RetinaFusion-XAI: Explainable Multimodal Oculomics for Cardiometabolic Risk Prediction and Remote Monitoring

The commonly used traditional methods for evaluating the risk of developing cardiovascular diseases, diabetes and hypertension often overlook early microvascular lesions. These changes which quietly occur in the body's smallest blood vessels are precisely the potential risk signals that fail to be detected in time. By the time traditional methods can identify abnormalities, the optimal window for intervention has usually already passed.

Taking retinal images allows doctors to examine the details of vascular tissues in the eyes without surgery which was originally an excellent way to detect risks in advance. However, most existing artificial intelligence tools used to analyze these images have unavoidable critical flaws: they only operate on the single data type of retinal images; after reaching a conclusion about a disease risk, they cannot explain which specific details they relied on to derive their judgment. What is more, they cannot build corresponding models to track changes in a single patient's vascular health for several years or even more than a decade, failing to accumulate usable data to support continuous monitoring.

This paper proposes RetinaFusion-XAI, a multimodal analysis framework capable of clarifying its reasoning logic. It no longer relies solely on a single type of data and can integrate various routine clinical indicators collected during medical visits, retinal biomarkers extracted from retinal images that reflect the body's state and feature data generated by deep learning models for joint analysis, filling the critical gaps that could not be addressed in previous risk assessments of this type.

Previous analytical tools have long failed to solve two core problems: first, they cannot properly analyze images captured by different imaging devices, resulting in extremely poor device compatibility; second, the conclusions they produce cannot explain the judgment logic to clinicians, making them unusable as reliable references for diagnosis and treatment. In response to these long-standing issues, this new framework incorporates targeted specialized designs: it has a built-in quality control mechanism that filters out qualified data, privacy protection measures for users, the ability to complete risk prediction tasks for several different diseases simultaneously, the function to hierarchically decompose the influence logic of each feature on the final conclusion, and a dedicated remote monitoring panel for medical teams allowing doctors to review all of a patient's monitoring data without needing to be present on-site.

RetinaFusion-XAI establishes a reliable non-invasive architecture that requires no invasive procedures to be performed on patients throughout the entire workflow. It can be used for early screening of related diseases in the general population, as well as for long-term continuous monitoring of a single patient's vascular health changes. Nevertheless, before this system can be formally integrated into hospitals' routine clinical work and deployed in settings such as outpatient departments and physical examination centers, multi-site prospective assessments must first be completed. Only after tracking and verifying its performance among asymptomatic populations across different sites and confirming its stable and reliable efficacy can it be officially put into clinical use.