Classification criteria developed using machine learning will allow doctors to conduct disease-specific research about this collection of inflammatory eye diseases.
The Standardization of Uveitis Nomenclature Working Group (SUN) has developed classification criteria for 25 of the most common uveitides, or diseases characterized by intraocular inflammation, using machine learning, according to research published in the American Journal of Ophthalmology and funded by the National Eye Institute.
Prior to the criteria developed by SUN, the classification of uveitis was determined by the location at which inflammation was clinically detected, whether it be in the anterior uveitis, intermediate uveitis, posterior uveitis, or panuveitis. However, inflammation detected at the same location can have different potential causes, courses, prognoses, and treatment needs, which leads to unclear diagnoses and a lack of comparability between cases.
The adoption of classification criteria to determine subtypes of uveitis will help clarify the uncertainty that exists among uveitis experts on the comparability of different cases and classification of diseases. The development will enable scientists to identify homogenous groups of patients that are generally accepted to have the same disease and conduct diagnostic research.
SUN developed the classification criteria in 4 phases, but ultimately utilized machine learning to distinguish each disease. During the informatics phase, experts standardized language to define subtypes of uveitis. The subsequent case collection phase sorted 5766 cases of uveitis and found an average of 100 to 250 cases for each subtype.
In the case selection phase, committees of 9 uveitis experts reviewed the cases and used formal consensus techniques to determine which disease subtypes to include in the final database. The remaining 4046 cases that received supermajority agreement from the experts were screened through machine learning and assigned to either a training set or a validation set.
Using multinomial logistic regression with lasso regularization, machine learning determined a set of distinguishing criteria for each disease belonging to the training set and evaluated the resulting criteria in the validation set. The rules established to express the criteria were evaluated by an observer for accuracy in a 10% random sample of cases and found to be above 95% for each subtype of uveitis.
The overall accuracy estimates for each uveitic class in the validation set were:
The high overall accuracies for the 25 uveitides in the validation set suggest that the established criteria have low misclassification rates and that it is acceptable for the criteria to be used in clinical and translational research. The proposed criteria were granted approval by the SUN Working Group to be used in future research.
“In the past, clinical research in the field of uveitis has been hampered by the lack of widely-accepted and validated diagnostic criteria,” said Douglas A. Jabs, MD, MBA, the SUN project leader and professor of epidemiology and ophthalmology, Johns Hopkins Bloomberg School of Public Health, Baltimore, in a statement. “These classification criteria are a major step forward for epidemiological studies, translational studies, pathogenesis research, outcomes research, and clinical trials. They hopefully will yield better disease-specific approaches to diagnosis and treatment.”
Jabs DA, McCluskey P, Oden N, et al. Development of classification criteria for the uveitides. Am J Ophthalmol. Published online April 9, 2021. doi: 10.1016/j.ajo.2021.03.061