AI Reads Retinal Images to Diagnose Diabetic Nephropathy Without Kidney Biopsies

For the hundreds of millions of people living with type 2 diabetes worldwide, one of the most consequential questions their doctors face is whether the kidneys are being damaged by diabetes itself or by some other renal disease. Answering that question correctly has long required an invasive kidney biopsy, a procedure that carries real risks and is unavailable in many settings. Now, a multicenter study published in the Chinese Medical Journal reports that an artificial intelligence model combining standard fundus photographs with routine clinical data can distinguish diabetic nephropathy from non-diabetic renal disease with high accuracy, and can also predict which patients are likely to progress toward end-stage renal disease.

The research, co-led by Dr. Sai Pan, Professor Xiaoniao Chen, Dr. Qian Wang, Dr. Guangyan Cai, and Professor Zheyi Dong, brought together nephrology and ophthalmology expertise from the Chinese PLA General Hospital and multiple cooperating medical institutions. The team enrolled a total of 397 patients with type 2 diabetes whose kidney lesions had been confirmed by biopsy, drawing participants from six hospitals across China. The population was divided into a primary cohort of 374 participants and an independent external validation cohort of 23 subjects, allowing the researchers to test their models both internally and on data from hospitals outside the primary dataset.

The biological rationale behind the approach rests on a striking anatomical fact: the retinal microvasculature is the only microcirculation in the human body that can be directly photographed and observed. Because the retina and the glomeruli, the filtering units of the kidney, are both susceptible to microvascular damage from long-term diabetes, morphological abnormalities visible in retinal vessels can mirror the microscopic injury occurring in kidney tissue.

Earlier studies in this field evaluated individual retinal vascular indicators in isolation, an approach the researchers argue ignores the internal correlations and synergistic mixed effects of multiple parameters acting together. To overcome that limitation, the team constructed an integrated machine learning analytical framework centered on weighted quantile sum regression, a statistical technique that allowed them to combine all retinal vascular geometric indicators into a single composite mixed-effect biomarker for comprehensive evaluation. Rather than asking which single vessel measurement matters most, the WQS index captures the joint signal across the full set of measurements.

Extracting those measurements automatically was itself a technical challenge. The researchers adopted unsupervised learning algorithms paired with ResNet deep neural networks to achieve automatic, labor-free retinal vessel segmentation, meaning the software could trace the vascular tree in each fundus image without manual annotation. From the segmented vessels they calculated a full set of core geometric metrics, including vessel diameter, fractal dimension, tortuosity density, and vascular branching angles. According to the study’s description of the work, retinal vascular geometric parameters differed significantly between patients with diabetic nephropathy and those with non-diabetic renal disease, providing the raw signal the diagnostic models would need.

To identify the best diagnostic model, the team compared seven mainstream machine learning algorithms. The top performer was a logistic regression model that integrated the WQS retinal composite index with seven vital clinical indicators. On internal test datasets it achieved an area under the receiver operating characteristic curve, or AUC, of 0.98 and an accuracy of 0.92. In cross-hospital external validation, the model maintained robust performance, recording an AUC of 0.95 and an accuracy of 0.91. Ablation experiments, in which components of the model are systematically removed to test their contribution, further demonstrated that incorporating retinal vascular parameters significantly improves overall diagnostic capacity compared with clinical data alone.

The prognostic arm of the study addressed a different question: which patients with confirmed diabetic nephropathy are most likely to experience rapid decline in renal function. For this, the researchers built a COX-Random Forest prognostic model based on 139 DN patients who were followed continuously until December 2023, with all-cause death or progression to end-stage renal disease as the critical clinical endpoints. That model reached an AUC of 0.88, effectively identifying patients at high risk of rapid renal function decline. SHAP interpretability analysis, a technique that quantifies each variable’s contribution to a model’s predictions, identified diabetes duration, diabetic retinopathy, and age as the three most influential predictive factors.

The study’s authors emphasized the practical implications of the pipeline’s design. Professor Chen explained that the WQS composite index maximizes the diagnostic potential of routine fundus images, while the unsupervised segmentation technology supports fully automatic image interpretation even in grassroots community hospitals, demonstrating the clinical value of fundus photographs in kidney disease diagnosis and management. First author Sai Pan stated that the multimodal AI pipeline breaks the limitation of single-index analysis and balances diagnostic accuracy with clinical accessibility, reflecting the advantages of artificial intelligence in integrating multidimensional medical data for disease screening.

Professor Dong added that the stable multicenter validation results verify the model’s reliability for large-scale clinical screening, laying a foundation, in his words, for its wide popularization and practical deployment in routine clinical practice in the future. The image analysis described in the study spans multicenter data collection, unsupervised deep learning retinal vessel segmentation, construction of the WQS mixed-effect index, and machine learning modeling for both DN diagnosis and ESRD risk prediction, with the combination of clinical data and retinal biomarkers greatly improving model performance over single-modality models.

The broader context underscores why a non-invasive alternative matters. Over 400 million people worldwide live with type 2 diabetes, and nearly half of these patients will develop chronic kidney disease over time. Diabetic nephropathy ranks as the leading trigger of irreversible end-stage renal disease, imposing heavy burdens on healthcare systems globally. The current gold standard for distinguishing DN from non-diabetic renal disease remains invasive renal biopsy, which brings unavoidable risks such as bleeding and secondary infection and is unavailable in remote and resource-limited medical facilities. Without pathological evidence, clinicians relying only on routine blood and urine tests can misdiagnose patients, and misclassification holds back precise targeted treatment, distorts disease prevalence data, and wastes limited medical resources, creating what the researchers describe as a pressing demand for safe, convenient, and scalable non-invasive diagnostic solutions.

Subject of Research: An AI model using retinal vascular geometric parameters from fundus images to diagnose diabetic nephropathy and predict renal outcomes without kidney biopsy

Article Title: AI analysis of retinal images: a novel non-invasive tool for diabetic nephropathy

Article References: AI analysis of retinal images: a novel non-invasive tool for diabetic nephropathy. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: diabetic nephropathy, artificial intelligence, retinal imaging, fundus photography, machine learning, kidney biopsy, end-stage renal disease, weighted quantile sum regression, deep learning, multicenter study, Chinese Medical Journal, non-invasive diagnosis