AI Advances in Kidney Disease: Retinal Imaging, Hospitalization Risk, and Lupus Nephritis

Artificial intelligence is increasingly being applied across kidney disease care, with new research demonstrating AI’s ability to predict diabetic kidney disease from retinal images, flag hemodialysis patients at imminent risk of hospitalization, and advance the diagnosis and treatment of lupus nephritis. The findings come from studies presented at medical meetings and published in peer-reviewed journals.

A retinal image–based deep learning system (DeepDKD Plus) accurately predicts both the incidence and progression of diabetic kidney disease (DKD), outperforming traditional clinical models, according to study results presented at the Association for Research in Vision and Ophthalmology (ARVO) 2026 annual meeting held in Denver, CO, May 3-7, 2026. The retrospective, multicenter study developed DeepDKD Plus to perform two tasks: predict 5-year DKD incidence — defined as decline in estimated glomerular filtration rate (eGFR) to <60 mL/min/1.73 m² or ACR >30 mg/g — in individuals with preserved renal function; and predict 5-year DKD progression — defined as ≥40% decline in eGFR or renal failure — in patients with preexisting kidney disease.

The model was trained on 11,116 retinal images from 5558 Chinese patients with diabetes, followed by internal validation and external validation across five multiethnic cohorts from China, Mexico, and the UK (n=13,935). For predicting 5-year DKD onset, DeepDKD Plus showed strong discrimination in internal testing with an area under the curve (AUC) of 0.814 (95% CI, 0.789–0.838), exceeding the performance of the clinical metadata model, which reached an AUC of 0.760 (95% CI, 0.733–0.786; P<.001). When retinal imaging was combined with clinical variables, performance improved further to an AUC of 0.828 (95% CI, 0.804-0.850). In external cohorts, the deep learning model maintained consistent accuracy, with AUCs ranging from 0.722 to 0.797, compared with 0.676 to 0.735 for the metadata-based approach. For predicting 5-year DKD progression, DeepDKD Plus achieved an internal AUC of 0.762 (95% CI, 0.732–0.788), again outperforming the metadata model (AUC 0.712; 95% CI, 0.684–0.743; P<.001), while the combined model reached an AUC of 0.797 (95% CI, 0.772–0.820). External validation results were similar, with DeepDKD Plus producing AUCs between 0.726 and 0.772, compared with 0.656 to 0.731 for the clinical model alone.

In a separate development, AI-driven machine learning models may identify patients on hemodialysis at risk for hospitalization for infection or fluid status-related abnormalities within 7 days. A large dialysis organization employed 2 imminent hospitalization prediction models (IHPMs) – 1 for fluid overload and 1 for infection – to identify patients who were likely to have a hospitalization after their in-center hemodialysis session. The models integrated dialysis electronic medical records and claims data, and each generated a risk score (range: 0-1), with a threshold risk score of 0.64 or above predicting immediate hospitalization. For high-scoring patients, case manager nurses reviewed the top 5 reasons for the score along with clinical data, which appeared on a dashboard. Based on the assessment, nurses could arrange for interventions, including fluid management (eg, reassessment of dry weight, extending or additional dialysis); medication review (eg, initiation, discontinuation, or adjustment of therapies such as antibiotics or diuretics); specialist referral; implementation of clinic treatment algorithms (eg, anemia management, nutritional supplementation, vascular access monitoring); and other supportive care (eg, referral for community resources, transportation, home health, behavioral health, or transplant evaluation). The investigators studied 10,294 patients supplying 83,928 risk scores. The odds of hospitalization were significantly decreased 8% with AI-driven intervention (OR 0.92; P=.025), including 7% with remote nurse-managed care and 9% with clinic-managed care. The findings were reported in NEJM Catalyst.

AI also shows promise in lupus nephritis (LN). A comprehensive review of the latest advancements in the diagnosis and treatment of LN, with a focus on the application of AI and machine learning technologies, was published in The Lancet. Renal biopsy is considered the gold standard diagnostic tool; however, because of the invasive nature of this procedure, only appropriate candidates should be considered. Observer variability stemming from human limitations is a concern, and digitization of renal biopsy analysis can enhance the accuracy of readings. When coupled with AI and ML, the minutiae of subgroup diagnosis may be identified, thereby enhancing the personalization of treatment pathways. The researchers referenced a novel spherical evolutionary algorithm, founded on a binary rough supervolume combined with grouped intelligent sampling (bRGSE), which achieved an accuracy level of 96.687% and sensitivity of 97.833% in discriminating between proliferative LN and membranous LN.

The review also covered targeted biologic treatments for LN, including biologics targeting B cells and plasma cells, T cells, cytokines, and the complement cascade. Treatment decisions could be informed by AI-based analysis of multiomic data, which integrates information from multiple biological levels, including genomics, transcriptomics, proteomics, and metabolomics. Two topics generating considerable buzz are targeted protein degradation and pharmacogenomics. Pharmacogenomics in the treatment of LN would be applied to understanding how genetic polymorphisms affect drug metabolism, efficacy, and the presentation of adverse reactions. The researchers conclude that precision medicine, which may one day be achieved through the application of AI and ML in the treatment of LN, holds great promise.

More broadly, artificial intelligence in nephrology spans a range of modeling approaches, according to a state-of-the-art review published in the International Journal of Molecular Sciences. For medical data recorded in tabular form (such as test results, age, and clinical parameters), models like logistic regression, random forests, and XGBoost perform very well. Intermediate solutions, such as the multilayer perceptron, combine the advantages of classical models with those of more complex methods. The most advanced models – deep neural networks – are used when the data is more complex, for example in medical image analysis, and can recognize structures and patterns regardless of their arrangement, which is particularly important in histopathological diagnostics. The most innovative direction involves combining artificial intelligence with modern biological analysis, such as proteomics or metabolomics, allowing for the detection of very early signs of disease before symptoms appear or changes become visible in standard tests. As emphasized by a professor at Wroclaw Medical University, "The greatest potential of these methods lies in their ability to analyze vast sets of biological data and identify patterns invisible in classical diagnostics. In practice, this means the possibility of earlier disease detection and better prediction of its course before irreversible kidney damage occurs."