FLUID: A Longitudinal Multi-Device Ultra-Widefield Fundus Dataset for Diabetic Macular Edema Prediction

The FLUID dataset is a longitudinal ultra-widefield (UWF) fundus imaging dataset designed to support research on the detection and prediction of diabetic macular edema (DME). It contains multimodal clinical and imaging data acquired in routine clinical practice using multiple imaging devices, primarily OPTOS and EIDON systems. The dataset includes retinal images collected across multiple visits per patient, enabling longitudinal analysis of disease progression. Each sample is associated with demographic and clinical metadata, including age, sex, imaging device, eye laterality, and DME labels. Two key annotation variables are provided: (i) the current DME status (DME_label) and (ii) the future DME status at a subsequent visit (next_DME_label), allowing the study of disease evolution and predictive modeling. A key strength of the FLUID dataset lies in its longitudinal structure, which enables the investigation of temporal dynamics and progression patterns of diabetic retinopathy-related complications. In addition, the inclusion of multi-device acquisitions introduces realistic domain variability, making the dataset particularly suitable for developing and evaluating robust and generalizable machine learning models. The dataset is organized at the image level, with each entry corresponding to a UWF fundus image linked to a unique patient identifier and visit information. Images have been standardized and anonymized, and patient identifiers have been replaced with synthetic IDs. The dataset also includes harmonized metadata fields to facilitate downstream analysis. Potential applications of the FLUID dataset include: Development of deep learning models for DME detection and prediction Longitudinal modeling of disease progression Domain adaptation and device-invariant representation learning Multimodal and clinical-data fusion approaches Benchmarking of predictive models for future disease onset Limitations of the dataset include class imbalance in DME labels and potential acquisition biases related to device type and clinical workflow. These aspects should be considered when designing and evaluating machine learning models. This dataset is intended to promote research in predictive ophthalmology and to support the development of clinically relevant AI systems for early detection and monitoring of diabetic macular edema.