Age-related macular degeneration (AMD) is an ordered, bilateral, and longitudinal disease, yet many artificial intelligence systems treat it as static binary image classification. We developed ORBIT-AMD, a multimodal trajectory-learning framework integrating color fundus photography and optical coherence tomography, bilateral eye-graph attention, concept bottlenecks, ordinal staging, cause-specific discrete-time survival prediction, and protocol alignment. In a UK Biobank development/internal-testing cohort of 58 214 participants and 109 691 eyes, and an external Tianjin Medical University Eye Hospital cohort of 1996 participants and 3780 eyes, ORBIT-AMD achieved AUROC values of 0.984 internally and 0.975 externally for prevalent late-AMD detection. Five-year late-AMD progression prediction achieved AUROC values of 0.825 and 0.767, respectively. Calibration and threshold analyses showed cohort-dependent workload and absolute-risk behavior, supporting site-specific calibration assessment and clinical-workflow evaluation before deployment. The concept bottleneck provided auditable lesion-level explanations, but these outputs should be interpreted as structured predictive explanations rather than causal evidence. ORBIT-AMD provides a trajectory-aware framework for AMD risk stratification and review prioritization, with prospective validation required before clinical implementation.
Cui et al. (Sat,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: