A systematic review with Bayesian modelling of the prevalence of pain with neuropathic characteristics
1Brain Function Research Group, Department of Physiology, School of Biomedical Sciences, Faculty of Health Sciences, University of the Witwatersrand, South Africa
*Corresponding author: Prof Peter R Kamerman, Brain Function Research Group, Department of Physiology, School of Biomedical Sciences, Faculty of Health Sciences, University of the Witwatersrand, 7 York Road, Parktown, 2193, South Africa Tel: +27 (0)11 717 2363, Email: peter.kamerman@wits.ac.zaAbstract
We performed the first systematic review and meta-analysis of the prevalence of pain with neuropathic characteristics using Bayesian methods to correct prevalence estimates for the use of screening tools with imperfect sensitivity and specificity (CRD42023416845). We searched major databases for national or regional epidemiological studies that reported the prevalence of pain with neuropathic characteristics, as identified by the PainDETECT, S-LANSS, or DN4-interview. Of the 1,251 unique records retrieved, 8 were finally extracted. The uncorrected (apparent) prevalence data were pooled using a random-effects meta-analysis for proportions. The corrected (true) prevalence was estimated using Bayesian models incorporating sensitivity and specificity distributions under non-informative [beta(1,1)] and informative priors [beta(4.389, 29.522); based on apparent prevalence]. Using the mean values from Bayesian credible intervals, a pooled estimate of true prevalence was generated using a random-effects model. The pooled estimate for the apparent prevalence was 10.6% (95% CI: 8.5; 12.9). The pooled estimate for true prevalence was 4.9% (95% CI: 3.8; 6.1) using informative priors, and 2.3% (95% CI: 1.5; 3.2) using non-informative priors. The use of imperfect screening tools may have overestimated the prevalence of neuropathic pain.
Perspective
The prevalence of neuropathic pain may be lower than previously estimated. A lower prevalence should not be equated with reduced societal or clinical significance, but it may have implications for healthcare resource allocation and research funding policies for neuropathic pain.
Article notes
Competing Interest Statement
P.R.K. is the sole proprietor of Blueprint Analytics and is a consultant for Partners in Research. The remaining authors have no conflicts of interest to declare.
Clinical Protocols
Funding Statement
N.M. was receiving funding from the National Research Foundation of South Africa.
1.Introduction
Estimates of the prevalence of neuropathic pain in the general adult population are flawed because the instruments used to screen for the presence of neuropathic pain only assess for pains with neuropathic characteristics, which does not necessarily mean the pain is indeed neuropathic, and the instruments are not perfect tools. For example, screening tools such as the self-report Leeds Assessment of Neuropathic Symptoms and Signs (S-LANSS), which was developed for epidemiological studies 7, and the interview-only Douleur Neuropathique en 4 questions (DN4-interview), which was suggested as being suitable for epidemiological studies 9, both have sensitivities and specificities greater than 0.7, but less than 0.9 in their validation populations. That is, they are imperfect diagnostic tools.
When the prevalence of a disease is low (e.g., 3 to 17% of the adult population is estimated to have neuropathic pain) 25, and screening tool sensitivity and specificity are moderate to high (e.g., tools developed to screen for pain with neuropathic features), then the number of false positives is typically numerically greater than the number of false negatives, and prevalence estimates are inflated. From a frequentist statistical standpoint, prevalence estimates can be corrected for the use of imperfect screening tools using the Rogan-Gladen estimator 35. However, this estimator has significant limitations, such as it assumes fixed sensitivity and specificity values for a questionnaire across populations, and at prevalence extremes, it yields invalid results (prevalence estimates less than zero or greater than one) 17,37. An alternative approach is to use Bayesian estimation, which allows the modelling of the sensitivity and specificity, and the prevalence estimate within the boundaries of zero and one. The method also allows the incorporation of prior knowledge on the prevalence in the estimate of true prevalence 17,37.
Consequently, we conducted a systematic review of the literature on the prevalence of pain with neuropathic features in the general adult population (using regional or country-level studies only), and then applied Bayesian estimation methods to estimate the corrected prevalence of pain with neuropathic features in the general adult population.
2.Methods
The protocol for this review was registered on the PROSPERO database (CRD42023416845), and the results are reported according to the recommendations of the Preferred Reporting Items for Systematic Review and Meta-Analysis (PRISMA) guideline 34; the checklist is available in the Supplementary file: Section 1.
2.1Eligibility criteria
Using the Population, Phenomena of Interest and Context (PICo) framework, our research question was, “What is the prevalence of neuropathic pain in the general adult population?”, and our selection criteria were as follows:
2.1.1Inclusion criteria
- An epidemiological study investigating disease (neuropathic pain) prevalence.
- The study was conducted at the level of a geographical region (e.g., Great Britain), country, or state/province.
- The study was conducted in the general adult population (≥18 years of age).
- Neuropathic pain was assessed as a primary or secondary outcome of the study.
- Neuropathic pain was screened for using the full Douleur Neuropathique en 4 questions (DN4) or the shortened interview only version (DN4-interview) 9; the Leeds Assessment of Neuropathic Symptoms and Signs (LANSS) 6 or the self-report version of the LANSS (S-LANSS) 7; or the painDETECT 18 screening tools.
- The study was published in English.
2.1.2Exclusion criteria
- The study focused on one type/cause of neuropathic pain only (e.g., painful diabetic neuropathy only).
- The study was conducted on a sub-population (e.g., presence of neuropathic pain in adults with chronic pain).
2.2Information sources and search strategy
We queried three electronic databases: PubMed, Scopus, and Web of Science (search strategies in the Supplementary file: Section 2). The initial search was conducted on 14 April 2023 and spanned the period: 1 January 2001 (the year the first screening tool—LANSS—was published) to 31 March 2023. A second search was conducted on 5 February 2025 and spanned the period: 1 April 2023 to 4 February 2025. Reference lists of papers that passed the initial author-abstract screening process (see below) were assessed to identify additional articles. Grey literature sources were not assessed.
2.3Selection process and data collection
The results returned from each database were uploaded to Covidence (https://www.covidence.org), an online platform for managing systematic reviews, and duplicates were removed using the built-in automatic duplicate removal tool that uses title, year, volume number, and author similarities to detect duplicates. Once duplicate articles had been removed, articles underwent title and abstract screening, followed by article retrieval and full text screening. Finally, the included articles underwent data extraction and quality assessment. The screening processes were performed in duplicate by four reviewers working independently of each other (PRK, TH, NM, and PCC). When there were conflicts, a third reviewer from the group made the final decision. Data extraction and quality assessment were also performed in duplicate by two independent reviewers (TH and NM), with PRK checking for consensus and resolving conflicts between the two reviewers.
A copy of the extraction template used in Covidence is available in the Supplementary file: Section 3. Key items extracted were: publication date; study location (country and region/state/province); sample frame; the criteria used to define whether pain was chronic; the neuropathic pain screening tool used and the threshold used for defining caseness; the number of people approached to take part; the number of people recruited and who completed the questionnaires; the number of people with no pain, acute pain, chronic pain, and chronic pain with neuropathic features. Where sample weights were used by the authors, the raw and weighted sample sizes were recorded.
Study quality and risk of bias were assessed on a Covidence template using the Joanna Briggs Institute’s (JBI) Checklist for Prevalence Studies [28,33. A copy of the quality assessment template is available in the Supplementary file: Section 4. The tool includes nine questions, each answered “yes”, “no”, or “unclear”. The checklist does not produce a score, and hence, there is no cut-off score for deciding whether a study should be included in a review. Rather, the evidence is weighed over the questions, and a subjective decision to include a paper is made based on the balance of the evidence.
2.4Effect measure
The effect measure of interest was the prevalence of pain with neuropathic features in the study sample.
2.5Reporting bias and certainty assessments
No assessment of reporting bias and certainty was made.
2.6Data analysis
All analyses and plotting were performed in the R statistical environment (v4.4.2) using functions from the following packages: dplyr [47, epiR38, forcats45, ggdist29, ggplot246, meta4, prevalence11, purrr48, and tidyr49.
All analysis scripts and the data used in these scripts are available at: https://github.com/kamermanpr/neuropathic-pain_prevalence/ and mirrored at Zenodo.org (https://doi.org/10.5281/zenodo.17510858). In addition, a Docker image has been created for the R environment used in these analyses (https://hub.docker.com/repository/docker/kamermanpr/neuropathic-pain_prevalence/).
Instructions on how to use the Docker image to recapitulate or modify our analyses are provided in the README file on the GitHub/Zenodo repositories. Outputs from the Bayesian models were too large to upload to GitHub, so they were uploaded to a separate Zendo repository (https://doi.org/10.5281/zenodo.15591085).
The analysis was conducted as follows:
2.6.1Apparent prevalence
- Apparent prevalence (i.e., no correction for screening tool sensitivity and specificity) was calculated using the number of people detected as having pain with neuropathic features (as detected using screening tools) and the total number of people reported to have completed the questionnaires. For the two studies where population weights were applied to the data, the scaled sample sizes were used instead of the raw numbers12,43. Following data extraction, the proportion (95% confidence interval; calculated using the exact binomial method) of people with pain with neuropathic features was calculated for each study.
- A random effects model meta-analysis for proportions was used to estimate the apparent prevalence of pain with neuropathic features across the eight studies. Details of meta-analysis method: i) an arcsine transformation was applied to individual study proportions before the calculation of the overall proportion using the inverse variance method (data were back transformed for reporting), iii) a restricted maximum-likelihood estimator was used to calculate the between-study variance (tau-squared), iv) the Q-Profile method was used to calculate the confidence interval of tau-squared and its square root, tau, and v) the calculation of I-squared was based on Q.
The decision to use a random effect model was made before the analysis was undertaken, and was selected ahead of a fixed effects model because we felt that differences in the populations assessed (i.e., population structure), the methods of recruitment (postal versus telephonic solicitation), the methods of measurement (three different screening tools were used), and the methods of questionnaire delivery (e.g., entirely self-complete versus directed verbal instructions) meant that it was unlikely that a single true effect could be assessed.
2.6.2True prevalence
- Because the sensitivity and specificity values of a screening tool are not universal, and apply only for the sample of people and context in which the tool was developed8,27, we generated distributions of possible sensitivity and specificity values for each questionnaire for inclusion in the Bayesian modelling of true prevalence (see below). The distributions of possible sensitivity and specificity values for each screening tool were modelled using a beta distribution, with the two shape parameters for each distribution being determined using the “beta.buster” function in the R “epiR” package. The R script to generate the distributions is located in the “03_betas.R” file at https://doi.org/10.5281/zenodo.17510858. In brief, for each published validation of the screening tools used in the studies included in this review, we extracted the point estimates for the sensitivity and specificity of the screening tool at the scoring threshold used in the reviewed studies. These values were specified as the most likely sensitivity and specificity values for a particular screening tool, with 90% confidence that the sensitivity/specificity value was greater than the sensitivity/specificity value - 0.1. For example, the published sensitivity for a score ≥ 19 on the PainDetect instrument is 0.8418, which was specified as the most likely value, and we were 90% confident that the sensitivity value was greater than 0.74.
- We used the “truePrev” function from the R “prevalence” package to calculate the true prevalence for each of the reviewed studies. The values specified for the number of people with pain with neuropathic features and the total sample size were the same as those used in the calculation of the apparent prevalences (see above). In addition, the distributions of sensitivity and specificity values for each screening tool were specified using the beta distributions that had been calculated (see above). The data were then modelled under two conditions:
The modelling started with 5 chains using a burn-in of 10,000 iterations, and an update of 100,000 iterations. The convergence of each model was then checked using the Brooks-Gelman-Rubin statistic, with convergence accepted at values < 1.2. If convergence was not attained, the burn-in and update iterations were each increased by a factor of 10 until convergence was achieved.The final model specifications for each study, and the model outputs, can be viewed in the R script files: i) “04_prevalence_true_bayesian_uniformative.R”, and ii) “05_prevalence_true_bayesian_informative.R” (https://doi.org/10.5281/zenodo.17510858)
- An uniformative prior distribution, specified as Beta(1, 1),
- An informative prior distribution, specified as Beta(4.389, 29.52249). The informative prior distribution was generated using the “beta.buster” function in the R “epiR” package, with distribution assumptions based on the results of the meta-analysis conducted using the apparent prevalence point estimate, namely, the most likely value = 0.1062 (10.62%), with 90% confidence that the true prevalence was less than 0.2062; see the Results section below for detailed findings of the apparent prevalence meta-analysis.
- The results of the Bayesian estimates of true prevalence for each study were plotted as the point estimate with 66% and 95% high-density intervals (certainty intervals). The boundaries of the high-density intervals were determined using the quantile method.
- To perform a meta-analysis of the true prevalence data, we multiplied the total number of people reported to have completed the questionnaires in each study by the calculated point estimate of true prevalence for a study. The resulting values were then included in the meta-analysis model as the number of people with pain with neuropathic features for each study. The meta-analysis was completed using the same methods described in point 2 of the apparent prevalence section.
3.Results
3.1Search results
Figure 1 provides an overview of the article selection process. After duplicate removal and title-abstract screening, 21 articles underwent full-text screening. Twelve of these 21 studies were deemed not to fit the inclusion criteria; a list of the 12 studies and the reason for their exclusion is provided in Section 5 of the Supplementary file1,3,13–15,20,24,32,36,39–41. The remaining 9 studies underwent data extraction5,10,12,21,23,26,42,43,50. However, during the extraction process, it became evident that sensitivity and specificity data for the screening tool used in one of the studies (Japanese version of the painDETECT) were not available26. Only the specificity of the tool was forthcoming after emailing the authors, so the study was excluded at the data extraction phase.
3.2Study characteristics
The characteristics of the eight studies that were included in the review are provided in Table 1. One of the studies, VanDenKerkhof et al., 201643, screened the recruited population for the presence of chronic pain with neuropathic features using the DN4-interview and the S-LANSS; we decided to treat the results from the two screening tools as two independent datasets, with both datasets being included in our analyses.
One study each was conducted in France10 and in Morocco21, and two studies were conducted in each of Canada42,43, the United Kingdom5,23, and the United States of America12,50. Five studies were national studies5,10,12,21,43, with the rest being regional. Six studies used the DN4-interview5,10,21,23,42,43, two studies used the S-LANSS43,50, and one study used the painDETECT12. In all cases, the published recommended screening tool scores were used to define whether a participant had pain with neuropathic features (DN4-interview: ≥ 39,22, painDETECT: ≥ 1918, and S-LANSS: ≥ 127). Similarly, all studies used pain lasting ≥ 3 months to classify whether pain was chronic. The study by VanDenKerkhof et al.43 included the additional criterion that the pain was associated with a likely (e.g., accident with nerve damage, amputation, and shingles) or probable (e.g., surgery more than 3 months ago, back problem, diabetes, cancer) insult to the somatosensory nervous system that could result in neuropathic pain. The point estimates of the sensitivity and specificity of the three screening tools were extracted from their respective validation studies (i.e., DN4-interview in French9, DN4-interview in Arabic22, S-LANSS in English7). Where no validation study was available for the language version of the screening tool used in a survey, the sensitivity and specificity values were taken from the original validation study (e.g., an unvalidated English translation of the original French version of the DN4-interview was used in four of the surveys5,23,42,43, so the sensitivity and specificity reported in the French validation study were used 9).
In total, across the 8 studies, approximately 1,411,552 people were approached to participate in the surveys (largest: ∼1,000,00012, smallest: 4,16742). Of those approached, 235,479 people responded (average response rate: 16.7%; lowest: 2.5%12, highest: 81.2%10), and 214,925 people had complete data that could be analysed (average completion rate: 15.2%; lowest: 2.5%12, highest: 78.6%10). Across the 9 datasets (VanDenKerkhof et al.43 contributed 2 datasets), the raw—no correction for population structure—total number of people fitting the criteria for chronic pain with neuropathic features was 20,217, giving a raw prevalence of 9.3%.
3.3Study quality and risk of bias
Results of the JBI Checklist for Prevalence Studies are provided in Table 2. Overall, there were limitations in several study quality metrics, particularly the reporting of standardisation methods for questionnaire administration, the appropriateness of the statistical analysis, and low response rates. However, no studies were excluded based on the overall balance of the evidence quality.
3.4Apparent prevalence
Figure 2 shows a forest plot of apparent prevalence (i.e., no correction for screening tool sensitivity and specificity) for the eight studies included in the analysis. The highest prevalence reported was by Toth et al. [17.9% (95% confidence interval: 18.8; 20.2)]42, and the lowest prevalence was reported by Bouhassira et al. [6.9% (95% confidence interval: 6.6; 7.2)]10. Overall, the random effects model returned a pooled estimate for apparent prevalence of 10.6% (95% confidence interval: 8.5; 12.9).
3.5True prevalence
Figure 3 shows the mean and credible intervals of true prevalence (i.e., after correcting for screening tool sensitivity and specificity) for the eight studies using an informative prior (top panel), and a non-informative prior (bottom panel). In both scenarios, the highest prevalence reported was by Toth et al. [informative prior: 8.7% (95% credible interval: 3.1; 15.7); non-informative prior: 6.0% (95% credible interval: 0.3; 15.5)]42, and the lowest prevalence was reported by Bouhassira et al. [informative prior: 3.3% (95% credible interval: 1.2; 6.0)]10, and VanDenKerkhof et al. [non-informative prior: 1.2% (95% credible interval: 0.0; 4.0)]43.
Figure 4 shows a forest plot of true prevalence for the eight studies using an informative prior (top panel), and a non-informative prior (bottom panel); the number of events was based on the total sample multiplied by the mean proportion shown in Figure 3 [e.g., Toth et al.: 0.087 (mean for informative prior) x 1,207 (sample size) = 105 events]. In both scenarios, the highest prevalence reported was by Toth et al. [informative prior: 8.7% (95% confidence interval: 7.2; 10.4); non-informative prior: 6.0% (95% confidence interval: 4.8; 7.5)]42, and the lowest prevalence was reported by Bouhassira et al. [informative prior: 3.3% (95% confidence interval: 3.1; 3.5)]10, and VanDenKerkhof et al. [non-informative prior: 1.2% (95% confidence interval: 1.2; 1.2)]43.
Overall (Figure 4), the pooled estimates from the meta-analysis was 4.9% (95% confidence interval: 3.8; 6.1) when using an informative prior, and 2.3% (95% confidence interval: 1.5; 3.2) when using a non-informative prior.
4.Discussion
We undertook to provide the first-ever estimate of the true prevalence of pain with neuropathic characteristics in the general adult population by conducting a systematic review of the field and then using Bayesian methods to provide a more accurate estimate of the true prevalence of the condition by correcting apparent prevalences generated through the use of imperfect screening tools. Using data from eight studies that met our inclusion criteria, the apparent prevalence was 10.6% (95% confidence interval: 8.5; 12.9). Correcting for screening tool sensitivity and specificity shifted the estimate of true prevalence substantially lower [informative prior: 4.9% (95% confidence interval: 3.8; 6.1); non-informative prior 2.3% (95% confidence interval: 1.5; 3.2)].
A limitation of our analytical approach when calculating the pooled estimate of true prevalence is that we only used the mean density of the credible interval for each study when back-calculating the number of cases for the meta-analysis. Thus, we did not perform meta-analyses for all possible combinations of major landmarks in the credible intervals (e.g., mean, 66% interval, 95% interval) across the studies. However, we judged that this large undertaking would only yield incremental improvements in the quality of our estimates. Instead, we present the credible intervals for each study using an informative and non-informative prior, and the pooled estimate from the meta-analysis represents the true prevalence based on the most likely estimates from the Bayesian analysis. A related limitation is our choice of an informative prior, which in this case was the apparent prevalence for each study. Doing so meant that we used prevalence estimates that we state are incorrect to inform a revised estimate, and the posterior distributions are thus informed by inappropriate priors. Nevertheless, we felt it important to include at least two Bayesian models using different prior assumptions.
While we believe our Bayesian approach provides better estimates of the prevalence of pain with neuropathic characteristics in the adult population, we still believe that our values may overestimate the true prevalence of pain with neuropathic characteristics. We reason that while our analysis approach corrects for the use of imperfect screening tools, inherent deficiencies across all of the included studies could lead to a positive reporting bias (Table 2). That is, there was a high prevalence of unclear or deficient sampling frames, unclear or absent controls for coverage bias and adjustment to reference populations, and response rates were generally low to very low. For example, the UK Biobank is not a representative sample of the UK population19, and the respondents to the pain rephenotyping sub-study differed from the parent sample5. Yet, no adjustments were made to the sample in the rephenotyping study, which would have improved the generalisability of the findings2,30,44. Another shortcoming of the data is that it is heavily biased to high-income, western countries (Europe and North America), with only a single study coming from a low- or middle-income country (Morocco21). The data are therefore representative of a Western, high-income country context and may not be globally representative.
We also draw attention to the fact that our true prevalence estimates of pain with neuropathic characteristics probably overestimate the true prevalence of neuropathic pain in the adult population. The co-occurrence of certain pain symptoms can be suggestive that a pain is neuropathic in origin; however, these symptom clusters are not pathognomic. Indeed, in the clinical context (as opposed to the epidemiological context)31, a combination of particular co-occurring symptoms, a relevant medical history, and a plausible neuroanatomical distribution are all required for the grading of a pain as being possible neuropathic pain; the lowest level of diagnostic certainty16. One study in our sample factored causes of neuropathic pain into their case definition43, but that still means that none of the studies met the minimum level of evidence to define the presence of possible neuropathic pain instead of pain with neuropathic characteristics. We propose that any epidemiological studies to estimate the prevalence of neuropathic pain as such, and not pain with neuropathic features, must include not only symptom screens, but also solicit information on medical history (e.g., the presence of a preceding injury or disease of the somatosensory nervosus system that could explain the pain) and site of pain (e.g., body charts)31.
In conclusion, we used a Bayesian methodology to estimate the true prevalence of pain with neuropathic characteristics. Using this method, we show that our pooled estimates of true prevalence are about 75% (non-informative prior) and 50% (informative prior) lower than the pooled estimate of apparent (uncorrected) prevalence. The substantial downward revision of prevalence estimates presented here compels a reassessment of how the burden of neuropathic pain is conceptualised in population health. However, a lower prevalence should not be equated with reduced societal or clinical significance. Neuropathic pain remains a debilitating condition that impacts quality of life, functional capacity and mental health. Its effects influence caregiver responsibilities, workforce participation, and healthcare resource use. As such, these revised estimates argue not for the reprioritisation of neuropathic pain in research and policy agendas, but for more precise, evidence-informed strategies to identify, manage and mitigate its impact.
Supporting information
CONTRIBUTIONS
PRK: Conceptualisation, data curation, formal analysis (including directly accessing and verifying the data), investigation, supervision, visualisation, and writing – original draft. TH: Formal analysis (including directly accessing and verifying the data), investigation, visualisation, and writing – review & editing. NM: Investigation, and writing – review & editing. PCC: Investigation, and writing – review & editing. All authors had full access to all the data in the study and had final responsibility for the decision to submit for publication.
CONFLICTS OF INTEREST
P.R.K. is the sole proprietor of Blueprint Analytics and is a consultant for Partners in Research. The remaining authors have no conflicts of interest to declare.
DATA SHARING
All data, analysis scripts, and the analysis environment are available at: https://github.com/kamermanpr/neuropathic-pain_prevalence (mirrored at https://doi.org/10.5281/zenodo.17510858), https://hub.docker.com/repository/docker/kamermanpr/neuropathic-pain_prevalence, and https://doi.org/10.5281/zenodo.15591085.
ACKNOWLEDGEMENTS
NM was receiving funding from the National Research Foundation of South Africa.