Abstract
Objectives Prompt diagnosis of giant cell arteritis (GCA) is important to avert visual loss. False-negative temporal artery biopsy (TAB) can occur. Without vascular imaging, GCA may be overdiagnosed in TAB-negative cases, but it is unclear how often this occurs. An unbiased test is a way to address an imperfect reference standard. We used the known Human Leukocyte Antigen (HLA) region genetic association of TAB-positive GCA to estimate the extent of overdiagnosis before widespread adoption of temporal artery ultrasound as a first-line test.
Methods Patients diagnosed with GCA between 1990-2014 consented to the UKGCA Consortium study. HLA region variants were jointly imputed from genome-wide genotypic data of cases and controls. Per-allele frequencies across all variants with p<1.0×10−5 were compared with population control data to estimate overdiagnosis rates in cases without a positive TAB.
Results Genetic data from 663 patients diagnosed with GCA were compared with data from 2619 population controls. TAB-negative GCA (n=147) and GCA without a TAB result (n=160) had variant frequencies intermediate between those of TAB-positive GCA and population controls. Making several strong assumptions, we estimated that around two-thirds of TAB-negative cases and around one-third of cases without TAB result may have been overdiagnosed. From these data, TAB sensitivity is estimated at around 88%.
Conclusions Conservatively assuming 95% specificity, TAB has a negative likelihood ratio of around 0.12. Genotyping alone cannot diagnose GCA at the individual level. Group-level HLA variant genotyping might be used to compare the overall accuracy of different diagnostic pathways or different classification criteria sets.
Key messages
Under certain conditions and assumptions, overdiagnosis can be estimated using genetic data.
The specificity of temporal artery biopsy was estimated as about 88%.
Without vascular imaging, giant cell arteritis may often be overdiagnosed in biopsy-negative patients.
Introduction
Giant cell arteritis (GCA) is a vasculitis of older people(1) which can present with a variety of symptoms including headache, scalp tenderness, jaw ache, ocular ischaemia, fever and weight loss(2). Where GCA is strongly suspected, it should be treated with high-dose glucocorticoids pending further investigation. The “diagnostic momentum” thus created can make it difficult to reverse the diagnosis if the temporal artery biopsy (TAB) subsequently proves negative and yet no better explanation for the presenting symptoms is found. Some patients with genuine GCA do have a negative TAB(3), often attributed to patchy arterial involvement or “skip lesions”(4, 5). Sensitivity of a unilateral TAB was estimated from bilateral TAB studies as 87%(6). Sometimes there is no evidence of any cranial artery involvement in GCA(7); this extracranial or “large-vessel” GCA subset is usually diagnosed with vascular imaging (8, 9). Vascular imaging is now also recommended as first-line investigation for cranial GCA, but where imaging and TAB give discordant or unexpected results, expert clinical judgement of GCA probability is crucial(10).
Where confirmatory tests are equivocal or negative, but pre-test probability is high, a clinical diagnosis of GCA may be justified by the desire to avoid preventable sight loss(11). The cost of this strategy is that some patients will be overdiagnosed. Adoption of better tests including vascular imaging should improve overall diagnostic accuracy of the pathway, but so far no method has been established that can demonstrate reduction in overdiagnosis.
We previously demonstrated that TAB-positive GCA is strongly associated with genetic variants in the Human Leucocyte Antigen (HLA) region including the single nucleotide polymorphisms rs9268905 (p=1.9×10−54, per-allele odds ratio(OR)=1.79) located between HLA-DRA1 and HLA-DRB1, rs477515 (p=4.0×10−40, OR=1.73) located in the intergenic region between HLA-DRB1 and HLA-DQA1, and the HLA-DRB1*04 allele (p= 6.8×10−38, OR=1.92) (12). HLA region genotyping is not currently performed in routine care for GCA diagnosis; therefore genotype cannot directly influence diagnostic decisions and can act as an “unbiased umpire” (13). In this study, we sought to estimate the extent of GCA overdiagnosis during the era before widespread adoption of vascular imaging as a first-line test in GCA diagnostic pathways.
Methods
Study population
The UK GCA Consortium was designed as a genetic epidemiology study. Ethical approval was received from the Yorkshire and Humber–Leeds West Research Ethics Committee (REC Ref. 05/Q1108/28). Patients diagnosed with GCA by a consultant rheumatologist or ophthalmologist were identified from clinic lists and searches of TAB records from histopathology databases. Participants from 31 centres provided written informed consent. Clinical data and TAB result were determined by casenote review. TAB result was determined from the clinical records as “positive”, “negative” or “no result”. The “no result” category included cases where no biopsy was done, no temporal artery biopsy was done (no arterial tissue received for processing), or where no clear TAB result was determinable despite complete review of the patient’s medical records.
Population control data came from two sources: 1,430 individuals from the 1958 British Birth Cohort (58BC), who were born in England, Wales and Scotland during one week in 1958; and 1,500 healthy blood donors from the United Kingdom Blood Service for the Wellcome Trust Case Control Consortium (WTCCC)(14, 15) (Supplementary Methods).
The genetic data from the TAB-positive GCA cases have been previously included in two international collaborative studies(12, 16). In this study, we focus on the cases that received a “clinical diagnosis” of GCA: the TAB-negative and “no TAB result” cases that still received a GCA diagnosis.
Genotyping and quality control
Genomic DNA was extracted from blood samples by standard methods(12). Cases and controls were genotyped separately with the use of two versions of the Illumina chip (Infinium Human core 24 Beadchip, HumanCore-12v1-0_A). The same quality filters were independently applied to the raw data from each cohort using PLINK version 1.9 (Supplementary Methods). The genome-wide Manhattan plot for GCA susceptibility is dominated by the signal from the HLA region in chromosome 6 (chr6: 29–34 Mb on build 36/hg18)(17) and therefore we chose to focus on the HLA region in our analysis. In this region, there is strong linkage disequilibrium. The allele type for each HLA gene can be inferred from SNP genotyping results by imputation.
HLA imputation
Classical HLA alleles were determined from the single nucleotide polymorphism (SNP) data by imputation. Genotypic data from both cases and controls were jointly imputed across the extended major histocompatibility complex (MHC) using the SNP2HLA imputation method, using the Beagle software package (18, 19) (Supplementary Methods). Post-quality control filter thresholds were set to remove variants with a minor allele frequency (MAF) <0.01 and variants with an information score <0.8. After this was done, 159 classical HLA alleles and 6730 SNPs remained. Variants from the association analysis of TAB positive cases versus controls spanning the entire HLA region with p < 1.0 × 10−5 were used in order to estimate the misclassification rate.
Association analysis using different case definitions
Firstly, to determine if the TAB result was associated with specific HLA alleles imputed across extended haplotypes, we compared the strength of association (odds ratio) and allele frequencies of susceptibility and protective HLA alleles in TAB-positive GCA, TAB-negative GCA and GCA with no TAB result.
Estimation of misclassification rate
Because of the extensive linkage disequilibrium within the HLA region, different genetic variants within this region are not independent of one another, and so cannot be treated as independent variables. On the other hand, selection of only a few, very strong HLA associations of TAB-positive GCA would have potentially introduced “winner’s curse” bias (20). To address this, for each TAB-defined subset of the GCA cases, the proportion of patients misclassified as GCA was estimated by taking the average effect size across all the variants with p < 1.0 × 10−5 (per-allele association of positive-TAB GCA cases with controls). This conservative approach acknowledges the non-independence of alleles and SNPs from each other, while avoiding loss of potentially informative data. The proportions of “genuine” GCA cases in the TAB-negative GCA and GCA without TAB groups, pn and px, were calculated from the allele frequencies in TAB-positive cases and controls (Supplementary Methods). Forsimplicity, we assumed an additive gene-dose model in our analyses: although HLA-DRB1*04 may have a dominant effect (21) we could not assume this was also true of other HLA variants.
Assumptions made
For the purposes of analysis we assumed (1) a 100% specificity of TAB (no false positive TAB), (2) the group of GCA cases without positive TAB comprises a mixture of “genuine” GCA and “overdiagnosed” GCA, (3) all “genuine” GCA cases share the same HLA associations, (4) “overdiagnosed” cases have a similar distribution of HLA variants as the control population, (5) HLA status did not influence clinical diagnosis.
Results
After quality control, we had genetic data for 663 cases of GCA and 2619 population controls. The GCA patients were diagnosed between 1991 and 2014. Only 65 GCA patients had imaging evidence of GCA: during this period, TAB was the first-line confirmatory test, and vascular imaging tests for GCA were not routinely performed in most UK hospitals. Table 1 gives clinical data on the TAB-positive GCA (n=356), TAB-negative GCA (n=147) and cases of GCA without any TAB result (n=160).
Clinical features of GCA patients included in the analysis
HLA allele associations
The HLA association of TAB-positive GCA patients was similar to that described in previous studies, including the large international study to which we had contributed TAB-positive cases(12). HLA-DQA1*01, HLA-DQB1*05 and HLA-DQB1*06 alleles had a protective effect; HLA-DQA1*03, HLA-DQB1*03 and HLA-DRB1*04 were associated with GCA susceptibility (Table 2, “TAB positive” column).
Significant association of classical HLA alleles selected from the TAB positive cohort, at two- and four-digit resolution, and then tested in the subsets of negative and no TAB result at 0.00001 level of significance without conditioning
TAB-negative GCA showed a pattern of HLA allelic associations in the same direction as TAB-positive GCA, but the effect was generally diluted: for each HLA allele, the strength of association was diminished (closer to the null odds ratio of 1.0). The odds ratios for GCA with no biopsy result were generally intermediate between those for TAB-positive and TAB-negative GCA (Table 2: columns for TAB negative, no TAB result).
Single-nucleotide polymorphisms
Associations with the three SNPs (rs9268969, rs9275184, rs477515) previously reported as being associated with biopsy-confirmed GCA (12, 16) were markedly diluted for TAB-negative GCA, to the extent that only rs9275184 (located between? HLA-DQA1 and HLA-DQA2) retained statistical significance at p<0.05 (Table 3, “negative TAB” column). Again, odds ratios for GCA with no biopsy result were intermediate between those observed for TAB-positive GCA and TAB-negative GCA (Table 3, “no TAB result” column).
Estimation of misclassification rates
470 HLA variants were associated with TAB-positive GCA compared to controls at a per-allele threshold of p<1.0 × 10−5. We estimated the proportion pn of “genuine” cases among TAB-negative GCA as 0.33 (SD:0.23) and proportion px of “genuine” cases in GCA without TAB result as 0.67 (SD:0.15).
Discussion
In this study, we used HLA genotyping data to estimate the proportion of clinically-diagnosed GCA patients who are overdiagnosed. Of 503 cases in our study who had a TAB, 356 (71%) were recorded as positive; this is in line with the 77% sensitivity of TAB reported in a meta-analysis(22). We found, by demonstrating dilution of the known genetic associations with biopsy-positive GCA, that subject to strong assumptions, about two-thirds of biopsy-negative GCA cases, and one-third of GCA cases without a biopsy result, may have been overdiagnosed.
The Merriam-Webster dictionary defines overdiagnosis as “the diagnosis of a condition or disease more often than it is actually present”. The problem here is that even if it is shown at the group level that overdiagnosis is occurring, overdiagnosis/overtreatment of an individual patient remains a rational response to diagnostic uncertainty, if the potential consequences of missing a case of genuine GCA include significant adverse outcomes such as visual loss. Under such circumstances, the greater the diagnostic uncertainty, the more overdiagnosis is likely to occur.
Our estimate depends on several assumptions which are limitations of our approach. (i) We assumed a 100% specificity of TAB (no false positive TAB). This assumption has been called into question by a reliability study in which discordant classifications of TAB images as positive/negative were made by a panel of pathologists(23). Standardised reporting templates for TAB might address this. (ii) We assumed that the group of GCA cases without positive TAB comprises a mixture of “genuine” GCA and “overdiagnosed” GCA. This could be an oversimplification if there are “halfway-house” disease states including potentially PMR.(9) (iii) We assumed that all “genuine” GCA cases share the same HLA associations, but there may be heterogeneity within this disease. Large-vessel involvement was previously suggested in small studies to have a different pattern of HLA association (7), but a more recent study had failed to confirm this(24). Takayasu arteritis (TAK) does appear to have a different HLA association compared to GCA (17, 25), but the cases in our study were all of an age-range compatible with GCA rather than TAK. (iv) We assumed “overdiagnosed” cases have a similar distribution of HLA variants to population controls. Although there is global variation of population HLA frequencies (21), our cases and controls both came from the UK. (v) We assumed that HLA genotype did not influence clinical diagnosis. This assumptions seems reasonable, since the clinical diagnosis of GCA was an inclusion criterion for UK GCA Consortium, and HLA typing is not an accepted test for GCA diagnosis in clinical practice. The HLA genotyping done for this study was only done in retrospect and the results of the HLA genotyping were not returned to the clinical care team.
Overdiagnosis can be a system-level phenomenon rather than necessarily implying error on the diagnostician’s part(26). Given an estimated GCA prevalence of 52 per 100,000 in the over-50s(1), with just over 25 million over-50s living in the UK (2021 data (27)), over 13,000 people in the UK have ever been diagnosed with GCA. Extrapolating from our findings reported here, up to 3,500 of these people may have been overdiagnosed and therefore been exposed to the harms of long-term glucocorticoid therapy. However, diagnostic pathways have rapidly evolved since the 2007-2014 era.
Although 71% of GCA patients in our cohort had a positive TAB, removing “overdiagnosed” cases from the denominator would give a TAB sensitivity for detecting “genuine GCA” of around 88%, comparable to the 87% sensitivity reported by studies of bilateral TAB(6). For the diagnostician, negative likelihood ratio (ratio of pre-test odds to post-test odds, given a negative test result) is of more practical use. Taking this 88% sensitivity and conservatively assuming a TAB specificity of 95%, the negative likelihood ratio associated with a negative TAB would be 0.12 (revised downward from the previously-reported estimate of 0.23(28)). To give an illustrative example, pre-test GCA probability of 20%, 50% or 80% would be downgraded by a negative TAB result to a post-test probability of 3%, 11% or 34% respectively.
Could HLA genotyping be useful in clinical practice to improve accuracy of estimation of pre-test probability? HLA-DRB1*04, the strongest allelic association at the group level, is fairly common in northern European populations and having this allele would produce a comparatively small diagnostic shift at the individual patient level: the positive and negative likelihood ratios of HLA-DRB1*04 in a northern European population are approximately 1.79 and 0.76 respectively (24); this could be combined with likelihood ratios for symptoms, signs and laboratory markers of inflammation to incrementally improve diagnostic accuracy(2). It is conceivable that HLA genetic data might be incorporated into a future multiparameter clinical decision aid, most likely, including just one or a few HLA variants into a model would be more feasible than analysing all 470 HLA variants we studied here. Because of strong linkage disequilibrium and multiplicity of testing, it is not possible to identify the optimal variant(s) from our data.
Our results provide a benchmark for estimating GCA overdiagnosis rates during the era before widespread adoption of temporal artery ultrasound. Future studies of HLA frequencies in GCA cohorts diagnosed via pathways involving vascular imaging could ascertain whether diagnostic accuracy has improved with contemporary diagnostic pathways. Within cohorts diagnosed during the same time period, the accuracy of different classification criteria for GCA could also be compared; HLA variant/allele frequencies in the subgroup classified as GCA by the 1990 ACR criteria, might be compared to HLA variant/allele frequencies in the subgroup classified as GCA by the 2022 ACR/EULAR classification criteria. Unbiased tests have a useful role to play in such comparisons(13).
Lastly, our approach might be extended to evaluate diagnostic pathways or classification criteria for other diseases where the gold-standard test is highly specific but insensitive, and where a strong genetic association is present.
Data Availability
The data used to support the findings of this study are included within the article.
Funding statement
Chatzigeorgiou: PhD was supported by a Emma and Leslie Reid Scholarship from the University of Leeds
Barrett: received salary support from the National Institute for Health Research (NIHR) Leeds Biomedical Research Centre (BRC)
Morgan: received salary support from the Medical Research Council (MRC) TARGET Partnership Grant, MR/N011775/1, NIHR Leeds BRC, NIHR Leeds Medtech and in vitro Diagnostics Co-operative (MIC) and NIHR Senior Investigator Award
Mackie: received salary support from NIHR Clinician Scientist Fellowship NIHR-CS-012-016 and NIHR Leeds BRC.
The UKGCA Consortium study received funding from the NIHR Leeds BRC, MRC TARGET Partnership Grant, MR/N011775/1, Academy of Medical Sciences/Wellcome Trust (AMS-SGCL4-Mackie), Mason Medical Research Foundation and Leeds Teaching Hospitals Charitable Trustees.
This study was supported in part by the NIHR Leeds BRC and the NIHR Leeds MIC. The views expressed are those of the authors and not necessarily those of the NIHR or the Department of Health and Social Care.
Author Contributions
Design, concept: CC, AWM, SLM, JHB
Analysis: CC, JHB, AWM, SLM
Clinical expertise: AWM, SLM
Writing manuscript: CC, AWM, SLM, JHB, JM
Manuscript review: CC, AWM, SLM, JHB, JM
Conflicts of interest
Chatzigeorgiou-None
Barrett – None
Martin -None
Morgan reports: consultancy fees payable to her institution from Roche/Chugai, Sanofi/Regeneron, Glazo Smith Kline and AstraZeneca, outside the submitted work. Reports research and/or educational funding were received from Roche/Chugai and Kiniksa Pharmaceuticals, outside the submitted work.
Mackie reports: Consultancy on behalf of her institution for Roche/Chugai, Sanofi, AbbVie, AstraZeneca; Investigator on clinical trials for Sanofi, GSK, Sparrow; speaking/lecturing on behalf of her institution for Roche/Chugai, Vifor and Pfizer; patron of the charity PMRGCAuk. No personal remuneration was received for any of the above activities. Support from Roche/Chugai to attend EULAR2019 in person and from Pfizer to attend ACR Convergence 2021 virtually.
Data availabilite statement
The data used to support the findings of this study are included within the article.
Acknowledgements
We would like to thank UK GCA participants, without whom this study would not have been possible.
References
