be assessed. Fourth, clinical validation of whether treatment outcomes are better when the tool is used to guide treatment matching is ultimately necessary. ### Methodological limitations There are several methodological considerations. First, features were assessed and allocated to a score based on clinical judgement. We did not assess the psychometric properties of the scoring which might have increased variability in the data. Second, the results apply to a diverse group with chronic pain attending a specific chronic pain management program and might not be representative of chronic pain patients in other cohorts or environments. Third, features were based on a preliminary version of a pain type discrimination tool and are likely to differ from the final feature specification. Fourth, the experienced clinician used the clinical data including test outcomes to classify the pain types. Although the initial cluster analysis does not use this information and is therefore unbiased in its classification, these data were used for the supervised methods for the supervised machine learning, and potential for bias needs to be recognised. Fifth, we cannot exclude that additional clusters would be identified of additional features were assessed. ## Conclusion In summary, this study provides initial validation of the approach to discriminate between pain types based on features of a patient’s pain presentation. This forms a foundation for future work to refine and test tool to deploy in clinical practice. ## Data Availability All data produced in the present work are contained in the manuscript View this table: [Supplementary Table 1](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/T3) Supplementary Table 1 Clinical criteria for Nociceptive pain (based on 15) View this table: [Supplementary Table 2](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/T4) Supplementary Table 2 Clinical criteria for Neuropathic pain (based on 18) View this table: [Supplementary Table 3](http://medrxiv.org/content/early/2024/08/14/2024.08.13.24311924/T5) Supplementary Table 3 Clinical criteria for Nociplastic pain (based on 16) * Received August 13, 2024. * Revision received August 13, 2024. * Accepted August 14, 2024. * © 2024, Posted by Cold Spring Harbor Laboratory The copyright holder for this pre-print is the author. All rights reserved. The material may not be redistributed, re-used or adapted without the author's permission. ## References 1. Global Burden of Disease Injury, Injury, Incidence Prevalence, Collaborators. Global, regional, and national incidence, prevalence, and years lived with disability for 310 diseases and injuries, 1990-2015: a systematic analysis for the Global Burden of Disease Study 2015. Lancet 388, 1545–1602, doi:10.1016/S0140-6736(16)31678-6 (2016). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/S0140-6736(16)31678-6&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=27733282&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) 2. Dieleman, J. L. et al. US Health Care Spending by Payer and Health Condition, 1996-2016. JAMA 323, 863–884, doi:10.1001/jama.2020.0734 (2020). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1001/jama.2020.0734&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=http://www.n&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) 3. Hayden, J. A., Ellis, J., Ogilvie, R., Malmivaara, A. & van Tulder, M. W. Exercise therapy for chronic low back pain. Cochrane Database Syst Rev 9, CD009790, doi:10.1002/14651858.CD009790.pub2 (2021). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1002/14651858.CD009790.pub2&link_type=DOI) 4. Hartvigsen, J. et al. What low back pain is and why we need to pay attention. Lancet 391, 2356–2367, doi:10.1016/S0140-6736(18)30480-X (2018). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/S0140-6736(18)30480-X&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=29573870&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) 5. Dima, A. et al. Identifying patients’ beliefs about treatments for chronic low back pain in primary care: a focus group study. Br J Gen Pract 63, e490–498, doi:10.3399/bjgp13X669211 (2013). [Abstract/FREE Full Text](http://medrxiv.org/lookup/ijlink/YTozOntzOjQ6InBhdGgiO3M6MTQ6Ii9sb29rdXAvaWpsaW5rIjtzOjU6InF1ZXJ5IjthOjQ6e3M6ODoibGlua1R5cGUiO3M6NDoiQUJTVCI7czoxMToiam91cm5hbENvZGUiO3M6NDoiYmpncCI7czo1OiJyZXNpZCI7czoxMToiNjMvNjEyL2U0OTAiO3M6NDoiYXRvbSI7czo1MDoiL21lZHJ4aXYvZWFybHkvMjAyNC8wOC8xNC8yMDI0LjA4LjEzLjI0MzExOTI0LmF0b20iO31zOjg6ImZyYWdtZW50IjtzOjA6IiI7fQ==) 6. Tagliaferri, S. D. et al. Classification Approaches for Treating Low Back Pain Have Small Effects That Are Not Clinically Meaningful: A Systematic Review With Meta-analysis. J Orthop Sports Phys Ther 52, 67–84, doi:10.2519/jospt.2022.10761 (2022). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.2519/jospt.2022.10761&link_type=DOI) 7. Shraim, M. A., Masse-Alarie, H., Hall, L. M. & Hodges, P. W. Systematic Review and Synthesis of Mechanism-based Classification Systems for Pain Experienced in the Musculoskeletal System. Clin J Pain 36, 793–812, doi:10.1097/AJP.0000000000000860 (2020). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1097/AJP.0000000000000860&link_type=DOI) 8. Gifford, L. S. & Butler, D. S. The integration of pain sciences into clinical practice. J Hand Ther 10, 86–95. (1997). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/S0894-1130(97)80063-4&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=9188028&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) 9. Woolf, C. J., American College of, P. & American Physiological, S. Pain: moving from symptom control toward mechanism-specific pharmacologic management. Ann Intern Med 140, 441–451, doi:10.7326/0003-4819-140-8-200404200-00010 (2004). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.7326/0003-4819-140-8-200404200-00010&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=15023710&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000220204800005&link_type=ISI) 10. International Association for the Study of Pain. Terminology, <[https://www.iasp-pain.org/resources/terminology/](https://www.iasp-pain.org/resources/terminology/)> (accessed 11th August 2024) 11. Smart, K., O’Connell, N. & Doody, C. Towards a mechanisms-based classification of pain in musculoskeletal physiotherapy? Physical Therapy Reviews 13, 1–10 (2008). 12. Kosek, E. et al. Do we need a third mechanistic descriptor for chronic pain states? Pain 157, 1382–1386, doi:10.1097/j.pain.0000000000000507 (2016). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1097/j.pain.0000000000000507&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=http://www.n&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) 13. Vardeh, D., Mannion, R. J. & Woolf, C. J. Toward a Mechanism-Based Approach to Pain Diagnosis. J Pain 17, T50–69, doi:10.1016/j.jpain.2016.03.001 (2016). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.jpain.2016.03.001&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=http://www.n&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) 14. Shraim, M. A., Masse-Alarie, H. & Hodges, P. W. Methods to discriminate between mechanism-based categories of pain experienced in the musculoskeletal system: a systematic review. Pain 162, 1007–1037, doi:10.1097/j.pain.0000000000002113 (2021). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1097/j.pain.0000000000002113&link_type=DOI) 15. Shraim, M. A. et al. Features and methods to discriminate between mechanism-based categories of pain experienced in the musculoskeletal system: a Delphi expert consensus study. Pain 163, 1812–1828, doi:10.1097/j.pain.0000000000002577 (2022). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1097/j.pain.0000000000002577&link_type=DOI) 16. Kosek, E. et al. Chronic nociplastic pain affecting the musculoskeletal system: clinical criteria and grading system. Pain 162, 2629–2634, doi:10.1097/j.pain.0000000000002324 (2021). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1097/j.pain.0000000000002324&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=http://www.n&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) 17. Nijs, J. et al. Nociceptive, neuropathic, or nociplastic low back pain? The low back pain phenotyping (BACPAP) consortium’s international and multidisciplinary consensus recommendations. Lancet Rheumatol 6, e178–e188, doi:10.1016/S2665-9913(23)00324-7 (2024). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/S2665-9913(23)00324-7&link_type=DOI) 18. Finnerup, N. B. et al. Neuropathic pain: an updated grading system for research and clinical practice. Pain 157, 1599–1606, doi:10.1097/j.pain.0000000000000492 (2016). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1097/j.pain.0000000000000492&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=27115670&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) 19. Lawson, R. G. & Jurs, P. C. New Index for Clustering Tendency and Its Application to Chemical Problems. Journal of Chemical Information and Computer Sciences 30, 36–41 (1990). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1021/ci00065a010&link_type=DOI) 20. Saxena, A. et al. A review of clustering techniques and developments. Neurocomputing 267, 664–681 (2017). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.neucom.2017.06.053&link_type=DOI) 21. Rousseeuw, P. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. J Comput Appl Math 20, 53–65 (1987). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/0377-0427(87)90125-7&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=WOS:A1987L11&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=A1987L111800005&link_type=ISI) 22. Fraley, C. & Raftery, A. E. Model-based clustering, discriminant analysis, and density estimation. Journal of the American Statistical Association 97, 611–631 (2002). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1198/016214502760047131&link_type=DOI) [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000176078300027&link_type=ISI) 23. Lloyd, S. Least squares quantization in PCM. IEEE transactions on information theory 28, 129–137 (1982). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=https://doi.org/10.1109/TIT.1982.1056489&link_type=DOI) 24. Szekely, G. J. & Rizzo, M. L. Hierarchical clustering via joint between-within distances: Extending Ward’s minimum variance method. Journal of Classification 22, 151–184 (2005). 25. Kearns, M., Mansour, Y. & Ng, A. Y. An information-theoretic analysis of hard and soft assignment methods for clustering. Learning in Graphical Models, 495–520 (1988). 26. Murtagh, F. & Contreras, P. Algorithms for hierarchical clustering: an overview. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 2, 86–97 (2012). 27. Geisser, M. E. et al. Comorbid somatic symptoms and functional status in patients with fibromyalgia and chronic fatigue syndrome: sensory amplification as a common mechanism. Psychosomatics 49, 235–242, doi:10.1176/appi.psy.49.3.235 (2008). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1176/appi.psy.49.3.235&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=18448779&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) 28. Clauw, D. J. Fibromyalgia: a clinical review. JAMA 311, 1547–1555, doi:10.1001/jama.2014.3266 (2014). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1001/jama.2014.3266&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=24737367&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) 29. Mayer, T. G. et al. The development and psychometric validation of the central sensitization inventory. Pain Pract 12, 276–285, doi:10.1111/j.1533-2500.2011.00493.x (2012). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1111/j.1533-2500.2011.00493.x&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=21951710&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000302468900003&link_type=ISI) 30. Cohen, M. L. Proposed clinical criteria for nociplastic pain in the musculoskeletal system are flawed. Pain 163, e606–e607, doi:10.1097/j.pain.0000000000002505 (2022). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1097/j.pain.0000000000002505&link_type=DOI) 31. Jensen, T. S. et al. A new definition of neuropathic pain. Pain 152, 2204–2205, doi:10.1016/j.pain.2011.06.017 (2011). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.pain.2011.06.017&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=21764514&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) [Web of Science](http://medrxiv.org/lookup/external-ref?access_num=000295213700008&link_type=ISI) 32. Nicholas, M. et al. The IASP classification of chronic pain for ICD-11: chronic primary pain. Pain 160, 28–37, doi:10.1097/j.pain.0000000000001390 (2019). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1097/j.pain.0000000000001390&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=30586068&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) 33. Smart, K. M., Blake, C., Staines, A., Thacker, M. & Doody, C. Mechanisms-based classifications of musculoskeletal pain: part 1 of 3: symptoms and signs of central sensitisation in patients with low back (+/- leg) pain. Man Ther 17, 336–344, doi:10.1016/j.math.2012.03.013 (2012). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.math.2012.03.013&link_type=DOI) [PubMed](http://medrxiv.org/lookup/external-ref?access_num=22534654&link_type=MED&atom=%2Fmedrxiv%2Fearly%2F2024%2F08%2F14%2F2024.08.13.24311924.atom) 34. Klyne, D. M., Moseley, G. L., Sterling, M., Barbe, M. F. & Hodges, P. W. Are signs of central sensitisation in acute low back pain a precursor to poor outcome? J Pain, doi:10.1016/j.jpain.2019.03.001 (2019). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1016/j.jpain.2019.03.001&link_type=DOI) 35. Meacham, K., Shepherd, A., Mohapatra, D. P. & Haroutounian, S. Neuropathic Pain: Central vs. Peripheral Mechanisms. Curr Pain Headache Rep 21, 28, doi:10.1007/s11916-017-0629-5 (2017). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1007/s11916-017-0629-5&link_type=DOI) 36. Wang, F. et al. Nociceptor-localized cGMP-dependent protein kinase I is a critical generator for central sensitization and neuropathic pain. Pain 162, 135–151, doi:10.1097/j.pain.0000000000002013 (2021). [CrossRef](http://medrxiv.org/lookup/external-ref?access_num=10.1097/j.pain.0000000000002013&link_type=DOI)