Contextualizing temporal arterial magnetic resonance angiography in the diagnosis of giant cell arteritis: a retrospective cohort study

Abstract

Objectives

Imaging modalities have become common in evaluating patients for a possible diagnosis of GCA. This study seeks to contextualize how temporal arterial magnetic resonance angiography (TA-MRA) can be used in facilitating the diagnosis of GCA.

Methods

A retrospective cohort study was performed on patients who had been previously referred to a rheumatologist for evaluation of possible GCA in Hamilton, Ontario, Canada. Data including clinical features, inflammatory markers, imaging, and biopsy results were extracted. Multivariable logistic regression model to predict the diagnosis of GCA. Using these models, the utility of TA-MRA in series with or in parallel to clinical evaluation was demonstrated across the cohort as well as in subgroups defined by biopsy and imaging status.

Results

In total 268 patients had complete data. Those diagnosed with biopsy- and/or imaging-positive GCA were more likely to demonstrate classic features including jaw claudication and vision loss. Clinical multivariable modelling allowed for fair discriminability [receiver operating characteristic (ROC) 0.759, 95% CI: 0.703, 0.815] for diagnosing GCA; there was excellent discriminability in facilitating the diagnosis of biopsy-positive GCA (ROC 0.949, 0.898–1.000). When used in those with a pre-test probability of 50% or higher, TA-MRA had a positive predictive value of 93.0%; in those with a pre-test probability of 25% or less TA-MRA had a negative predictive value of 89.5%.

Conclusion

In those with high disease probability, TA-MRA can effectively rule in disease (and replace temporal artery biopsy). In those with low to medium probability, TA-MRA can help rule out the disease, but this continues to be a challenging diagnostic population.

Introduction

GCA is the most common large-vessel vasculitis, with an estimated prevalence of 0.2% in North America in adults over 50 [1]. It can be difficult to diagnose due to non-specific symptoms, and must be frequently considered as, if missed, it can complicate to anterior ischaemic optic neuritis that causes blindness in up to 65% of those untreated [2–4]. Temporal artery biopsy (TAB) has long been the ‘gold standard’ investigation to diagnose GCA with pathognomonic giant cells and arteritis; however, it is invasive and insensitive—meta-analyses suggesting it has a sensitivity of 25–77% and ‘biopsy negative’ GCA diagnosed in the majority [5, 6]. While the ACR classification criteria published in 1990 were thought to be highly sensitive and specific, subsequent studies suggest that these estimates were optimistic [7–9]. Ultimately, it is a clinical diagnosis based on clinical features, inflammatory markers, response to glucocorticoid therapy, and, more recently, various imaging modalities [10].

Temporal artery ultrasound has become well explored over the last two decades as it is easily available and is cost-efficient [11]. A positive ‘halo sign’ and/or non-compressible arteries have been found to have sensitivity of 68–81% and specificity of 81–91% [10, 12]. Temporal arterial magnetic resonance angiography (TA-MRA) has been found to be similarly effective for detecting vasculitic changes with similar diagnostic utility, potentially better negative predictive value (up to 98.2%), and the capacity to visualize more vessels than ultrasound [13]. Positron emission topography (PET) has also demonstrated promise, and can be used to assess for the larger extra-cranial vessels; however, there is significant variability in accessing this modality [10, 14].

Diagnosing GCA, despite these advances, continues to be challenging. Several models have emerged that have sought to diagnose biopsy-positive [15–18] or ‘clinical’ [19, 20] GCA, though as of yet there are no algorithms or models that incorporate the use of imaging alongside clinical risk stratification despite it being included in the modern standard of care. This study seeks to explore how TA-MRA can be best integrated into diagnostic models, either in parallel as an up-front assessment or in serial after appropriate clinical evaluation, and the efficacy of these models in diagnosing TAB-positive and TAB-negative GCA.

Patients and methods

Patients

Patient records were found from two sources: a previously published study demonstrating the utility of TA-MRA in diagnosing GCA by Rhéaume et al. (data was provided by study authors), and a prospective cohort that consists of patients who were referred to academic or community rheumatologists in Hamilton, Ontario, Canada for assessment of a potential diagnosis of GCA between November 2013 and June 2018 and who opportunistically consented to have their information collected [13].

All patients in the Rhéaume et al. population underwent TAB as well as TA-MRA; those in the second cohort only underwent investigations that were deemed appropriate by the assessing clinician. Details concerning the performance of TA-MRA are available in the protocol of the original paper; in brief, a 3 Tesla MRI was used with three-dimensional time-of-flight magnetic resonance angiography and fat-saturated spin-echo T1-weighted images to collect images of the arteries and vessel walls, respectively, and demonstrate inflammation [13]. TA-MRAs were assessed to be abnormal based on a four-point scale where a score of 2 or greater was considered abnormal. All TA-MRAs were read by the same radiologist (R.R.), across both cohorts; in total he has read over 500 TA-MRAs. For all patients, information concerning a patient’s clinical presentation, inflammatory markers, relevant imaging, use of steroids, and TAB was recorded. A diagnosis of GCA was made clinically using all information available by the assessing rheumatologist.

Analysis

Descriptive analysis of demographics, presenting symptoms, examination findings, and investigations with odds ratios (ORs) and 95% CIs were calculated. Odds for continuous variables (age, ESR, and CRP) were analysed using univariate logistic regression; binary variables were compared using χ². Descriptive subgroup analysis was performed on those who were diagnosed with GCA with either a positive or negative TAB, as well as those who had a positive or negative TA-MRA.

Predictive models were constructed to predict the diagnosis of GCA, either with or without TA-MRA across the cohort, as well as the biopsy and TA-MRA subgroups, using multivariable logistic regression modelling. All patient data available was considered for analysis. As several smaller models have been previously constructed, several categorizations of age and inflammatory markers were generated and included in the variable selection process [15, 21]. Continuous variables (age, CRP, ESR) were analysed for skew and appropriately normalizing transformations were applied.

Models were constructed using a two-step variable selection process. Initially, a best forward subset selection algorithm was used on sets of variables that contained anywhere from one to a maximum number of variables equal to 10% of the number of events to avoid overfitting [22]. The final variable sets with the highest χ² value for each number of variables then underwent regression to determine the model’s Akaike information criterion (AIC), area under the receiver-operating curve (AUROC), Spiegelhalter test score, and Hosmer-Lemeshow test score using a quintile approach. The model selected for final analysis was the one with the lowest AIC [23]. Goodness-of-fit testing minimum criteria were not required due to the fragility with low event numbers [24].

The selected models were then analysed for variable coefficients, corresponding ORs, and deviance residuals. Multicollinearity was assessed by evaluating the variance inflation factor of a linearized model. Using the cohort data, a set of probability cut-offs was used to determine each model’s performance to rule out a diagnosis of GCA if the patient’s probability was lower than the 10th, 25th, or 50th percentile; or to rule in a diagnosis of GCA if the patient’s probability was equal to or greater than the 50th, 75th, or 90th percentile. An optimal cut-off for the model was also calculated using the method described by Unal for comparison [25]. In addition to the models predicting GCA either inclusive or exclusive of the performance of TA-MRA (parallel and exclusive models, respectively), a serial TA-MRA model was constructed, which demonstrated the test performance of TA-MRA at given post-test probabilities calculated from the TA-MRA-exclusive clinical model. A sensitivity analysis was performed across all models where individuals with a negative TA-MRA, negative TAB, and not meeting the ACR 1990 classification criteria for GCA were excluded.

All calculations and analyses were generated using SAS^® Studio software, version 3.8 for Windows [26].

Ethical approval was obtained from the local institutional ethical board review for both cohorts (Approval numbers 15–403-D and 06–2732), and each patient provided written consent to be included in the cohort.

Results

Demographics, clinical features, investigations, and biopsy results can be seen in Table 1; outcomes for patients from each cohort can be seen in Supplementary Data, available at Rheumatology online, Fig. 1. Of the individuals assessed, those diagnosed with GCA were older and had higher prevalence of jaw claudication, vision loss, and increased inflammatory markers. When stratified by biopsy status, biopsy-positive patients were more likely to have jaw claudication, vision loss, constitutional symptoms, weight loss, elevated inflammatory markers, abnormal TA-MRA, and satisfy the 1990 ACR classification criteria. Whenstratified by TA-MRA positivity, those who were diagnosed with TA-MRA-positive GCA were likely to be older, male, have an elevated CRP, and demonstrate ischaemic symptoms including jaw claudication and vision loss (Table 2). As not all individuals in the prospective cohort underwent TAB, there were more who could be assessed as TA-MRA positive or negative GCA rather than TAB positive or negative.

Across the cohort, TA-MRA had a sensitivity of 64.7% (95% CI: 55.4-73.2) and specificity of 91.5% (86.6–95.1), it had 82.8% PPV and 80.5% NPV. The 1990 ACR criteria were found to have a sensitivity of 54.6% (45.7–63.4) and specificity of 66.0% (58.9–72.6), and TAB had a sensitivity of 38.9% (29.9–48.6) with 100% (96.6–100.0) specificity. Given the NPV seen in the Rhéaume et al. study [13], many of the patients in the prospective cohort with low-moderate pre-test risk underwent TA-MRA first and then only a TAB if the former was abnormal, hence there were lower rates of biopsy in the opportunistic cohort.

Model characteristics

For the purposes of model construction, 268 (82.7%) of the 324 patient records had sufficient covariates to be used in model generation. Both CRP and ESR were found to have a skewed distribution that was able to be normalized through logarithm transformation. Across all models, no variance inflation factor was larger than 1.5 using linearized models (not shown).

The results of the model selection process can be seen in Tables 3 and 4. Across all models, log (CRP) was found to be a predictor of GCA (OR 2.22–3.62), as well as a positive TA-MRA for all models except the clinical model, which did not include TA-MRA as one of the eligible variables. TA-MRA was found to be the most consistent predictor of any diagnosis of GCA, regardless of biopsy status (OR 9.29–43.40). Temporal artery tenderness, jaw claudication and vision loss were the clinical features most commonly seen to be positively associated with a diagnosis of TAB-positive GCA. Headache was not predictive of a diagnosis of GCA in any model.

The AUCs for each model demonstrated that there were clear differences in the facility of predicting biopsy-positive or biopsy-negative GCA. While the addition of TA-MRA to the clinical model increased the AUROC of the model, when compared across a standardized subset of the data the AUC was not significantly higher (χ² = 3.42, P = 0.06). Within the clinical diagnosis model, there were only two patient records that had deviance residuals >2; both were individuals with non-specific symptoms, a normal MRI, no TAB, normal inflammatory markers, and had model probabilities of 0.068 and 0.071 that were both ultimately clinically diagnosed as having GCA. In the clinical diagnosis and TA-MRA model the same two individuals as well as two others with similar features (except for elevated ESRs of 31 and 72 mm/h, respectively) were ultimately diagnosed with GCA.

In the model predicting biopsy-positive GCA (Table 4), an AUROC of 0.95 (0.90–1.00) was seen with the three variables used which each demonstrated large magnitudes of association. The TAB model was the only model in which the Hosmer-Lemeshow test required rejection of the null hypothesis suggesting that there was poor fit (χ² = 20.58, P < 0.01), though given the low number of events even with the lower number of bins, the validity of the test is difficult to assess; the Spiegelhalter test, which is less limited by the impact of small numbers demonstrated goodness of fit (as well as with other models). The only individual with a high deviance residual was a patient presenting with non-specific symptoms, an elevated ESR of 46 mm/h, a normal CRP of 3.5 ng/ml and a negative TA-MRA and calculated model probability of 0.002, yet still demonstrated TAB positivity. The model predicting biopsy-negative GCA was the only model that included weight loss, which demonstrated negative predictive capacity but was not clearly independently predictive of the diagnosis (OR 0.27, 0.05–1.47). There were no patients who had a deviance residual >2.

Model performance and cut-offs

Using purely clinical features and markers of inflammation (Table 5A) there was low sensitivity seen with progressively more extreme probabilities, specificity was well preserved. Despite this, the misclassification rate was between 30.0% and 56.9%, with the majority of these rates being driven by false negatives. The ideal cut-off, which seeks to minimize the sum of the absolute value of the differences between the sensitivity, specificity, and AUROC demonstrated a better NPV with optimal misclassification (at P = 0.364, sensitivity 71.4%, and specificity 64.9%).

The model that included TA-MRA in parallel (Table 5B) did not demonstrate a statistically superior AUROC but did demonstrate greater diagnostic discriminability across all parameters compared with the model with only clinical findings, and lower misclassification rates except in using very low probabilities to rule out disease. The ideal cut-off demonstrated greater specificity to rule in disease and had higher PPV where sensitivity and negative predictive value were largely maintained.

Sensitivity analyses of both did not demonstrate changes in the discriminability of the model; however, the misclassification rate (MCR) was lowered, more so when ruling out disease (MCR 37.9% to rule out disease at <10% post-test probability on sensitivity analysis compared with 55.0% across all patients), and there was an improved NPV (NPV 90.7% with a calculated probability ≤25%).

In comparison to the parallel model, TA-MRA can be used in serial with clinical assessment using a decision model seen in Fig. 1. Clinical assessment is used to create a pre-test probability. This can then be classified into levels of risk stratification where TA-MRA performance allows one to rule in or rule out a diagnosis of GCA with various levels of confidence. In those with a 20% or lower pre-test probability of GCA, TA-MRA can be used to effectively rule out disease whereas in those with a pre-test probability of 40% or more it can be used to rule it in. While specificity was 92.6% for those with a pre-test probability of 20–40% to rule in a diagnosis of GCA, a 26.0% MCR was noted.

Discussion

This cohort demonstrates that ischaemic symptoms, increasing age and increased inflammatory markers are most predictive of a clinical diagnosis of GCA, consistent with other studies [15–17]. Neither headache nor non-ischaemic visual symptoms were found to be associated with the diagnosis of GCA, suggesting these findings are non-specific. Other findings that have been variably associated with GCA including scalp and temporal artery tenderness were not found to be different in comparative analysis but were found to be important considerations in the discriminative algorithm.

This cohort provides further information on two under-explored groups: those who are biopsy negative and those who have negative TA-MRA. The former of these likely composes the majority of GCA populations; yields of TAB in two systematic reviews varied from 25% to 77.3% depending on the review methodology, and other papers that compare biopsy-positive vs biopsy-negative cases have suggested anywhere from 15–78% [5, 6, 27, 28]. Across these previous cohorts as well as this one, biopsy positivity appears to indicate a more severe phenotype or an advanced stage of disease [27, 29].

A cohort of those who are both biopsy and imaging negative has not been well described, and based on the findings of the 42 individuals within this cohort, may indicate an earlier stage of disease as this group tended to have more non-specific symptoms and a lower burden of inflammation, though as noted within the sensitivity analysis, eight individuals were diagnosed despite negative imaging, biopsy and ACR criteria (the other 34 met ACR criteria), where with such limited symptoms some diagnostic ambiguity is likely to exist. Indeed, these ambiguities highlight the difficulty in developing effective diagnostic criteria. It also highlights the utility of imaging in providing better alternative diagnoses, in addition to demonstrating that prognostic models can be effectively used in conjunction with, but not in replacement of, clinical gestalt.

While much of the research to date has focussed on ultrasound as the modality of choice, TA-MRA has been demonstrated to have likely diagnostic equivalence, though a true head-to-head diagnostic trial has yet to be completed [10–12, 30]. It is thought, however, that TA-MRA may provide greater sensitivity, particularly for TAB-negative GCA, as TA-MRA allows for a larger field of view and other vessels not typically visualized with ultrasound. While the cohort analysed by Rhéaume et al. is a part of this cohort, the drop in NPV (from 98.2% to 80.5%) likely reflects the differences between study use of an investigation, where all individuals are investigated, and real-world use, where only the individuals whom the clinician feels are best served by the investigation are imaged [13]. Indeed, this real-world usage may also be responsible for the lowered sensitivity (64.7%, 95% CI: 55.4, 73.2) with preserved specificity, where referred patients have a different distribution of disease probability, and some have alternative diagnoses that lead to false positives.

The use of parsimonious methods of model derivation within this cohort allows for unique insight into how imaging can interact with and assist clinical evaluation in diagnosing GCA. The addition of TA-MRA into the models found that within this cohort, imaging has similar diagnostic utility compared with clinical evaluation for GCA as seen in Table 3, as TA-MRA was able to replace virtually all features as a single variable (OR 15.03, 7.50–30.12) within the parallel model. When considered across various diagnostic cut-offs in Table 5A and 5B, it was clear that the addition of TA-MRA in parallel to clinical evaluation improved all diagnostic parameters, especially in those who had mid-range probabilities (0.25 ≤ P ≤ 0.75), with the greatest changes seen in the positive predictive and misclassification rate.

When used in serial to clinical evaluations, as demonstrated in Table 3 and Fig. 1, the utility of TA-MRA was evident. Indeed, for those with a 40% or higher pre-test probability of GCA the disease can be ruled in with a positive scan; a negative scan will likely rule out the disease but given that 12 of 105 TA-MRAs were false negatives, a second investigation would likely be merited to provide clinical certainty. In those with 20% or lower pre-test probability, a negative TA-MRA essentially rules out GCA, and a positive scan, similar to a negative scan in the high pre-test probability, mandates further evaluation. For those with 20–40% pre-test probability, however, a positive or negative TA-MRA can add significant diagnostic weight but would likely benefit from multiple modalities of investigations and a possible trial of glucocorticoids to provide better diagnostic certainty. This is particularly pertinent in GCA where electing to either treat or not treat for GCA can have dramatic consequences—either a long course of steroids or a risk of vision loss.

Conversely, the subgroup models (Table 4) indicate that in the presence of one of the suggestive symptoms for GCA, positive imaging is equivalent to biopsy positivity (ROC 0.95, 95% CI: 0.90, 1.00). A model that did not include TA-MRA, where vision loss, jaw claudication and log CRP were used had AUROC of 0.91 (0.86–0.97), further demonstrating that in individuals with clearly suggestive symptoms, imaging can be used to confirm the diagnosis. These findings allow for an alternative and potentially accelerated diagnostic pathway when TAB may be difficult to obtain or a patient may be a poor surgical candidate; as a positive TAB does not appear to change outcomes, it may be obviated altogether [28]. Further, by minimizing the chance that treatment may be delayed, outcomes can be maximized for those with a more severe phenotype who are more likely to suffer vision loss [31].

In biopsy-negative GCA patients, it appears that ‘classic’ clinical features are more important for triangulating a diagnosis (as shown both in Table 1 and in their lack of selection in the parsimonious model). This may suggest that the natural history of GCA is that symptoms occur either before or concomitant to changes on imaging, and positive biopsy is seen in those with sufficiently advanced disease to manifest vascular occlusive phenomenon, though not all individuals with GCA may develop such a severe burden of disease.

These findings validate the suggested diagnostic strategies proposed by Mackie et al. and Ponte et al., which were used as the basis for the clinical algorithm developed in Fig. 1, who suggest imaging paired with clinical assessment in the initial diagnostic evaluation of individuals being considered for a possible diagnosis of GCA [32, 33]. These algorithms place ultrasound as the first investigation, which is not unreasonable given the increased availability of ultrasound compared with 3 Tesla MRI. This cohort, as evidenced in Fig. 1, demonstrates that TA-MRA has at least similar performance to ultrasound, and may have better utility for indivdiuals with low-to-medium probabilities of GCA, who are often the subject of the greatest clinical uncertainty. This suggests that a ‘right person, right test’ approach to the diagnosis of GCA will be the most fruitful strategy. Further, ultrasound only visualizes superficial cranial arteries while TA-MRA reveals the intracranial vessels . As such, for those with high pre-test probability of GCA, any positive imaging may be sufficient to rule in GCA, and in those with medium or lower probability TA-MRA may be the optimal first investigation, and is likely to be the optimal second investigation in those who have no clear features of GCA on ultrasound.

The limitations of this study include the fact that the diagnoses of GCA were made by an individual clinician based on their clinical judgement according to disease characteristics, diagnostic tests and response to treatment. This will assuredly lead to a degree of misclassification of those without true GCA as having GCA. This is likely to negatively bias the performance of TA-MRA as we would expect that more people without GCA would be diagnosed with it rather than individuals with GCA being diagnosed as not having the disease, thus reinforcing the findings seen here. Our confidence that the misclassification errors are acceptable is further improved by the comparability of the variables selected and their magnitude to other studies that utilized more rigorous methods, such as independent review, to classify patients with GCA [17, 19, 34]. Despite our confidence, replication of our results are warranted before wider use of TA-MRA. A further limitation is that we used relatively simple TA-MRA data and that other potential benefits of TA-MRA are not considered, such as finding alternative diagnoses such as temporomandibular joint synovitis as well as the opportunity to visualize other vessels including the ophthalmic arteries [10, 35–37].

Funding: No specific funding was received from any bodies in the public, commercial or not-for-profit sectors to carry out the work described in this article.

Disclosure statement: Dr Khalidi reports personal fees and non-financial support from Roche, non-financial support from Bristol Meyers Squibb, and Abbvie outside the submitted work. Dr Garner reports grants/research support from Roche.

Data availability statement

Upon reasonable request.

Supplementary data

Supplementary data are available at Rheumatology online.

References