ChronoDR-Net: A Multimodal Temporal Transformer Framework for Longitudinal Prediction of Diabetic Retinopathy Progression

Forecasting diabetic retinopathy progression cannot be reliably addressed as a sequence of isolated classification tasks, but instead requires modeling disease evolution as a longitudinal and patient-specific process. Most existing approaches continue to rely on static fundus image analysis or limited forms of temporal aggregation, which restrict their ability to capture long-range dependencies and conditionally evolving interactions between retinal pathology and systemic risk factors. In this study, diabetic retinopathy progression is treated as a multimodal temporal learning problem in which future disease severity depends jointly on historical retinal appearance and longitudinal clinical context. ChronoDR-Net is introduced as a multimodal temporal transformer framework designed to integrate fundus image sequences with visit-level clinical metadata through structured crossmodal attention. Spatial retinal features are extracted using a Vision Transformer, while temporal self-attention is employed to model disease dynamics across irregularly sampled clinical visits. Clinical variables are represented as contextual tokens that modulate the temporal saliency of retinal features through bidirectional attention, allowing patient-specific progression signals to emerge naturally from the data. Multi-horizon forecasting is performed under a jointly optimized objective that constrains ordinal progression, temporal smoothness, and cross-modal alignment. Evaluation on a longitudinal subset of the EyePACS dataset, followed by external validation on a multi-center cohort, indicates that ChronoDR-Net achieves a Quadratic Weighted Kappa of 0.915 for two-step-ahead severity prediction, outperforming recurrent and metadata-agnostic temporal baselines by a clear margin. Ablation experiments further suggest that deep cross-modal interaction plays a central role in modeling heterogeneous progression trajectories. Taken together, these findings highlight the importance of explicit temporal and conditional multimodal modeling for reliable diabetic retinopathy forecasting and support the use of ChronoDR-Net in personalized longitudinal screening settings.