An Explainable Stepwise Framework for Etiological Diagnosis of Uveitis

Uveitis remains difficult to diagnose etiologically because heterogeneous clinical information must be integrated and the remaining uncertainty resolved through sequentially ordered complementary examinations. We developed an explainable stepwise framework for this staged process. In the pre-diagnostic stage, a Probable Etiologies Diagnostic model (ProED) ranks the Top-N most likely etiologies from demographic, historical, ophthalmic and clinical data that are available before laboratory work-up. In the interactive stage, an explainable algorithm recommends complementary examinations and progressively excludes unlikely etiologies, ranking unperformed tests by expected information gain under a burden proxy that combines examination count, invasiveness level and examination modality. The framework was evaluated on a single-center retrospective cohort covering 25 uveitis etiologies, with a public dermatology dataset used as a proxy for feature-restricted conditions. ProED achieved a Top-5 accuracy of 96.37%. The interactive stage achieved an accuracy of 81.3%, a weighted precision of 82.2%, a weighted recall of 80.3% and a weighted F1-score of 81.6%, improving to 90.6% accuracy and a 91.7% weighted F1-score when the two most likely etiologies were accepted. However, the framework quantifies only a simulated reduction in recommended examinations; financial cost, waiting time, patient discomfort and resource use are approximated by the weighting scheme rather than measured, so its real-world clinical, operational and economic impact requires prospective multi-center evaluation with explicit cost, workflow and patient-experience endpoints.