Cross-modal evidence progression for interpretable ophthalmic diagnosis under incomplete clinical observations

Abstract

The integration of cross-modal clinical data, such as optical coherence tomography, color fundus photography, and electronic health records, has significantly advanced automated ophthalmic diagnosis. However, real-world clinical environments frequently suffer from incomplete observations due to varying equipment availability, patient compliance, and clinical workflow constraints. Moreover, black-box deep learning models lack the interpretability required for clinical trust and regulatory approval. This paper presents a novel framework for cross-modal evidence progression designed specifically for interpretable ophthalmic diagnosis under conditions of incomplete clinical observations. The proposed methodology leverages an evidential reasoning mechanism that dynamically accumulates diagnostic evidence from available modalities while estimating and mitigating the uncertainty introduced by missing data. Instead of relying on traditional data imputation techniques, our approach maps available modality-specific features into a shared evidential space, projecting multimodal representations into interpretable diagnostic trajectories. Comprehensive evaluations on large-scale ophthalmic datasets demonstrate that the proposed framework maintains high diagnostic accuracy and robustness even when up to half of the modalities are missing. Furthermore, the evidence progression mechanism provides transparent, step-by-step diagnostic reasoning, offering clinicians clear insights into how the model arrives at its predictions. The findings suggest that dynamic evidence accumulation is a highly viable pathway for deploying trustworthy artificial intelligence in resource-constrained clinical settings.