What does IGRA testing add to the diagnosis of ocular tuberculosis? A Bayesian latent class analysis

Tuberculosis (TB) still remains a major public health problem in most countries, including developed nations. With a large number of immigrants in the United Kingdom (UK), TB is disproportionately distributed amongst new migrants and certain ethnic groups who carry social risk factors and tend to be more adversely affected. [1] Seven thousand, eight hundred and ninety-two cases of TB were notified in the UK in 2013 with an incidence of 12.3/100,000. London accounted for the highest proportion of cases in the UK (37.8%, 2985/7892), with a rate of 35.5/100,000. [1] A significant proportion of these patients presents with extrapulmonary latent mycobacterium TB. Over the last decade, we have witnessed an apparent increase in the incidence of ocular TB. [1] This apparent increase may be attributable to a higher rate of diagnosis due to availability of the interferon gamma release assay test (IGRA). Ocular TB, as a form of latent extrapulmonary TB is faced with the multifaceted problem of diagnosis from inconclusive signs and corraborative investigations. It can present as anterior uveitis, intermediate uveitis, vitritis, retinal vasculitis, neuroretinitis, solitary or multiple choroidal tubercles, serpiginous-like choroiditis, subretinal abscess, endophthalmitis and panophthalmitis [2–6].

Ocular TB has been postulated to arise either as a result of direct infection with TB bacilli or as a hypersensitivity reaction to latent TB infection. [4, 6–8]. In-vitro assessment of interferon-gamma (IFNγ) via IGRA has been an important addition to the diagnosis of TB and has increased the sensitivity in diagnosing latent TB. [9, 10] QuantiFERON-TB Gold In-tube test (QFT; Cellestis Limited, Carnegie, Victoria, Australia) is Food and Drug Administration (FDA) approved IGRA test used for identifying latent TB infection. [9–11] According to the manufacturer’s instructions, QFT is considered positive for values > 0.35 IU/mL. Based on Centers for disease control and prevention (CDC), QFT was approved to be used for diagnosis of latent TB globally and same was adopted by National Institute for Health and Care Excellence (NICE) in the UK. [12].

The absolute values of QFT was interrogated by Gineys et al. and values of 2 IU/ml was suggested to be associated with better outcome in patients with ocular TB from low endemic settings. [13] Positive QFT results can occur in patients with non uveitic entities (optic neuritis/ orbital inflammation/ myositis/ sarcoidosis) and also in patients with non tuberculous bacilli (M. Kansasii, M. szulgai, M. marinum) or even on false handling of samples (with erroneously large number of TB antigens in QFT tube) leading to lack of specificity of the test. [14, 15] The reported specificity of the QFT for pulmonary and latent TB ranges from 91 to 99%, but reported sensitivity is somewhat lower 89–91%. [16, 17] The accuracy of IGRA in patients with ocular TB in a BCG-vaccinated, non-endemic population was recently evaluated and reported sensitivity and specificity were 80 and 85%, signifying impact of IGRA in this cohort of population. [18] In a recent review published by Getahun et al., the authors have looked into the worldwide literature and concluded that there is no perfect method for the diagnosis of latent mycobacterial infection [19].

Rosenbaum et al. has described in their study on utility of routine screening of patients with uveitis for TB, post-test probability of a test can be deduced, via Bayes’ theorem, by using the known sensitivity and specificity of test as well as pre-test probability of the presence of ocular TB. [20] In the present study, we will investigate the predictive value of the QFT for the diagnosis of presumed ocular TB using a Bayesian latent class analysis of the diverse population in a tertiary eye care setting, so as to determine if QFT is justifiable as a routine screening tool for patients with presumed ocular TB. Our aims were to combine all available data into a statistical model predicting each patient’s ocular TB status, and to compare the ability of this model, with and without the QFT results, to predict treatment failure (steroid taper, defined below).

This was a retrospective cohort study of patients seen at a large tertiary referral centre in the UK. Medical records of patients with presumed ocular TB were reviewed following approval by the Institutional Review Board (IRB). Patient records/information was anonymized and de-identified prior to analysis. This research was performed in adherence to the tenets of the declaration of Helsinki. Records of all the cases between 2005 to 2013 who had IGRA test and Mantoux test done were identified based on data from clinical laboratory at the centre.

The clinical case definition of presumed ocular TB was based on broad non-specific clinical signs proposed by Gupta et al. in 2007. [4] Clinical phenotypes suggestive of TB include anterior uveitis (granulomatous, non-granulomatous), intermediate uveitis, posterior uveitis (choroidal tubercle or tuberculoma, subretinal abscess, serpiginous-like choroiditis), panuveitis, retinitis, retinal vasculitis and neuroretinitis. Patient data were retrieved and data collected included information on demographics, morphological diagnosis, investigations and therapeutic regimens. Patients at presentation were examined by the residents, along with senior uveitis specialists, for signs suggestive of ocular TB. All patients underwent comprehensive examination and investigations to exclude other infectious diseases and immunologic disorders, including chest radiograph. Angiotensin converting enzyme (ACE) and serology tests (HIV, syphilis, hepatitis B and C) were done, if deemed necessary. Patients with a clinical diagnosis of uveitis suggestive of TB in one or both eyes were included in the study. The QFT assay was performed as per manufacturer’s recommendations and results were interpreted as per the recommendations from manufacturer. [21] Patients with one of the ocular phenotypes mentioned above, and an elevated QFT or other features suggestive of ocular TB, were referred to a respiratory physician for initiation of ATT. All attempts were made to exclude cases with sarcoidosis using information from combination of biochemical test or radiological test.

Moreover, patients with insufficient follow-up visits were excluded from the study. Data were subsequently double-entered into a database by two co-authors (RA and JC) and cross-checked for accuracy; discrepancies resolved by referring back to medical records.

Statistical analysis

In line with manufacturer’s recommendations, values of QFT < 0.35 units were regarded as being equivalent to zero. A transformation of the QFT data was sought to facilitate model a mixture of small, positively skewed values (representing TB-negative patients) and larger, symmetrically-distributed values (representing TB-positive patients) [22].

Exploratory analyses were conducted to find differences in the use of ATT and treatment failure, as well as QFT or ACE between ethnic groups, gender, or presence/absence of binary signs, using non-parametric statistical tests. Correlation between age and QFT, as well as ACE and QFT, was assessed using Spearman’s rho coefficient. All exploratory analyses were conducted in R software (version 3.2.3) [www.r-project.org]. Raised ACE may signify a sarcoid kind of presentation and as per our recently published classification. [23] It represents tuberculous-sarcoid kind of presentation and not purely a tubercular type of presentation [23].

A Bayesian latent class model was then constructed using Stan software (rstan interface version 2.8.0) [mc-stan.org]; code is given in the technical appendix (Additional file 1). Bayesian methods are useful for statistical modelling in the context of our data because they can accommodate complex models and provide intuitive outputs regarding values of interest in terms of probability. They are solved by computer-intensive simulation to find values that fit the observed data. [24] (Additional file 1). Statistically, the estimated patient status is called a latent class, but to avoid confusion between latent class and latent TB infection, we use the term “estimated probability of ocular TB”. All associations in the model are shown together in Fig. 1. The latent class approach allows for relationships that both predict the estimated probability of ocular TB (for example, ethnicity is a proxy risk factor for exposure to Mycobacterium Tuberculosis), and are a result of, ocular TB (for example, the risk of choroiditis might differ between the patients who truly have ocular TB and those who do not). For each relationship there was a regression coefficient, and all the relationships are linked and evaluated simultaneously, along with each patient’s estimated probability of ocular TB. This approach is increasingly popular to evaluate diagnostic tests in the absence of a gold standard [25, 26].

We did not include subjective opinions as prior probabilities, but instead used the weakly informative prior approach advocated by Gelman and colleagues, [26] which avoids extreme values that are implausible and potentially problematic for computation. Residual QFT values (the mean of the observed QFT minus predicted QFT, across all steps from Stan) were plotted against observed values and other potential predictors such as chest X-ray evidence of prior active pulmonary TB to diagnose potential problems in the models. Each patient’s estimated probability of ocular TB was plotted with its confidence interval and against the recorded level of QFT to highlight strengths and weaknesses of the model. The dataset is provided in Additional file 2.

The 58 (21.7%) patients with panuveitis features (with uveitis at more than one location) appeared differed significantly from those with uveitis in the following ways:

Raising QFT to the power of 0.25 (the ‘quartic root’) produced a distribution amenable to further statistical modelling (Fig. 2); we refer to this as QFT4. It can be seen to fall into two parts, one positively skewed with mode close to zero, and the other symmetric and centred near 1.2 (QFT = 5.28). The higher part appears bimodal (has two peaks) in this chart but with relatively small numbers, and in the absence of a biological explanation, we regard this as one common group of patients. Future research with larger samples can confirm the distribution of QFT values in this population. The least common region, which separates these two distributions and is the best candidate for a clinical cut-off level, is near 0.4 (QFT = 0.25).

The model does not provide satisfactory predictive ability. Each patient’s estimated probability of ocular TB, with 95% confidence intervals, is shown in Fig. 3 . Of the 267 patients, 68 had CIs straddling either side of 0.5, while 47 were entirely below and 152 entirely above. Based on the estimated probability, taking 0.5 as a cutoff, 68 were below 0.5 and 199 above, which might be regarded as probably not TB and probably TB respectively. However, this classification corresponds poorly to treatment failure, which we expected in those who had TB and were not treated with ATT, or did not have TB but were treated with ATT. The positive predictive value (PPV) was 58.7% and the negative predictive value 56.1%, which is not much better than chance.

When the model is run again without QFT, the confidence intervals of Fig. 3 are much wider, with 259 patients straddling 0.5, 8 with CIs entirely below, and none entirely above. The estimated probability was above 0.5 for 148 patients and below for 119, and from this prediction, the PPV was 51.4% and NPV 40.3%.

Despite the means of the two-part distribution being well represented by the estimates in Table 3 , the standard deviations are very wide. The estimated point of crossover between the distribution of TB-negative and -positive patients was at QFT = 0.05 with 95% CI 0.04 to 0.07. However, as seen in Fig. 4, the relationship between QFT4 and the estimated probability of ocular TB (which is informed in part by QFT4 itself) breaks down at high probabilities.

Currently there is no clear consensus on the diagnosis of ocular TB and this remains a challenge for clinicians around the world. There are many postulations behind the pathophysiology of ocular TB. It may be an immune mediated ocular inflammation that is initially triggered by mycobacterium TB, without any underlying active infection. [27] Alternatively, it may result from an inoculum directly within the eye. [28] Although such direct TB infection may manifest such as a choroidal abscess, the majority of patients displayed clinical features, which were not exclusively associated with TB infection. None of the clinical phenotypes in our series from the low endemic region showed a statistically significant association for reduced risk of recurrence with positive or better response to ATT. In our current study, we have investigated the possible predictive value of QFT for the diagnosis of presumed ocular TB from a cohort of patients with signs strongly suggestive of ocular TB, based on Bayesian model analysis.

The best estimate of a QFT level separating TB-positive and TB-negative patients is extremely low, leading most likely to high sensitivity but poor specificity. Although it is not inconsistent with the least common region seen in the observed data around 0.23, the uncertainty is too great to permit a firm conclusion separating the patients into two groups. This indicates the considerable uncertainty about true TB status, even taking all variables into account in a predictive model. In our current model, patients with higher QFT levels and treated with ATT had lower treatment failure as against those patients who did not had ATT with lower QFT values.

Ang et al. have earlier researched the usefulness of IGRA in diagnosis of presumed ocular TB, however the authors have not specified the clinical phenotype associated with positive IGRA and have instead enlisted the anatomic location of uveitis [17]. They have reported a poor sensitivity (36%) and modest specificity (75%) of IGRA test alone in presumed tuberculous uveitis; the area under the receiver-operator curve (AUC: a measure of diagnostic ability combining sensitivity and specificity that ranges from 0 to 1) of 0.555. Combining the IGRA with the results of tuberculin skin testing improved the performance somewhat (AUC of 0.644). The authors concluded that both IGRA and tuberculin skin test had poor sensitivities and that neither test could reliably exclude ocular TB if there were highly suggestive clinical features [17].

Rather than assume a gold standard and use ROC analysis, we have taken a Bayesian approach, as this allows us to express uncertainty around each patient’s estimated probability of ocular TB. Our data suggest that QFT alone does not adequately separate TB uveitis from non-TB uveitis. There are very limited studies on the possible predictive clinical signs from low endemic regions, further compounding the diagnostic conundrum. The challenge may be considered greater in countries such as the UK in which TB is much less prevalent but have a very high number of immigrant population from high endemic zones.

La Distia Nora et al. evaluated the clinical manifestations of patients with intraocular inflammation with positive QFT in nonendemic settings. [29] In the series of 77 patients, 85% of patients were found to be immigrants from endemic region. Retinal occlusive vasculitis and serpiginous choroiditis were commonly noted phenotypes with mean QFT value of 7.5 IU/ml. Like the current study, the authors did not find any positive association between QFT positivity and clinical signs. The authors and the editorial from Pepple et al. have concluded the limited role and risk of false positivity in patients with signs suggestive of ocular TB in low endemic settings and the raised QFT may albeit represent sarcoidosis. [29, 30] To further investigate the association between TB and sarcoidosis, we did correlation analysis between potential biomarkers of sarcoidosis (serum ACE) and TB (QFT). In our study we found no statistically significant association between raised serum ACE and QFT.

In another series by Ahn et al. from Korea (low endemic setting), the authors found strong correlation of positive IGRA in presumed ocular TB. [31] They also found retinal vasculitis and posterior uveitis to be the clinical phenotypes associated with tubercular uveitis (TAU) or ocular TB. Cordero-Coma et al. established immune response prevalence of 32.25% in 31 uveitis patients from Spain with estimated sensitivity of 82% and specificity of 100% with NPV of 86% for positive QFT. [32] Mackensen et al. in 2008 and Gupta et al. in 2010 had established the association of QFT with serpiginous like choroiditis. [6, 33] Babu et al. have demonstrated the use of QFT in the diagnosis of presumed ocular TB in a prospective pilot study from a high endemic region. [34] Sensitivity of QFT was shown to be 82% with specificity of 76% for diagnosis of intraocular TB based on clinical signs suggested by Gupta et al. [6, 34] However, in patients with a definite culture or biopsy proven systemic TB, the test has shown only 58% sensitivity and 77% specificity, highlighting the limitation of QFT in diagnosing active systemic TB [34].

In a population where TB is non-endemic, such as the UK, sensitivity of IGRAs is thought not to be as good as TST. [35, 36] Furthermore, specificity of the test has also been questioned. If IGRAs have helped clinicians with a new window of opportunity to diagnose latent TB it has also caused a reasonable degree of confusion with regards to its false positive and false negative results, with particular variation in low-endemic regions. It is hence essential to review the application and utility of IGRAs in low-endemic countries such as the UK.

The use of ATT to manage ocular TB is regarded as an effective treatment for ocular TB and response to therapy can be a good surrogate for diagnosis of ocular TB. Bansal et al. showed a reduced recurrence of ocular TB following ATT, in addition to steroids. [37] ATT clears latent TB infection and subsequently protects against the manifestation of active TB in up to 90% of patients. [38] Overall it would be expected that ATT would have been effective in treating ocular TB in this study and thus preventing any recurrence. It is unlikely that treatment failure in this study group could be attributed to failure of the ATT itself, and it can be compounded due to the heterogeneity in treatment regimen, inadequate treatment length, antibiotic resistance or non-compliance with the drug regimen.

Based on our observation, it can be postulated that patients with suspected ocular TB get started on steroids early if symptoms are more severe, and for the same reason they are more likely to really have ocular TB hence higher QFT. With past history of pulmonary TB, patients are enrolled in the dataset as suspected ocular TB. As the current study was not a trial of ATT and attempting to analyse the data in terms of treatment efficacy runs the risk of obtaining falsely significant results by multiple testing and confounding by indication. We sought to explore patterns in the data and generate rather than test hypotheses. Likewise, the effect of response to therapy in patients with peripheral retinal vasculitis following panretinal laser photocoagulation could not be assessed in this series, but we have highlighted the outcome of treatment in patients with peripheral retinal vasculitis in our recently published report [39].

This study has several limitations as data was collected retrospectively and researchers have made a decision based on the facts present at their disposal, which can introduce investigator’s bias. There was no information about proportion of TB uveitis patients (amongst all the new uveitis cases seen in a given year), about how many patients had new episodes of inflammation and how many had old ongoing inflammation. There was also a significant variation between oral corticosteroids, immunosuppression and ATT treatment type and length prohibiting us from drawing definitive conclusions. The definition of treatment outcome requires further validation as there is currently no gold standard or consensus guidelines defining the treatment outcome in patients with ocular TB. Furthermore, as the study was from a single centre, it may be difficult to generalize the results of this study. We propose that future directions should include conducting prospective well-designed multicentre (between low and high endemic regions) studies to understand the true formulate and accurate global depiction global picture of presumed ocular TB. A carefully designed multicentre study with clear case definition for ocular TB, non ambigous treatment outcomes, defined follow up with standardized therapy (if not randomized control trial) can tell us more information and perhaps lead to a definite conclusion.

None of the defined clinical phenotypes of ocular TB or positive QFT were associated with favourable outcome with ATT. There was no evidence for a cut-off value of QFT that successfully discriminated between the two groups of patients. Extrapulmonary mycobacterium TB manifesting in the eye remains a diagnostic conundrum.