Statistical shape modeling of the hip and the association with hip osteoarthritis: a systematic review.

Objective: To summarize available evidence on the association between hip shape as quanti ﬁ ed by statistical shape modeling (SSM) and the incidence or progression of hip osteoarthritis. Design: We conducted a systematic search of ﬁ ve electronic databases, based on a registered protocol (available: PROSPERO CRD42020145411). Articles presenting original data on the longitudinal relationship between radiographic hip shape (quanti ﬁ ed by SSM) and hip OA were eligible. Quantitative meta- analysis was precluded because of the use of different SSM models across studies. We used the Newcastle e Ottawa Scale (NOS) for risk of bias assessment. Results: Nine studies (6,483 hips analyzed with SSM) were included in this review. The SSM models used to describe hip shape ranged from 16 points on the femoral head to 85 points on the proximal femur and hemipelvis. Multiple hip shape features and combinations thereof were associated with incident or progressive hip OA. Shape variants that seemed to be consistently associated with hip OA across studies were acetabular dysplasia, cam morphology, and deviations in acetabular version (either excessive anteversion or retroversion). Conclusions: Various radiographic, SSM-de ﬁ ned hip shape features are associated with hip OA. Some hip shape features only seem to increase the risk for hip OA when combined together. The heterogeneity of the used SSM models across studies precludes the same SSM model and de ﬁ nition of hip OA are needed to allow for the comparison of outcomes across studies, and to validate the found associations.

s u m m a r y Objective: To summarize available evidence on the association between hip shape as quantified by statistical shape modeling (SSM) and the incidence or progression of hip osteoarthritis. Design: We conducted a systematic search of five electronic databases, based on a registered protocol (available: PROSPERO CRD42020145411). Articles presenting original data on the longitudinal relationship between radiographic hip shape (quantified by SSM) and hip OA were eligible. Quantitative metaanalysis was precluded because of the use of different SSM models across studies. We used the Newcastle eOttawa Scale (NOS) for risk of bias assessment. Results: Nine studies (6,483 hips analyzed with SSM) were included in this review. The SSM models used to describe hip shape ranged from 16 points on the femoral head to 85 points on the proximal femur and hemipelvis. Multiple hip shape features and combinations thereof were associated with incident or progressive hip OA. Shape variants that seemed to be consistently associated with hip OA across studies were acetabular dysplasia, cam morphology, and deviations in acetabular version (either excessive anteversion or retroversion). Conclusions: Various radiographic, SSM-defined hip shape features are associated with hip OA. Some hip shape features only seem to increase the risk for hip OA when combined together. The heterogeneity of the used SSM models across studies precludes the estimation of pooled effect sizes. Further studies using

Introduction
Hip osteoarthritis (OA) is one of the most common types of OA, and is a major contributor to the number of years lived with disability worldwide 1 . Hip shape has been recognized as an important risk factor for hip OA 2 . For this reason, the influence of hip shape has been increasingly studied over the last decade 3e9 . Hip shape variants that are known to significantly increase the risk for hip OA are acetabular dysplasia and cam morphology 2,7,10 . These hip shape variations are typically quantified by predefined radiological measurements such as the center-edge angle (CEA) and the alpha angle. However, other hip shape variants that are currently not captured by predefined radiological measurements may also play a role in the etiology of hip OA. The sole use of predefined measurements for hip shape analysis may therefore impede the discovery of further hip shape variants that increase the risk for hip OA.
This limitation has been partially circumvented by the emergence of statistical shape modeling (SSM) 11 as a novel shape analysis technique. SSM allows quantification of the whole shape of the hip and/or pelvis, in contrast to predefined measurements 12,13 . The application of SSM yields a set of shape variants, called shape modes, that are present in the studied population. When SSM is applied to radiographic images of the hip, the association between each hip shape mode and hip OA can be measured.
SSM has been increasingly used, and many different hip shape modes have so far been associated with hip OA. However, the interpretation of the SSM shape modes can be difficult and there is no thorough overview of the related literature yet. The purpose of this systematic review was to summarize which hip shape variants were found to be associated with incident or progressive hip OA, and to determine if there are any consistent patterns of similar shape variants to be recognized across different studies.

Protocol and registration
We reported this systematic review according to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) guidelines 14 . The review protocol was first submitted to PROSPERO on September 23, 2019, and was registered on April 28, 2020 (available from: www.crd.york.ac.uk/prospero/display_ record.php?ID¼CRD42020145411).

Eligibility criteria
All publications presenting original research on the association between hip shape and hip OA in human adults were considered eligible, as were conference abstracts published in 2016 or later. The inclusion criteria were: -Assessment of the longitudinal association between hip shape and OA had to be an aim of the study; -Hip shape had to be assessed with some form of SSM; -Hip OA should be either incident or progressive; -The definition of hip OA could be radiological, clinical, by total hip replacement (THR) status, or a combination of those; -Studies had to have control subjects that did not develop incident or progressive hip OA during the study.
The exclusion criteria were: -Hip shape was measured contralaterally to the hip that developed the outcome (e.g., the shape of the contralateral hip in case of THR); -The studied hip shape variant was explicitly described to be secondary to other conditions (e.g., childhood hip disease, trauma, avascular necrosis, tumors, previous hip surgery); -The primary outcome was biomechanical injury, or the validation of a novel diagnostic technique; -The OA outcome reflected 'early osteoarthritic changes', such as cartilage damage during arthroscopy or novel magnetic resonance imaging (MRI) techniques like delayed gadoliniumenhanced MRI of cartilage (dGEMRIC), Scoring Hip Osteoarthritis with MRI (SHOMRI), and T1r mapping.

Search and deduplication
An experienced information specialist (WB) searched the databases Embase (via Embase.com, since 1971), MEDLINE (Medline ALL via Ovid, since 1946), Web of Science Core Collection (since 1975) and the Cochrane Central Register of Trials (via Wiley, since 1992) from inception until April 25, 2020 (date last searched). A previously published method was used for search development and optimization 15 . The searches combine terms (both thesaurus terms where available, and terms in title and/or abstract) for hip osteoarthritis with terms for anatomy or morphology and terms for risk or pathology. Search results were limited to exclude (1) animal and child-only studies, (2) conference abstracts published before 2016, and (3) publications in other languages than English. The full search strategy can be found in Supplement 1. Additionally, we searched Google Scholar and screened the reference lists of the included references for any other relevant articles. The search results from all databases were imported in EndNote and deduplicated 16 .

Study selection
Two reviewers (MvB and RA) independently screened the titles and abstracts of all search results, and after having compared the included references, independently reviewed the full text of all potentially eligible studies. This process was done in EndNote with a predefined method 17 . Subsequently the reviewers held a consensus meeting to discuss each full-text article separately, and to select the final studies to be included. A third reviewer (MN) was consulted to resolve any disagreements.

Data collection/extraction
A custom open-ended electronic data extraction form was developed and pilot-tested with a sample of the included studies. The used data extraction form, including the full list of extracted variables, can be found in Supplement 2. Data extraction was independently performed in duplicate by two reviewers (MvB and RA), and the results were compared in a consensus meeting. For one conference abstract of which the full text was not published yet, the reviewers requested and received the full text manuscript from the authors.

Risk of bias assessment
We used the NewcastleeOttawa Scale (NOS) to assess the risk of bias of the individual studies 18 . We used either the cohort version or the caseecontrol version as appropriate. The questions and the scoring key can be found in Supplement 3. The two reviewers (MvB and RA) independently appraised the quality of the individual studies, and disagreements were resolved in a consensus meeting. Publication bias was reduced by searching for recent conference abstracts and by searching Google Scholar for gray literature.

Statistical shape analysis
The application of SSM requires all images (e.g., radiographs) to be annotated by placing a set of points around the outline of the bone. To negate the effect of size and orientation, the outline of the bone (the shape) across images is usually aligned first using a technique called Procrustes analysis. Principal component analysis (PCA) is then applied to identify the main variations in shape (called shape modes) within the given population (i.e., across all images), summarized as a statistical shape model. Shape modes are stored as a set of continuous variables, usually standardized to have a mean of 0 and a standard deviation of 1, and are linearly independent of each other. These shape modes represent the apparent radiographic shape, and may not always match the true anatomic shape due to the influence of subject positioning and radiographic projection effects. Shape modes are ordered by their contribution to the total shape variance, the lower mode numbers being the most contributing. Because the SSM process arbitrarily assigns deviations from the mean shape as either positive or negative, a certain shape variant can either be positively or negatively (inversely) associated with the outcome. Furthermore, due to the nature of PCA the definition of individual shape modes will be data dependent and thus will vary across datasets/studies.

Data synthesis
The main outcome measures that we extracted were the measures of association for the relationship between SSMdefined hip shape and OA. These could be odds ratios (OR), relative risk (RR), prevalence ratios (PR), or any other association measures. If present, the covariate-adjusted measures were extracted. We only performed qualitative data synthesis, as the use of SSM models resulting from different studies precludes statistical pooling and thus meta-analysis. To still be able to summarize associations, we qualitatively compared the descriptions (as provided in the original papers) of the different hip shape modes from across studies. The reported shape descriptions are therefore either the literal descriptions by the original authors, or the reviewers' interpretation of the original figures if these were unambiguous. If neither was the case, we did not report a shape description.

Study selection
The initial database searches yielded 4,618 unique references, which were screened by title and abstract. Twenty-five of these had used SSM to quantify hip shape and were retrieved for full-text reading. The screening and inclusion process as well as the reasons for exclusion are shown in Fig. 1. Finally, we included nine articles in this review 19e27 .

Study characteristics
The main characteristics of the nine included studies (published between 2007 and 2017) are presented in Table I. The study by Mezhov et al. 27 has only been published as a conference abstract as of yet, but we received the full manuscript from the authors upon request. The included studies present data on a total population of 4,706 subjects, with 6,483 hips analyzed with SSM. Not all subjects were unique, since some parts of study populations were used in two separate articles 20,23,25,27 . The Rotterdam Study population was also used twice, but random samples were drawn, making duplicate entry of subjects unlikely 19,24 . Factoring in the use of data from these study populations in separate articles, the number of unique hips analyzed with SSM was 4,584. Median sample size was 664 subjects (range 110e831) and median follow-up period was 6.5 years (range 5e19). The overall proportion of females was 69.0%, ranging from 51% 23,27 to 100% 20,26 . The mean age of included subjects ranged from 53.6 20 to 70.7 26 , with a pooled mean age of 61.8 years across all studies.

Risk of bias
A summary of the risk of bias assessment is presented in Table II, whereas an extensive overview can be found in Supplement 2. Eight of the included studies were deemed as having good methodological quality, with a low risk of bias 19e22,24e27 . When strictly following the NOS guidelines, one study scored poorly because of self-reported THR assessment and the lost to follow-up rate 23 . However, the reviewers considered the overall quality of this study sufficient to regard the findings as reliable.

Assessment of exposure and outcome
An overview of the assessment of exposure and outcome in each study can be found in Table III. Seven studies 19e22,24e26 used pelvic radiographs to assess hip shape, whereas the other two 23,27 used Dual-energy X-ray absorptiometry (DXA). The SSM points used to outline the hip shape varied from 16 19 to 85 23,27 . Three studies only described the femoral head 19 or part of the proximal femur 21,26 . Three studies additionally included the acetabular roof 22,23,27 . The remaining three studies also included the ipsilateral lower pelvis, consisting of the acetabulum, the pelvic teardrop, and the pubic and ischial bones 20,24,25 . All studies 19e27 used the ASM toolkit (University of Manchester, Manchester, UK) to annotate the images. Seven studies also used this toolkit to create the SSM, while two studies 23  Osteoarthritis andCartilage grade 2 26 , and meeting the American College of Rheumatology (ACR) criteria 25 . Some studies included multiple definitions of hip OA, either creating subgroups per outcome definition 19,25 or pooling multiple definitions into one group 24,26 . Six studies 19e21,24e26 used incident hip OA as the outcome, meaning all cases had baseline OA scores (e.g., KL, Croft) of 0e1. In the remaining three studies 22,23,27 , the distinction between incidence and progression could not be made because part of the study sample already had OA scores 2 at baseline. All studies corrected for two or more covariates in their analyses 19e26 .

The association between hip shape and THR
The results from the studies that used THR as a separate outcome definition 19,20,22,23,25,27 are summarized in Table IV, whereas the complete results (including non-significant associations) can be found in Supplement 2. All six studies that used THR as a separate outcome measure found at least one shape mode that was statistically significantly associated with THR (median 2 modes, range 1e6) at the chosen alpha level. The indication for THR was incident hip OA in three studies 19,20,25 , and incident or progressive hip OA in the other three studies 22,23,27 . One study 25 used Bonferroni correction for multiple testing.
A total of 18 hip shape modes were associated with future THR across the different studies. One of these modes (describing a flattened headeneck junction, a flat major trochanter and a prominent acetabular posterior wall) showed a consistent association in two different populations, namely the CHECK and Chingford populations 20 . Five studies 19,20,23,25,27 (out of the six that used THR as a separate outcome measure) found at least one shape mode consistent with cam morphology; and four 20,22,23,27 out of six studies found a mode representing acetabular dysplasia. A hip shape variant possibly representing pincer morphology was associated with THR in one study 25 out of the six studies that included the acetabular roof in their model 20,22e25,27 . The description of this shape mode was "more pronounced lateral acetabular rim" in this study. Deviations in acetabular version were associated with THR in both studies that included the acetabulum in their shape model 20,25 . One study describes a shape mode with "a prominent acetabular posterior wall", possibly representing excessive acetabular anteversion, combined with "a flattened headeneck junction and a flat major trochanter" 20 . The other study describes a mode with "acetabular retroversion", combined with a "flat headeneck junction and broad femoral neck" 25 .
The association between hip shape and radiographic hip OA Studies that mainly used radiographic hip OA as outcome measure 19,21,24,26 are summarized in Table V, whereas the complete results (including non-significant associations) can be found in Supplement 2. At least one shape mode per study was statistically significantly associated with hip OA (median 3 modes, range 1e6) at the chosen alpha level. In all four studies the outcome was incident hip OA (baseline OA scores of 0e1). Two studies 24,26 used a combined definition of hip OA, where THR and radiographic hip OA were pooled into a single endpoint. However, one of those studies only seemed to present radiographic hip OA cases in their results, and no THRs 24 . Two studies 24,26 used Bonferroni correction for multiple testing.
Thirteen hip shape modes were associated with incident radiographic hip OA. One study 21 presented two hip shape modes that showed different associations in different subgroups. In this study, mode 2 (representing alterations in the transition between greater trochanter and femoral neck and femoral neck length and thickness) was inversely associated with symptomatic radiographic hip OA in the entire study population, but positively associated with radiographic hip OA in males only. Positive mode 2 scores represented flattening of the femoral head, suggestive of cam morphology. Two 19,21 out of four studies found shape modes representing cam morphology; and the only study that included the acetabulum in their model 24 found shape modes representing dysplasia. Acetabular version was also associated with radiographic hip OA in that study, but the type (ante-or retroversion) was unspecified 24 .
The association between hip shape and clinical hip OA One study 25 used a clinical definition of hip OA, namely the ACR criteria, next to another definition (THR). They found no statistically significant associations between baseline hip shape modes and ACR criteria at follow-up. Another study 21 made the distinction between symptomatic radiographic hip OA and overall radiographic hip OA. This study found associations between different shape modes and symptomatic radiographic hip OA in the overall     *These shapes are positively associated with the outcome, unless stated otherwise. For a visual impression of what these shape modes look like, we refer to the original articles. Effect sizes are shown per 1 SD increase in shape mode value. An effect size ratio between 0 and 1 indicates that the negative SDs are associated with the outcome, and ratios above 1 indicate that positive SDs are associated with the outcome. Descriptions in regular typeface are taken literally from the original papers, while descriptions in italics are interpreted from the figures of the original papers; **Exact P-values were not given, but were under the alpha level of 0.05; a This study described a combined outcome definition (THR or KL 2) in their methods, but only presented KL 2 cases in their results; b This study used a combined outcome definition (THR or Croft 2); yThis study did not describe what the actual differences between positive and negative SDs were; zIn the group with baseline symptoms a decrease in mode 6 score was associated with the outcome, while in the group without baseline symptoms an increase in mode 6 score was associated with the outcome; xOR for symptomatic radiographic hip osteoarthritis; CI: confidence interval; OR: odds ratio; KL: KellgreneLawrence grade; BMI: body mass index; BMD: bone mineral density; SD: standard deviation.

Table V
Hip shape modes significantly associated with radiographic hip osteoarthritis outcomes Osteoarthritis andCartilage population, as well as in subgroups with or without baseline symptoms (Table V).

Discussion
In this systematic review we have summarized all available evidence from the published literature on the association between SSM-defined apparent radiographic hip shape and hip OA. Our results show that every published study on this topic that was included in this review found at least one hip shape mode statistically significantly associated with incident or progressive hip OA or future THR. Most studies found multiple (up to six) linearly independent hip shape modes associated with hip OA. Most of the included studies used different populations and different SSM point positions for their modeling, which complicates the comparison of hip shape modes between studies. However, in the following we attempt to discuss the overall patterns in radiographic hip shape that were found to be associated with hip OA.
Shape variants that likely represent cam morphology and acetabular dysplasia were consistently found to be associated with future THR and/or incidence or progression of radiographic hip OA. Shape modes that might represent cam morphology were described as "cam-type change of femoroacetabular impingement" 21 , "pistol-grip deformity" 23 , "less concavity of superior headeneck junction" 25 , "less pronounced curve from upper femoral neck into the head" 19 , "less head-neck offset" 25 , "nonspherical femoral head" 20,23,27 , "flattening of the head-neck transition" 21 , and "flattening of the femoral head" 21 . Modes that may represent acetabular dysplasia were described as "less/poor/ decreasing acetabular coverage" 22e24,27 , and "a shallow acetabulum" 20 . The associations between hip OA and both cam morphology and acetabular dysplasia have already been proven in other studies that used traditional measurements, such as the alpha angle and the CEA 7,8,10,28e35 . Two cross-sectional studies that used SSM also found associations between cam morphology and the presence of hip OA 36,37 . These studies were not included in our systematic review due to their cross-sectional design. Because there were no baseline OA measurements, it remains unclear whether the shape modes found in these studies preceded hip OA or resulted from it.
A shape mode possibly representing pincer morphology was also associated with THR in one of the studies included in this review 25 . Other studies, using traditional measurements such as the CEA and the crossover sign, did not find a positive association between pincer morphology and hip OA so far 7,8,10 . Maybe the risk for hip OA is only increased when a pincer morphology is mixed with other shape features, or for certain subtypes of pincer morphology that are not captured with traditional measurements. A crosssectional study 36 (excluded from our systematic review) also found an association between pincer morphology and hip OA. In the shape mode of that particular study, the "pincer-type variation" was combined with a "larger femoral head and wider femoral neck". This combination could theoretically aggravate femoroacetabular impingement. However, since no baseline OA measurements were done in this study, the "pincer-type variation" shape mode could have also represented an osteophyte of the lateral acetabulum, secondary to hip OA.
Multiple studies included in this systematic review found associations between acetabular version and hip OA. This is in line with studies using traditional measurements, which have also suggested that both acetabular anteversion and retroversion could be associated with hip OA 38e40 . A cross-sectional study that used SSM to define hip shape also found associations between two shape modes, possibly representing acetabular retroversion and anteversion respectively, and the presence of hip OA 41 .
Because one statistical shape mode often consists of more than one shape feature, extra caution has to be taken when singling out just one shape feature. The association with hip OA may only be present when there is a combination of multiple shape features. This is precisely the advantage of SSM. One combination that consistently appears to be associated with hip OA is cam morphology combined with dysplastic acetabular features 20,25 , a combination that has been previously described in the literature 42,43 . It is still not entirely clear why this combination would increase the risk for hip OA, because theoretically a cam would be less likely to impinge with a dysplastic acetabulum. However, one computer simulation study has demonstrated that impingement can still occur, but more proximally and more medially than with a normal acetabulum 44 . It remains unknown whether the higher risk is due to the cam morphology alone, the dysplastic acetabulum, or the interaction between the two. Another reported shape combination was the presence of a cam morphology with acetabular retroversion 25 , which could be theoretically explained by femoroacetabular impingement happening earlier during hip flexion and internal rotation. The combination of a valgus hip with acetabular dysplasia 22,23 was associated with hip OA in two studies. From a biomechanical perspective, this could be explained by higher vertical joint reaction force 45 acting on a smaller surface during weight bearing. This combination has also been previously described 43 . Besides the aforementioned combinations, variations in the size of the trochanters, the length and width of the femoral neck, and the apparent rotation of the femur and pelvis were found, but no obvious patterns were seen in these variations.
The magnitude of the reported associations between hip shape modes and hip OA varied greatly between studies. Due to the different SSM point positions and different outcome definitions, the association measures are not directly comparable. Large ORs or RRs can be interpreted as a strong association nevertheless.

Strengths and limitations
This is the first systematic review on the association between SSM-defined radiographic hip shape and hip OA. It offers an overview of the patterns of hip shape features that are associated with hip OA in multiple populations. The interpretation and implications of the results were carefully discussed within the review group, which contains experts in the fields of both hip OA and SSM. Strengths of the included studies are the relatively large sample sizes and the various populations of differing ages and ethnicities that were included. Overall, the included studies scored well on methodological quality.
One limitation of this review is that we were not able to conduct a meta-analysis. This is inherent to SSM, because the shape modes will be defined by the population from which they were created. This was already taken into account when designing the review protocol. The lack of a meta-analysis makes validation of associations difficult. We therefore subjectively described patterns of hip shape that seemed to be consistently associated with hip OA across the included studies. A second limitation is that the interpretation and description of the shape modes are relatively subjective processes, which were left to the authors of the included papers. Still, we purposefully reported only the literal descriptions from the original articles to reduce bias by our own interpretation. Another limitation is that none of the included studies have validated the found associations in an independent test dataset. Internal validation would have been possible if the datasets had been divided into a training set and a test set. This is something that future SSM studies could possibly address. One more consideration is the influence that hip OA may have on hip shape. As some studies have shown, hip OA may not only result from certain hip shape variants but can also cause changes in hip shape 46 . This is not a problem in incidence studies 19e21,24e26 where all analyzed hips were free of OA at baseline, but the hip shape modes found in progression studies 22,23,27 could already be a result of early hip OA. Further limitations of the included studies are the heterogeneity of pelvic radiograph protocols and outcome definitions, and the varying use of covariate adjustment. Further research is required to investigate whether significant covariates (e.g., gender) may require independent shape models instead of simply adjusting for them. Lastly, most studies only described shape modes that were significantly associated with hip OA at their chosen alpha level, but some studies used Bonferroni correction, whereas others did not. This may have led to some reporting bias, even more so because statistical significance does not always translate to clinical significance. In our opinion, the use of multiple testing correction in SSM analysis should depend on the goal of the analysis. When SSM is used for hypothesis generation, you could argue not using a correction because you would want to find any possible leads. The associations found in this way should not be taken as evidence though, but have to be investigated further. In other cases, a method like the Bonferroni correction is warranted. In any case, authors should preferably explain their reasoning for (not) using multiple testing correction.

Conclusion
This systematic review suggests that several radiographic hip shape features and combinations thereof are associated with the incidence or progression of radiographic hip OA and with future THR. Associations of both cam morphology and acetabular dysplasia with hip OA have been found by SSM in multiple studies. In addition, hip shape features other than these well-known variants also appear to be associated with hip OA. Moreover, certain combinations of (sometimes subtle) hip shape features, rather than single features, may increase the risk for development or progression of hip OA when present together. More research with SSM is needed to validate these associations, and a standardized set of SSM point positions should be used to allow comparison between studies. When SSM is used to generate hypotheses, the found associations could be tested with traditional radiographic measurements in an independent sample. This would both validate the associations and make them more easily transferrable to clinical practice. Outside the submitted work: NA reports personal fees from Pfizer and Eli Lilly and Company, with research grants from Merck. SBZ reports personal fees from Pfizer and Osteoarthritis and Cartilage, with research grants from the European Union, The Netherlands Organisation for Health Research and Development, the Dutch Arthritis Society, and Foreum. GJ reports personal fees from BMS, Roche, Abbvie, Amgen, Eli Lilly and Company, Novartis, Janssen, with research grants from Covance. AN reports personal fees from GSK, Flexion Therapeutics, MedScape and Health Press Ltd. HW reports research grants from the European Union, the Dutch Arthritis Society, and the Dutch Government. RA reports a research grant from the Dutch Arthritis Society.

Author contributions
The remaining authors report no competing interests.

Role of the funding source
The funding sources had no role in study design, in collection, analysis, and interpretation of the data, nor in the preparation of the manuscript or the decision to submit the manuscript for publication.