AutoOPH: A Multi-Agent System for Glaucoma Research

Glaucoma research requires more than fitting a model to a single image set. A study may involve fundus photographs, optical coherence tomography, structured examinations, and longitudinal records, while its conclusions depend on choices about the cohort, analysis unit, data split, and evaluation. Existing medical agents mainly support diagnosis or individual clinical tasks, whereas general research agents are not designed specifically for glaucoma datas. We introduce AutoOPH, a locally deployable multi-agent system that carries a glaucoma study from literature search and data assessment to experiment design, implementation, reflection, review, and reporting. Its Research Analyst combines verified literature retrieval with a meta-learning strategy that adapts search priorities from earlier accepted research trajectories. A study contract fixes the question, analysis unit, split, metric, and validity checks before iterative development. In a round-matched, three-round reflection comparison on fundus-based glaucoma subtype prediction, AutoOPH achieved a patient-level out-of-fold macro-F1 of 0.5925, compared with 0.565 for an AutoResearch baseline. The controller also rejected a candidate with higher score after the reviewer detected reuse of the evaluation set, and the SEC prevented the rejected result from entering confirmatory memory. These findings shows that the prototype can carry a question through an end-to-end research workflow on a small dataset.