News
Article
The use of retinal images can help investigators noninvasively identify chronic kidney disease and assess patient prognosis.
A new screening tool that can help clinicians diagnose chronic kidney disease (CKD) without the need for invasive renal biopsies. The developers of the model say it is highly accurate and flexible, making it feasible for use in a variety of settings. The study was published in the journal Nature Communications.1
The “gold standard” for pathological diagnosis of CKD is the renal biopsy, the authors noted. A biopsy can not only diagnose the disease itself, but can also help investigators understand the state of the patient’s disease and the extent of active and chronic lesions.
“However, renal biopsy is often limited by high-risk anatomical characteristics, systemic diseases with high mortality, and the lack of biopsy techniques and nephrology services in less developed regions,” the authors wrote.
In a primary care setting, the model could be used to identify potential cases of CKD, allowing for early referral to specialty clinics. | Image credit: filins - stock.adobe.com
A 2004 study of more than 750 adults who underwent renal biopsies at a single academic medical center found that 13% of patients experienced biopsy-related complications, about half of which were considered major.2 Patients with serum creatinine levels above 5.0 mg/dl were more than twice as likely to experience complications, the investigators found.
Recent efforts to solve the CKD diagnosis problem have focused on utilizing the proteome, microRNA, and molecular imaging—along with the help of artificial intelligence—to derive noninvasive CKD diagnostic models.
However, those efforts have still not been able to displace the renal biopsy as the definitive diagnostic tool.
In the new study, the investigators outlined their work developing the Kidney Intelligent Diagnosis System (KIDS). The model was trained on 13,144 retinal images from 6773 participants who underwent renal biopsy. Other clinical data was also incorporated into the model.
The authors noted that alterations in microvascular structure and function are often observable prior to organ damage. They added that microvascular dysfunction often mirrors dysfunction in visceral beds.
“As a convenient window to observe microvascular and neural tissues, the retina can be used for noninvasive detection, diagnosis, staging, and management of systemic diseases,” they explained.
Moreover, they said deep learning algorithms have been able to identify links between retinopathy severity and declining renal function, which in turn can help detect CKD.
In training and testing their model, the authors said the model was able to achieve an area under the receiver operating characteristic curve (AUC) of 0.839-0.993 for CKD screening. The model was also able to identify the 5 most common pathological types of CKD with an AUC of 0.790-0.932. They said their models were externally validated at multiple centers representing multiple patient ethnicities. In addition, the model was compared to the analysis of 12 nephrologists in a prospective data set. The KIDS model outperformed the nephrologists by 26.98%.
The model translates well to real-world settings, the authors explained. In a primary care setting, they said it can be used to identify potential cases of CKD, allowing for early referral to specialty clinics. In a specialty nephrology setting, the tool can provide objective pathological diagnosis and predict adverse outcomes without the need for renal biopsy. The tool can be especially helpful in under-resourced areas like Somalia, which was one of the locations used for external validation, the investigators said.
The authors identified a number of limitations to their research. Since the data were based on patients who underwent renal biopsy, the investigators said their sample size was limited. In addition, while the model was developed and validated using patients from China and Somalia, more research is needed to confirm whether it is equally accurate for patients of other ethnicities. The investigators said they are deploying the model on a cloud platform in order to expand its availability and validate its accuracy.
References
1. Wu Q, Li J, Zhao L, et al. A noninvasive model for chronic kidney disease screening and common pathological type identification from retinal images. Nat Commun. 2025;16(1):6962. doi:10.1038/s41467-025-62273-0
2. Whittier WL, Korbet SM. Timing of complications in percutaneous renal biopsy. J Am Soc Nephrol. 2004;15(1):142-147. doi:10.1097/01.asn.0000102472.37947.14
Stay ahead of policy, cost, and value—subscribe to AJMC for expert insights at the intersection of clinical care and health economics.