Explainable artificial intelligence may help identify early kidney involvement and support kidney-risk assessment in patients with type 2 diabetes, but current evidence is insufficient for broad routine deployment, according to a narrative review published in Diabetes, Metabolic Syndrome and Obesity. 

“The most defensible current role is as a complementary risk-stratification and audit layer within guideline-based care,” the authors wrote. 

Explainable artificial intelligence (XAI) uses methods that make model predictions more interpretable, such as identifying which clinical variables influence a patient’s estimated risk. The review examined whether these approaches could support kidney-risk assessment in patients with type 2 diabetes (T2D), focusing on which models are closest to practical use and where their clinical boundaries remain. 

Review approach 

The authors conducted structured searches of PubMed, Semantic Scholar, and OpenAlex, supplemented by citation tracing and full-text assessment. The detailed synthesis included 18 purposively selected, clinically relevant evidence records: 17 model- or deployment-oriented records and one systematic review and meta-analysis. Primary model evidence was concentrated in 2023 to 2026.  

Selection prioritized clinical relevance, endpoint clarity, explainability, validation, calibration, missing-data reporting, fairness, and workflow relevance. The authors organized findings into five clinical areas, assessing performance separately from readiness for use. 

1. Detecting occult, early, or existing kidney disease 

An occult diabetic kidney disease model targeted early microalbuminuria that could be missed when routine creatinine or estimated glomerular filtration rate appeared normal. The model used Shapley additive explanations (SHAP), a method that identifies how input variables contribute to predictions, to guide simplification to eight features. According to the review, the resulting logistic model achieved areas under the receiver operating characteristic curve (AUROC) of 0.824 in training and 0.786 in external validation. 

Another model defined early diabetic kidney disease (DKD) as a urinary albumin-to-creatinine ratio of 30 to 300 mg/g with preserved estimated glomerular filtration rate. Despite its explicit early-disease endpoint, the authors described its discrimination as modest and its sensitivity as low, raising concerns about its suitability as a screening tool. 

Other classifiers used kidney-status markers, including microalbuminuria, creatinine, and proteinuria. The authors cautioned that their high discrimination may partly reflect recognition of existing disease rather than prediction beyond standard assessment. 

A model using nonlaboratory inputs achieved moderate discrimination for albuminuria risk. The authors positioned it as a potential prompt for confirmatory kidney testing rather than a replacement for laboratory screening. 

2. Predicting future kidney disease 

The authors distinguished future-risk prediction from classification of current disease, emphasizing that predictors should precede kidney outcomes.  

NephraRisk predicted three-year incident or progressive DKD or diabetic nephropathy and reported an AUROC of 0.852. Its evaluation included calibration, decision-curve analysis, fairness checks, temporal validation, SHAP explanations, and a web application. However, the authors said its literature-informed imputation and kidney-status predictors required independent assessment in other populations. 

A Korean multicenter electronic health record study achieved external validation across hospitals, although logistic regression matched or exceeded XGBoost externally. A single-center retrospective Chinese study involving 12,190 adults with T2D found that LightGBM had the highest AUROC among the evaluated models for DKD prediction, but external validation remained necessary before clinical use. 

3. Forecasting progression and referral need 

The review examined models predicting worsening kidney function, end-stage kidney disease, rapid progression, and the need for nephrology referral. 

One model predicting one-year kidney function progression in patients with T2D and established chronic kidney disease (CKD) combined SHAP-guided simplification, temporal validation, and a web tool. 

The authors distinguished these applications from early detection. Such models may support closer monitoring, escalation of care, or specialist triage, but their performance does not establish that they can detect silent early DKD. 

4. Using retinal images and multimodal data 

Retinal models sought to capture microvascular and systemic risk signals that may not be fully represented by routine laboratory values. 

DeepDKD included large-scale training, external validation, and differentiation of DKD from non-DKD in biopsy-linked datasets. A separate multimodal model combined retinal fundus images with clinical variables to predict five-year CKD risk in patients with T2D. It reported explanation methods, calibration, decision-curve analysis, and external validation, but performance was lower in the external UK Biobank cohort than in South Korea. 

The latter model predicted CKD broadly rather than diabetes-specific kidney disease. The authors emphasized that CKD in patients with diabetes may have diabetic, hypertensive, vascular, or other causes. They also cautioned that retinal attention maps identify image regions influencing model predictions; they do not establish kidney pathology. 

5. Incorporating social determinants of health 

A model using clinical, behavioral, and social variables assessed self-reported CKD risk among adults with T2D using SHAP to explain predictions.  

The authors cautioned that self-reported CKD does not establish confirmed DKD. The cross-sectional survey design also could not establish causality. They said social variables should guide support and resource allocation. They also emphasized the need to assess subgroup calibration, false-negative rates, false-positive burden, and consequences for access to monitoring and referral. Strong overall discrimination alone does not establish fairness, they wrote. 

Evidence needed before routine use 

The authors noted that endpoints varied across the evidence base, external validation was uneven, and prospective workflow evidence was rare. Missing-data reporting was inconsistent, fairness and uncertainty reporting were limited, and explanations had not consistently been shown to improve decisions or outcomes. The authors emphasized that web applications and retrospective decision-curve analyses do not by themselves establish clinical benefit. 

They called for prospective validation, testing in clinical workflows, and assessment across clinically relevant subgroups, including age, sex, kidney function stage, albuminuria status, medication exposure, and access-related variables. They added that high-risk outputs should prompt confirmatory kidney assessment, clinical review, and shared decision-making rather than automated diagnosis or treatment. 

“XAI will matter clinically only if it helps diabetes and kidney-care teams make safer, more equitable, and more timely decisions in real practice,” the authors wrote. 

They disclosed using ChatGPT and OpenAI Codex for language editing, organization, formatting checks, and refinement of nonsubstantive wording. They reviewed and verified the outputs and take full responsibility for the final content. 

 The authors reported no other conflicts of interest.