To determine classification criteria for tubercular uveitis.
Machine learning of cases with tubercular uveitis and 14 other uveitides.
Cases of noninfectious posterior uveitis or panuveitis, and of infectious posterior uveitis or panuveitis, were collected in an informatics-designed preliminary database, and a final database was constructed of cases achieving supermajority agreement on the diagnosis, using formal consensus techniques. Cases were analyzed by anatomic class, and each class was split into a training set and a validation set. Machine learning using multinomial logistic regression was used on the training set to determine a parsimonious set of criteria that minimized the misclassification rate among the intermediate uveitides. The resulting criteria were evaluated on the validation sets.
Two hundred seventy-seven cases of tubercular uveitis were evaluated by machine learning against other uveitides. Key criteria for tubercular uveitis were a compatible uveitic syndrome, including (1) anterior uveitis with iris nodules, (2) serpiginous-like tubercular choroiditis, (3) choroidal nodule (tuberculoma), (4) occlusive retinal vasculitis, and (5) in hosts with evidence of active systemic tuberculosis, multifocal choroiditis; and evidence of tuberculosis, including histologically or microbiologically confirmed infection, positive interferon-γ release assay test, or positive tuberculin skin test. The overall accuracy of the diagnosis of tubercular uveitis vs other uveitides in the validation set was 98.2% (95% confidence interval 96.5, 99.1). The misclassification rates for tubercular uveitis were training set, 3.4%; and validation set, 3.6%.
The criteria for tubercular uveitis had a low misclassification rate and seemed to perform sufficiently well for use in clinical and translational research.
G lobally, disease caused by Mycobacterium tuberculosis is one of the 10 leading causes of death. It is estimated that 10 million persons globally developed tuberculosis (TB) in 2017, and that 1.6 million persons died from TB. Although TB is worldwide in its distribution, in 2017 two-thirds of cases occurred in just 8 countries (India, China, Indonesia, the Philippines, Pakistan, Nigeria, Bangladesh, and South Africa), and 87% of cases occurred in the World Health Organization’s list of 30 high-TB-burden countries. Nine percent of cases of TB disease are co-infected with the human immunodeficiency virus (HIV), of which 72% occurred in Africa. Only 6% of cases occurred in the WHO European region and 3% in the WHO region of the Americas. In the United States in 2017, there were 9,093 cases reported, 70% of which occurred in individuals born outside the United States. Although 10 million persons developed tubercular disease (active TB) globally in 2017, latent TB is estimated to affect 1.7 billion people, about 23% of the world’s population; these individuals are at risk for developing active TB.
Several ocular uveitic presentations have been attributed to ocular TB. These include (1) anterior uveitis with iris nodules; (2) serpiginous-like tubercular choroiditis; (3) choroidal granuloma (ie, tuberculoma); (4) occlusive retinal vasculitis; and (5) in immunocompromised persons, multifocal choroiditis in the context of active systemic TB. Complicating the diagnosis of ocular TB is the fact that a minority of persons with ocular TB have evident active systemic TB, and that approximately one-half or fewer have evidence of current or previous pulmonary infection on chest imaging. As such, diagnosis of ocular TB usually employs evidence of TB infection with a tuberculin skin test (TST) or an interferon-γ release assay (IGRA), neither of which distinguishes between latent and active TB. , There are no randomized clinical trials demonstrating a response to antitubercular therapy, but there are several large case series. , , However, evaluation of treatment studies often has been confounded by the concomitant use of corticosteroids and antitubercular therapy. , , Nevertheless, most patients with the syndromes listed above seem to respond to antitubercular therapy, , , and at least 1 study demonstrated a superior outcome among patients presumed to have ocular TB treated with antitubercular therapy compared to those not treated with antitubercular therapy. Treatment regimens recommended for ocular TB typically presume underlying active TB and treat as such (eg, 4 drugs for 2 months, followed by 2 drugs for an additional 4-10 months). Nevertheless, there is a lack of consensus on the diagnosis and treatment of tubercular uveitis.
The Standardization of Uveitis Nomenclature (SUN) Working Group is an international collaboration that has developed classification criteria for the leading 25 uveitides using a formal approach to development and classification. Among the uveitides studied was tubercular uveitis.
The SUN Developing Classification Criteria for the Uveitides project proceeded in 4 phases, as previously described: (1) informatics, (2) case collection, (3) case selection, and (4) machine learning. , ,
As previously described, the consensus-based informatics phase permitted the development of a standardized vocabulary and the development of a standardized, menu-driven hierarchical case collection instrument.
Case Collection and Case Selection
De-identified information was entered into the SUN preliminary database by the 76 contributing investigators for each disease, as previously described. , , Cases in the preliminary database were reviewed by committees of 9 investigators for selection into the final database, using formal consensus techniques described in the accompanying article. , Because the goal was to develop classification criteria, only cases with a supermajority agreement (>75%) that the case was the disease in question were retained in the final database (ie, were “selected”). ,
The final database was analyzed by anatomic class, and each class was randomly separated into a learning set (∼85% of the cases) and a validation set (∼15% of the cases) for each disease, as described in the accompanying article. Machine learning was used on the learning sets to determine criteria that minimized misclassification. The criteria then were tested on the validation sets; for both the learning set and the validation set, the misclassification rate was calculated for each disease. The misclassification rate was the proportion of cases classified incorrectly by the machine learning algorithm when compared to the consensus diagnosis. For tubercular uveitis the diseases against which it was evaluated in the machine learning phase included infectious posterior uveitis and panuveitides (acute retinal necrosis, cytomegalovirus retinitis, syphilitic uveitis, and toxoplasmic retinitis), noninfectious posterior uveitides (acute posterior multifocal placoid pigment epitheliopathy, birdshot chorioretinitis, multiple evanescent white dot syndrome, multifocal choroiditis with panuveitis, punctate inner choroiditis, serpiginous choroiditis), and noninfectious panuveitides (Behçet disease, sarcoidosis-associated uveitis, sympathetic ophthalmia, Vogt-Koyanagi-Harada disease), respectively. No cases of tubercular anterior uveitis were collected.
Comparison of Cases With Evidence of Systemic Tuberculosis vs Those Without and Comparison of Cases From High-Tuberculosis-Burden Countries vs Those Not From Such Countries
Cases with evidence of TB in an extraocular organ were compared to those with ocular disease alone and a positive TST or IGRA, and cases from high-TB-burden countries were compared to those not from high-TB-burden countries. For categorical variables, comparison was performed with the χ 2 test or the Fisher exact test when the count of a variable was less than 5. Continuous variables were summarized as medians and compared with the Wilcoxon rank sum test. For characteristics with multiple categorical grades, values above and below the median were compared. P values are nominal and 2-sided.
The study adhered to the principles of the Declaration of Helsinki. Institutional Review Boards (IRBs) at each participating center reviewed and approved the study; the study typically was considered either minimal risk or exempt by the individual IRBs.
Three hundred fifty-eight cases of tubercular uveitis were collected, and 277 (77%) achieved supermajority agreement on the diagnosis during the “selection” phase and were used in the machine learning phase. Cases of tubercular uveitis were evaluated in the machine learning for infectious posterior uveitides and panuveitides, noninfectious posterior uveitides, and noninfectious panuveitides. The details of the machine learning results for these diseases are outlined in the accompanying article. The characteristics of cases with tubercular uveitis are listed in Table 1 . Four patterns of cases emerged: (1) serpiginous-like tubercular choroiditis; (2) choroidal nodule (tuberculoma); (3) occlusive retinal vasculitis; and (4) in a small proportion of cases—all with systemic, extraocular TB—multifocal choroiditis. A small number of cases had both retinal vasculitis and multifocal choroiditis. A comparison of cases with evidence of systemic TB and those without (ie, with only a positive TST or IGRA for TB) is shown in Table 2 . Patients in cases with evidence of systemic TB were less likely to be of Asian origin, more likely to have vitreous inflammation, and less likely to have serpiginous-like tubercular choroiditis. A comparison of cases from high-TB-burden countries and those not from high-TB-burden countries is shown in Table 3 . Cases from high-TB-burden countries had a greater proportion of patients with serpiginous-like-choroiditis, whereas those from “low-TB-burden” countries had a greater proportion of patients with retinal vasculitis and with higher grades of anterior chamber and vitreous inflammation. The criteria developed after machine learning are listed in Table 4 . The key features of the criteria are a compatible ocular uveitic syndrome and evidence of infection with TB. Compatible uveitic syndromes included anterior uveitis with iris nodules, serpiginous-like tubercular choroiditis ( Figure 1 ), a choroidal nodule (“tuberculoma”; Figure 2 ), and occlusive retinal vasculitis ( Figure 3 ). The overall accuracies by anatomic class were infectious posterior uveitides and panuveitides, learning set 92.1% and validation set 93.3% (95% confidence interval [CI] 88.1, 96.3); noninfectious posterior uveitides, learning set 93.9% and validation set 98.0% (95% CI 94.3, 99.3); and noninfectious panuveitides, learning set 96.3% and validation set 94.0% (95% CI 89.0, 96.8). The overall accuracy of the diagnosis of tubercular uveitis vs other uveitides in the validation set was 98.2% (95% CI 96.5, 99.1). The misclassification rates for tubercular uveitis in the learning set were as follows: against noninfectious posterior uveitides 7.5%, against noninfectious panuveitides 5.3%, and against infectious posterior uveitides and panuveitides 1.3%. Overall the misclassification rate for tubercular uveitis in the learning set was 3.4%. In the validation set the misclassification rates were as follows: against noninfectious posterior uveitides 0%, against noninfectious panuveitides 6.7%, and against infectious posterior uveitides and panuveitides 5.0%. Overall, the misclassification rate for tubercular uveitis in the validation set was 3.6%.
|Number of cases||277|
|Age, median, years (25th, 75th percentile)||32 (25, 44)|
|Asian, Pacific Islander||80|
|Uveitis course (%)|
|Keratic precipitates (%)|
|Anterior chamber cells, grade (%)|
|Anterior chamber flare, grade (%)|
|Iris atrophy (sectoral, patchy, or diffuse)||0|
|IOP, involved eyes|
|Median, mm Hg (25th, 75th percentile)||15 (14, 18)|
|Proportion patients with IOP > 24 mm Hg either eye (%)||5|
|Vitreous cells, grade (%)|
|Vitreous haze, grade (%)|
|Vitreous snowballs (%)||11|
|Pars plana snowbanks (%)||1|
|Serpiginous-like tubercular choroiditis (%)||43|
|Choroidal nodule (ie, tuberculoma) (%)||4|
|Retinal vasculitis (%) a||53|
|Multifocal choroiditis (%) a||6|
|Bilateral hilar adenopathy||6|
|Immunocompromised patients (%)||2|
|Evidence of infection with Mycobacterium tuberculosis b (%)||100|
|Histologic or culture confirmation of infection in another organ||17|
|Positive tuberculin skin test (eg, PPD)||88|
|Characteristic||Evidence of Systemic TB a||Positive TST or IGRA b Only||P Value|
|Number of cases||48||229|
|Age, median, years (25th, 75th percentile)||32 (23, 48)||32 (25, 44)||.69|
|Uveitis course (%)||.08|
|Keratic precipitates (%)||.07|
|Anterior chamber cells, grade (%)||.83|
|Anterior chamber flare, grade (%)||.13|
|IOP, involved eyes|
|Median, mm Hg (25th, 75th percentile)||14 (12, 17)||16 (14, 18)||.15|
|Percent cases with IOP > 24 mm Hg either eye||4||5||.84|
|Vitreous cells, grade (%)||.62|
|0 or ½+||50||50|
|Vitreous haze, grade (%)||.02|
|Serpiginous-like tubercular choroiditis (%)||17||49||<.001|
|Choroidal nodule (ie, tuberculoma) (%)||10||3||.03|
|Retinal vasculitis (%)||60||52||.29|