Neurobiology of osteoarthritis: a systematic review and activation likelihood estimation meta-analysis

Osteoarthritis (OA) affects 240 million people worldwide. Neuroimaging has been increasingly used to investigate brain changes in OA, however, there is considerable heterogeneity in reported results. The goal of this systematic review and meta-analysis was to synthesise existing literature and identify consistent brain alterations in OA. Six databases were searched from inception up to June, 2022. Full-texts of original human studies were included if they had: (i) neuroimaging data by site of OA (e.g. hand, knee, hip); (ii) data in healthy controls (HC); (iii) > 10 participants. Activation likelihood estimation (ALE) was conducted using GingerALE software on studies that reported peak activation coordinates and sample size. Our search strategy identified 6250 articles. Twenty-eight studies fulfilled the eligibility criteria, of which 18 were included in the meta-analysis. There were no significant differences in brain structure or function between OA and healthy control contrasts. In exploratory analysis, the right insula was associated with OA vs healthy controls, with less activity, connectivity and brain volume in OA. This region was implicated in both knee and hip OA, with an additional cluster in the medial prefrontal cortex observed only in the contrast between healthy controls and the hip OA subgroup, suggesting a possible distinction between the neural correlates of OA subtypes. Despite the limitations associated with heterogeneity and poor study quality, this synthesis identified neurobiological outcomes associated with OA, providing insight for future research. PROSPERO registration number: CRD42021238735.

Data extraction. Four reviewers independently extracted data (MH, NEB, CJZ, YL) and verified data by cross-checking from all included studies. The following information was extracted: authors, publication year, type of study design, number of participants by sex, age, body mass index, disease severity, outcomes, brain regions of interest, networks, stereotactic coordinates. If multiple related contrasts were reported, we included all contrasts but handled them as one experiment, thereby using only one set of coordinates in the meta-analysis. If further information was needed, authors were contacted at least twice via email, after which data were considered irretrievable.
Data synthesis and analysis. To perform coordinate-based meta-analysis, activation likelihood estimation (ALE) analysis was conducted using GingerALE, version 3.0.2 (https:// www. brain map. org/ ale/). Studies included in the ALE analysis reported peak activation coordinates in Montreal Neurological Institute (MNI) or Talairach space and sample size. Within each experiment, the reported activation foci were treated as centres of a three-dimensional Gaussian probability distribution, whose width is determined by the study's sample size and thus reflective of spatial uncertainty of the foci 12 . As larger samples model smaller Gaussian distributions, they are also likely to produce more reliable approximations of the "true" activation effect. Then, these modelled possibilities were combined across foci, producing a modelled activation map for each experiment. To test for spatial convergence of neuroimaging findings, voxel-wise ALE scores were calculated by taking the union of all modelled activation maps. Statistically significant convergence between experiments was identified by comparing the ALE scores against a null distribution of random spatial association, with the outcome clusters representing above-chance convergence between experiments. Correction level was set to p < 0.001, 1000 permutations and p < 0.05 cluster-level family-wise error (FWE). For illustration, the resulting ALE maps were imported to MRIcron (https:// people. cas. sc. edu/ rorden/ mricr on/ insta ll. html) and plotted over a standardised anatomical MNI-normalised template.
First, the primary contrast between OA vs healthy controls was performed. Six subsequent, exploratory analyses were performed as in previous similar ALE analysis, e.g. in fibromyalgia 13 , to contrast OA vs. healthy controls in the direction of effect as follows: (1) OA greater than healthy controls contrast (e.g. greater activation or brain volume in OA compared to healthy controls) and (2) OA less than healthy controls contrast (e.g. less activation or brain volume in OA compared to healthy controls). We also evaluated if there were imaging method-specific differences between OA and healthy controls, as follows: (3) contrast between OA vs. healthy controls as measured by resting-state functional magnetic resonance imaging; (4) contrast between OA vs. healthy controls as measured by structural MRI. Finally, we compared specific osteoarthritic joints: (5) knee OA vs. healthy controls; and (6) hip OA vs. healthy controls.
In light of the best practice guidelines for neuroimaging meta-analysis 10,11 , studies that did not report wholebrain analyses were not included. Furthermore, we conducted a pre-registered sensitivity analysis from the meta-analysis by removing studies that did not provide sufficient detail about their multiple comparison correction methods or that were not adequately corrected for multiple comparison, e.g. by reporting activation at a voxel-level threshold of p < 0.001 (uncorrected) with an additional cluster-level correction of p < 0.05. reviewers (FD and DMK) using the 14-item National Institute of Health Quality Assessment Tool for Observational Cohort and Cross-Sectional Studies 14 . Consensus on items was achieved by a staged learning approach where the reviewers met to check understanding initially after independently rating three studies, then again after completing all ratings. All conflicted items were then resolved by a final consensus. Overall study quality was assessed based on the potential risk of bias across four domains; (1) information bias (item 1), (2) selection bias (items 2, 3, 4), (3) measurement bias (items 5,6,7,8,9,10,11,12,13), and (4) confounding bias (item 14), according to the tool guidelines 14 . A domain was considered to be a "potential risk of bias" if at least one item within the domain received a "no" response, or most items in the domain received a "cannot determine/not reported" response. As the analyses included in this review were cross-sectional, items 6, 7, 10 and 13 within the measurement bias domain were either rated as "no" or "not applicable" for all studies. To account for this issue, we omitted these items for the scoring of domain bias and enforced a maximum study quality score of "moderate" (rather than "high"), i.e. when no domains were considered a potential source of bias. Study quality was downgraded to "low" or "very low" if one or more domains, respectively, were considered a potential source of bias.
Conference presentation. The results reported in this manuscript have been previously presented at the Australian Brain and Psychological Sciences Meeting, with the abstract published in the Conference Booklet. www. abps2 022. org/ wp-conte nt/ uploa ds/ 2022/ 07/ ABPS2 022-confe rence-bookl et. pdf.

Results
Our search strategy identified 6250 articles (Fig. 1). Twenty-eight studies fulfilled the eligibility criteria studies and 18 of these studies were included in meta-analysis. Characteristics of studies are described in Table 1 [36][37][38] , and two evaluated hip and knee OA 39,40 . A summary of the imaging outcomes is described in Table 2.

Discussion
The aims of this systematic review and meta-analysis were to (1) establish the evidence for alterations in structure and function of the brain in people with OA and (2) investigate the association between changes in brain structure and function and OA joints, pain severity, and duration. Our primary ALE meta-analysis did not show any differences in the brain structure or function between people with OA and healthy controls. Findings from our sensitivity analysis implicated the left post central gyrus in OA. Most studies evaluated knee OA, with only a few studies focusing on hip and hand OA. Findings for our exploratory ALE meta-analysis of studies that reported OA less than healthy controls contrasts converge with the narrative synthesis to suggest that the right anterior insula is the brain region that may be implicated in OA. People with OA may have less brain activity, connectivity and volume compared to healthy controls in this brain region. Indeed, the right anterior insula was implicated in knee OA and hip OA when compared separately to healthy controls. Notably, differences between hip OA compared to healthy controls were also observed in the medial prefrontal cortex. There was minimal evidence to suggest that pain intensity or pain duration associate with changes in brain structure and function. This systematic review was conducted in accordance with best practices of neuroimaging analysis 10,11 , yet the quality of studies informing the body of evidence was considered low. Thus, we have limited certainty in the robustness of our findings. The impetus for this systematic review and ALE meta-analysis was the observation of inconsistent results in studies investigating brain structure and function in OA, and the subsequent difficulty in selecting a marker(s) of brain structure and function to understand response to treatments for OA. Indeed, pooling data from all available studies for analysis did not reveal significant differences between those with OA and healthy controls. Although this finding may indeed suggest no difference in brain or function, differences may be undetectable. Symptom heterogeneity along with the heterogeneity of techniques used to assess the brain are possible explanations for the overall absence of differences between OA and healthy controls in our primary analysis. To overcome the issue of various MRI approaches, we isolated the ALE meta-analysis to specific techniques (e.g. MRI structural). However, no differences were observed which perhaps stems from the remaining issue of heterogeneity among the participants. One approach may be to assess subgroups of OA based on symptoms, as it could be reasonably speculated that people with more intense pain and/or longer duration of symptoms may have more pronounced brain adaptations. However, the challenge of identifying homogenous subgroups of people with OA is highlighted by the general lack of association between brain measures and pain characteristics including intensity and duration ( Table 3). The absence of association between potential markers of OA and clinical pain is an issue that extends beyond the brain imaging field (e.g. biomechanics 41 ), and again questions our rudimentary tools to assess pain (e.g. VAS, NRS). Notably, patients with OA struggle to self-describe pain with just "intensity" and describe numerous characteristics that vary in duration, depth, type of occurrence, impact and rhythm 42 .
The insula was most consistently implicated in several studies comparing OA and healthy controls, and also in association with pain intensity in our narrative review. Moreover, the right anterior insula emerged as significantly different between OA and healthy controls when including only studies that report differences of OA less than healthy controls. Although confirmatory studies are needed, we speculate these findings collectively suggest the insular cortex, and particularly the right anterior aspect may be implicated in the pathophysiology of OA. The insular cortex plays a role in somatosensory and pain processing in the central nervous system 43 and the anterior insula plays a role in emotion experience and subjective feeling associated with nociception 43 . Hence, the potentially lower right insular volume in OA compared to healthy controls might imply the dysfunction of the right insula in interoceptive awareness and emotionally relevant context for sensory experience that contributes to OA pain. The insula is connected to various other structures associated with pain processing including but not limited to the cingulate, para hippocampal, precuneus, amygdala, medial prefrontal cortex and occipital regions 44,45 , that were also identified as different to healthy controls, albeit less consistently. It remains unclear whether the potential alterations in the insula associated with OA drive adaptations to other structures and functions of the brain through its elaborate connectivity to many other structures.
We observed that knee and hip OA exploratory contrasts did not yield completely identical results. Specifically, hip OA was additionally associated with the cluster in medial prefrontal cortex, suggesting that there might be differences between OA types in the brain. Although people with hip and knee OA are often studied together 39,40 , there are differences between hip and knee OA 46 . For example, robust qualitative research (143 participants) suggests that people with hip OA often use more intense language to describe their pain compared to those with knee OA 47 . The affective and cognitive components of the pain sensation are processed in subregions of the medial prefrontal cortex, which may link to differences in pain experiences between hip and knee OA 48 . The differences in medial prefrontal cortex were informed by contrasts to healthy controls and due to limited number of studies available we were not able to conduct a direct comparison between hip and knee OA. Studies typically do not exclude participants if they have OA in joint beyond the joint of investigation. Hence, caution should be used interpreting these findings between potential differences in osteoarthritic joints and controls, as OA often affects more than one joint. Future studies should specifically study differences between the brain organisation of different OA sites.
Limitations and future directions. Our findings should be interpreted with caution considering some key limitations. First, we used a meta-analytic algorithm to integrate existing data and delineate consistent association across studies. However, this analytic approach can only include results from experiments that reach significance. Although, this limitation biases the meta-analysis toward finding significant results it adds confi- www.nature.com/scientificreports/ dence in our null finding from our primary ALE meta-analysis as we did not observe an association even when null experiments were included. Second, several factors such as sex 49 and medication may play a role in brain structure and function adaptation, specific to the insula 49 in people with chronic pain. However, the insufficient number of eligible experiments limited our ability to robustly assess the influence of these factors. The diverse inclusion criteria relating to medication used across experiments precludes subgroup analysis focused on medication. Third, most studies excluded participants with depression and anxiety. This may limit the generalisability of findings given the prevalence of depression and anxiety is approximately 20% in people with knee OA 50 , and evidence on the neural correlates of pain and depression 51 . Finally, limiting our focus to cross-sectional studies to better understand alterations associated with OA neglects understanding of longitudinal changes or changes in response to treatments. For example, longitudinal studies might provide insight into neuroplastic features associated that complement understanding of neuroplastic adaptations in OA beyond the brain 52 . Despite the difficulties associated with assessing pain, future research is encouraged to consider subgroups potentially based on pain characteristics. In fibromyalgia, Liu et al. 53 eloquently demonstrated the neuroplastic potential of the right anterior insular cortex when subgrouping patients by number of painful sites. More studies with sample sizes appropriately powered to detect potentially meaningful differences will reduce heterogeneity in estimates and increase confidence in the estimate ranges of possible differences for different measures of neurobiology associated with OA. This is a rapidly changing field of research, and inclusion of new experiments may change our findings. www.nature.com/scientificreports/

Conclusions
In summary, our pre-registered analysis did not find evidence of significant differences in OA neurobiology compared to healthy controls. However, findings from our exploratory quantitative analysis converge with our narrative synthesis to suggest that the right insula -namely interoceptive awareness and emotionally relevant context for sensory experience that contributes to OA pain may be implicated in knee and hip OA. Some limited evidence also potentially implicates the medial prefrontal cortex in hip OA. Despite the limitations associated with heterogeneity and study quality, these regions are potentially relevant to OA provide avenues for future research.

Data availability
All datasets generated and analysed during the current study, such as specific coordinates for the ALE analysis, are available in Supplementary Appendix 2.
Received: 20 February 2023; Accepted: 21 July 2023 Table 4. Study quality assessment according to the National Institute of Health Quality Assessment Tool. ✓ represents yes, × represents no,-represents could not determine, -represents not reported, --represents not applicable. Item 13 was not applicable for all studies.