Development and Validation of a Bayesian Network for Supporting the Etiological Diagnosis of Uveitis

Abstract

The etiological diagnosis of uveitis is complex. We aimed to implement and validate a Bayesian belief network algorithm for the differential diagnosis of the most relevant causes of uveitis. The training dataset (n = 897) and the test dataset (n = 154) were composed of all incident cases of uveitis admitted to two internal medicine departments, in two independent French centers (Lyon, 2003–2016 and Dijon, 2015–2017). The etiologies of uveitis were classified into eight groups. The algorithm was based on simple epidemiological characteristics (age, gender, and ethnicity) and anatomoclinical features of uveitis. The cross-validated estimate obtained in the training dataset concluded that the etiology of uveitis determined by the experts corresponded to one of the two most probable diagnoses in at least 77% of the cases. In the test dataset, this probability reached at least 83%. For the training and test datasets, when the most likely diagnosis was considered, the highest sensitivity was obtained for spondyloarthritis and HLA-B27-related uveitis (76% and 63%, respectively). The respective specificities were 93% and 54%. This algorithm could help junior and general ophthalmologists in the differential diagnosis of uveitis. It could guide the diagnostic work-up and help in the selection of further diagnostic investigations.

Keywords

• uveitis
• diagnosis
• algorithm
• artificial intelligence
• Bayesian network