A Multi-Task Deep Learning Framework for Concurrent Chronic Kidney Disease Staging and Hypertensive Retinopathy Grading from Retinal Fundus Images

Abstract

Oculomics represents a transformative paradigm in non-invasive diagnostics, leveraging deep learning to extract quantitative vascular and structural biomarkers from retinal fundus photographs for systemic health assessment. This study presents a Multi-Task Deep Learning Framework for concurrent Chronic Kidney Disease (CKD) staging and serological blood group typing from a single retinal fundus image. The proposed architecture employs an ImageNet-pretrained EfficientNetB0 as a shared feature-extraction backbone, followed by two parallel task-specific fully connected classification heads. The first head predicts CKD status, while the second estimates the patient’s blood group phenotype. Retinal images were resized, normalized to the [0,1] range, and augmented in real time to improve generalization and reduce overfitting. In the demonstrated results, the model predicted Normal kidney status with approximately 33.7% confidence and the O+ blood group with approximately 71.4% confidence. Class-wise probability distributions and gradient-based visualization maps provide additional interpretability by highlighting retinal regions contributing to predictions. Implementation challenges involving absolute path reconciliation, dataframe target structuring, label encoding, and multi-output tensor alignment were resolved to ensure stable gradient propagation. The proposed end-to-end pipeline demonstrates the technical feasibility of unified retinal image analysis for low-cost, non-invasive systemic screening and blood group prediction. Further validation using larger and clinically diverse datasets is required to establish diagnostic accuracy, robustness, and clinical utility.