Explainable Artificial Intelligence for Early Diabetic Kidney Disease Risk Profiling in Type 2 Diabetes: Translational Readiness and Clinical Boundaries

Abstract

Purpose

Early kidney-risk profiling in type 2 diabetes is difficult because albuminuria, estimated glomerular filtration rate, comorbidity, and care access capture different aspects of diabetic kidney disease. This review examines where explainable artificial intelligence may support clinically defensible kidney-risk profiling.

Methods

This narrative review used structured searches of PubMed, Semantic Scholar, and OpenAlex using kidney, diabetes, and artificial intelligence/explainability terms, supplemented by citation chasing and full-text assessment of key records. Broad query returns were documented for PubMed (600) and Semantic Scholar (956). Records were screened for topic fit and prioritized for clinical relevance to kidney-risk tasks, endpoint validity, explainability, validation, calibration, missing-data handling, fairness, and workflow relevance; 18 clinically relevant evidence records were retained for detailed synthesis in Supplementary Table 1.

Results

Evidence was organized by clinical task: occult or early diabetic kidney disease detection, incident kidney disease prediction, progression and referral forecasting, retinal and multimodal profiling, and equity-aware risk assessment. Current studies show promising discrimination and increasing use of explanation methods, calibration, decision-curve analysis, web tools, and external or temporal validation. Persistent limits include heterogeneous endpoints, diagnostic circularity, uneven external validation, sparse workflow evidence, and limited fairness or uncertainty reporting.

Conclusion

Explainable artificial intelligence is best viewed as a risk-stratification, communication, and audit layer within guideline-based care, not as a standalone diagnostic or treatment-decision system.

Introduction

DKD is one of the most common and consequential complications of T2D.^1,2 It remains a major cause of CKD, kidney failure, cardiovascular events, and premature mortality worldwide.³ Its burden also appears in healthcare use, multimorbidity, and long-term care complexity.⁴ Yet DKD is not a single linear phenotype. Disease-primer and consensus sources describe heterogeneous albuminuria/eGFR patterns and emphasize the need to consider alternative causes of kidney disease.^1,5 Current diabetes-kidney guidance is already clear: repeat eGFR and albuminuria assessment, risk-factor control, and kidney-protective therapy remain the foundation of care.^6,7 Broader CKD guidance adds risk-based staging and follow-up.⁸ The unresolved clinical problem is earlier and more individualized profiling of patients whose risk is not obvious from routine markers alone, including those with discordant albuminuria and eGFR, overlapping cardiometabolic disease, medication exposure, or uneven access to care.⁹

Pathophysiologically, DKD reflects interacting metabolic, hemodynamic, inflammatory, endothelial, oxidative, and fibrotic mechanisms.⁵ This heterogeneity helps explain discordant albuminuria and eGFR trajectories.^10–12 It also contributes to the clinical under-recognition of early disease.^13,14

Contemporary kidney-protective therapy is also changing the setting in which prediction models are developed and used. Expanding use of sodium-glucose cotransporter-2 inhibitors and glucagon-like peptide-1 receptor agonists may alter baseline risk and progression patterns.^6,7 Broader CKD guidance and nonsteroidal mineralocorticoid receptor antagonist evidence further underscore changing model transportability across treatment eras.^8,15

Clinicians therefore need more than a binary answer about whether DKD is present. In routine diabetes care, a more useful output is who needs closer kidney assessment, why an estimate appears high, whether the estimate is trustworthy locally, and what action should follow. A kidney-risk profile should point toward confirmatory UACR/eGFR testing, monitoring intensity, nephroprotective medication review, risk-factor optimization, patient communication, or referral, while treatment decisions remain anchored to diabetes-kidney consensus and CKD-diabetes guideline care.^7,15 For prediction models, this shifts attention from discrimination alone to endpoints, calibration, clinically sensible thresholds, interpretable risk drivers, and workflow fit.^16,17

AI and machine-learning models are useful candidates in this setting because they can combine signals that are difficult to integrate through simple clinical rules.¹⁸ Current studies in T2D and kidney disease have used routine laboratory variables, retinal images, and SDOH variables.^19,20 Others address incident DKD, multimodal CKD prediction, or deployment-oriented DKD risk prediction.^21,22 This range is clinically useful, but it also makes the literature easy to overread. The same AUROC has different implications when a model predicts incident disease, classifies current kidney involvement, or supports downstream escalation.

XAI is relevant because a kidney-risk estimate is unlikely to influence care if clinicians cannot inspect what drives it. SHAP, partial dependence plots, feature importance, Grad-CAM, and web-based risk interfaces can make influential variables or image regions visible.^19,22 Local explanation and SHAP-based attribution methods support this approach in tabular models, while Grad-CAM provides a route for image-region visualization.^23,24 These tools improve visibility, not validity. A SHAP ranking can clarify model behavior without proving biological causality, treatment benefit, calibration, transportability, or fairness.^25,26

Several recent reviews show that AI/ML for DKD prediction, diagnosis, and progression is active and often reports promising discrimination, but they also describe substantial heterogeneity and limited external validation.^27,28 Reviews of diabetic nephropathy prediction and diabetic microvascular complications show similar overlap.^29,30 This review does not add another broad algorithm catalogue. It asks which explainable models are closest to practical kidney-risk profiling, what features make them translationally more mature, and where their clinical boundaries remain.

This review takes a clinical task perspective. Evidence is grouped by task rather than by algorithm family, and model performance is interpreted separately from readiness for use. The aim is to help endocrinology, nephrology, and digital-health readers judge where XAI may support T2D kidney care and where it remains a research-stage tool.

Clinical Problem: Defining Early DKD Risk Profiling Before Discussing Algorithms

Why the Clinical Task Must Come First

The clinical value of AI in DKD depends first on the question being asked. In T2D care, kidney-risk models may search for unrecognized kidney involvement, predict future DKD, forecast rapid deterioration, triage nephrology referral, or estimate broad CKD risk. These tasks occur on the same care pathway but are not interchangeable. A model that separates advanced DKD from earlier DKD can describe severity, but it should not be presented as preclinical detection.³¹ Similarly, ESRD or referral prediction may support escalation planning without solving early DKD screening.^32,33

For this reason, this review uses “early DKD risk profiling” as a task-oriented concept rather than a single endpoint label. The most direct evidence comes from models that explicitly target early or occult kidney involvement, such as occult DKD screening with routine variables or early DKD defined by UACR 30–300 mg/g with preserved eGFR.^19,34 Incident-risk models at or near T2D diagnosis are also relevant because predictors precede future kidney outcomes.²¹ NephraRisk and multicenter EHR DKD occurrence prediction strengthen the incident/progressive layer, although both require local validation and workflow testing.^22,35 Broad CKD prediction models in T2D can inform risk profiling, especially when retinal, clinical, or SDOH inputs are used, but they should not be read as DKD-specific diagnosis unless the endpoint supports that claim.^20,36

DKD, CKD, and the Etiology Problem

A persistent problem is the use of CKD in T2D as a practical surrogate for DKD. The shortcut is understandable in routine data but clinically fragile. DKD nomenclature and consensus sources emphasize albuminuria, eGFR, and evaluation for alternative causes.^1,37 Many adults with T2D have hypertensive, vascular, age-related, inflammatory, medication-related, primary renal, or mixed contributors to CKD, and US population data show heterogeneous kidney manifestations in diabetes.^10,38 Renal insufficiency can also occur without albuminuria or retinopathy, and nonalbuminuric renal-impairment studies reinforce the endpoint problem.^11,12 RIACE and nonalbuminuric progression analyses show that this phenotype is clinically meaningful.^13,14 A model trained on CKD codes or reduced eGFR may therefore learn kidney impairment in a diabetes population without distinguishing DKD from non-diabetic kidney disease. The distinction matters because the next step may be monitoring, albuminuria confirmation, medication review, nephrology referral, or evaluation for a non-diabetic cause.

The distinction is especially important for retinal and multimodal models. Retinal fundus images may contain microvascular information relevant to systemic vascular injury, and recent retinal AI studies show promising discrimination for DKD or CKD-related outcomes.^36,39 Prior retinal deep-learning and kidney-eye literature support the broader idea that fundus images can encode systemic risk signals, including CKD, T2D, and cardiovascular-risk information.^40,41 Still, a retinal heatmap or multimodal CKD probability does not establish diabetic kidney etiology. In the Diabetes Care multimodal study, the model predicted five-year CKD risk in T2D and used both SHAP and Grad-CAM, but the endpoint remained broad CKD and did not differentiate DKD from non-DKD.³⁶ DeepDKD moved closer to this boundary by addressing DKD detection and DN vs non-DKD differentiation in biopsy-linked datasets, although Grad-CAM saliency should still be read as model attention rather than confirmed pathology.³⁹

Predictor Timing and Diagnostic Circularity

Early risk profiling also depends on when predictors are measured and what they already encode. Long-term T2D cohort evidence and DKD risk-factor reviews show that renal-risk factors can be identified before or during early disease, but baseline predictors still need to be separated from concurrent disease-defining variables.^42,43 In many high-performing tabular models, influential variables include eGFR, creatinine, urea, UACR, proteinuria, microalbuminuria, diabetic retinopathy, or diabetes duration.^44,45 These features are clinically meaningful, but they may already carry part of the outcome.

Such models can still be useful. They may organize risk, combine multimorbidity patterns, prioritize follow-up, or support transparent communication. The key issue is framing. A model built from baseline features at initial T2D diagnosis and followed for incident DKD answers a different question from a model classifying current DKD status with concurrent renal laboratory data.²¹ The same distinction applies to WRF, established-CKD progression, or referral models.^46,47

What XAI Adds to the Clinical Problem

Explainability helps because kidney-risk predictions need to be auditable in clinical terms. Clinicians need to see whether a high-risk estimate is driven by kidney-function markers, glycemic control, blood pressure, retinal information, vascular comorbidity, low-burden hemodynamic proxies, or social vulnerability. SHAP-based models can expose variables influential for global model behavior and patient-level predictions, while partial dependence plots summarize average feature-response patterns.^19,48 Retinal XAI adds spatial explanation through Grad-CAM or related saliency approaches, and retinal vascular-parameter models offer more directly interpretable image-derived measurements.^39,49 Healthcare-XAI literature also warns that explanation quality depends on the clinical question, user, and validation context.^25,50

Explanation is one part of readiness, not readiness itself. A model can be transparent and still be poorly calibrated, externally fragile, insensitive at clinically relevant thresholds, or awkward in workflow. For early DKD, sensitivity and alert burden matter because a screening-support tool must avoid missing patients who need confirmatory kidney assessment while not generating unmanageable false positives. The routine-variable early-DKD model illustrates one position on this spectrum.³⁴ The occult-DKD model, NephraRisk, and the Diabetes Care multimodal model illustrate others.^19,22,36 Table 1 summarizes the terminology boundaries needed to interpret this evidence.

Table 1.

Terminology Boundaries for XAI-Based Kidney-Risk Profiling in T2D

Term or Concept	Clinical Boundary/Interpretation
DKD, DN, and CKD in T2D	DKD refers to kidney disease occurring in diabetes, commonly identified through albuminuria, reduced eGFR, or both, after alternative causes are considered.1,37 DN is often used as an overlapping term and may imply more diabetes-attributable pathology.51 CKD in T2D is broader than DKD because kidney impairment may be diabetic, hypertensive, vascular, obstructive, medication-related, or mixed.
Early DKD and occult DKD	Early DKD should be reserved for explicitly defined early kidney involvement, such as microalbuminuria or abnormal urinary albumin excretion with preserved eGFR. Occult DKD describes kidney involvement that may be missed by conventional presentation or routine markers alone.
Non-proteinuric/nonalbuminuric progression and downstream outcomes	Albuminuria and eGFR trajectories should not be assumed to move together. WRF, ESRD/ESKD, and nephrology referral are downstream outcomes and should not be treated as synonyms for early DKD.10,14
Risk profiling	Risk profiling means estimating and explaining kidney-risk patterns to support monitoring intensity, communication, prevention-oriented treatment review, referral timing, or trial enrichment; it is not diagnosis.
XAI	XAI refers to methods that make model predictions more interpretable or auditable, including SHAP, partial dependence plots, feature importance, and Grad-CAM.24,52 XAI explains model behavior, but does not by itself establish causality, fairness, clinical utility, or treatment benefit.25,26
Low-burden risk enrichment	Low-burden risk enrichment uses non-invasive or easily measured variables to identify patients who may warrant confirmatory kidney assessment.
Translational readiness	Translational readiness means that a model has moved beyond retrospective performance reporting toward validated, calibrated, useful, fair, monitored, and workflow-compatible clinical decision support.53,54

Working Definition for This Review

In this review, explainable AI for early DKD risk profiling refers to AI or machine-learning approaches that estimate kidney-risk patterns in people with T2D and provide interpretable outputs that could plausibly support earlier monitoring, confirmatory testing, risk communication, referral decisions, or prevention-oriented treatment review. The scope includes direct early or occult DKD models, incident DKD models, broad CKD risk models in T2D, retinal and multimodal models, low-burden albuminuria-risk enrichment, progression and referral models, and SDOH-aware CKD risk profiling. The review treats these as related but different tasks within a kidney-risk continuum. Figure 1 maps these related tasks across the kidney-risk continuum and links each task to plausible XAI-enabled outputs, including confirmatory UACR/eGFR testing, monitoring intensity, medication eligibility review, risk-factor optimization, and referral consideration.

Figure 1.

Clinical task map for XAI kidney-risk profiling in T2D. XAI kidney-risk evidence spans T2D without known kidney disease, occult or early DKD, incident DKD or CKD, established CKD/DKD, rapid progression or WRF, and downstream outcomes such as ESRD/ESKD or nephrology referral. Endpoint timing determines clinical meaning: occult/early models support screening enrichment and confirmatory UACR/eGFR testing, incident-risk models support preventive monitoring and risk-factor review, retinal and multimodal models provide non-invasive microvascular-risk signals, progression/referral models support escalation and specialist triage, and SDOH-aware models support equity-aware outreach. Tabular clinical/laboratory XAI, retinal or multimodal AI, and SDOH-aware profiling may contribute to actionable outputs including confirmatory testing, monitoring intensity, medication eligibility review, risk-factor optimization, and referral consideration.

Evidence Identification and Review Approach

Search Sources and Search Strategy

Literature was identified through structured searches of PubMed, Semantic Scholar, and OpenAlex, supplemented by targeted DOI/PMID follow-up, citation chasing, and full-text assessment of key records. The PubMed search combined kidney, diabetes, and artificial-intelligence concept blocks with Boolean operators: (“type 2 diabetes” [tiab] OR “diabetic kidney disease” [tiab] OR “diabetic nephropathy” [tiab]) AND (“machine learning” [tiab] OR “artificial intelligence” [tiab] OR “deep learning” [tiab] OR “explainable” [tiab] OR “interpretable” [tiab] OR “SHAP” [tiab]) AND (“kidney” [tiab] OR “renal” [tiab] OR “nephropathy” [tiab]). Broad query returns were documented as PubMed n=600 and Semantic Scholar n=956. These returns document search scope. Records were selected purposively, consistent with the narrative-review design, according to clinical task relevance, endpoint clarity, explainability, validation, calibration, missing-data reporting, fairness signals, and translational usefulness. OpenAlex was used to cross-check coverage and support citation chasing.

Search results were supplemented by citation chasing from recent reviews, guidelines, prediction-model reporting sources, and the reference lists of clinically relevant model papers. Review overlap was assessed to clarify novelty: existing reviews and meta-analyses cover AI and machine learning for DKD prediction, diagnosis, biomarkers, or progression.^27,28 Reviews of diabetic nephropathy risk prediction and diabetic microvascular complications add related overlap.^29,30 A recent narrative review similarly surveys the broader DKD AI landscape.⁵⁵ However, these sources generally do not center XAI for early kidney-risk profiling in T2D through translational readiness and clinical boundaries.

Eligibility Criteria and Study Selection

Eligible evidence consisted of peer-reviewed, published, English-language studies relevant to XAI or machine-learning kidney-risk tasks in T2D. Clinical tasks included occult or early DKD detection, incident DKD or CKD prediction, progression or referral forecasting, retinal and multimodal profiling, and equity-aware or SDOH risk assessment. Unpublished studies, non-peer-reviewed preprints, non-English records, articles outside the kidney-risk task, and reports with insufficient information for clinical interpretation were not used for detailed synthesis. Primary model evidence in the detailed synthesis was concentrated in 2023–2026, while earlier clinical, pathophysiological, guideline, and reporting-framework references were retained where needed for background and appraisal.

Study selection was purposive and clinically oriented, consistent with the narrative-review design. Records from the broad retrieval set were first screened for topic fit, language, peer-reviewed publication status, and relevance to kidney-risk tasks in T2D. They were then prioritized when they provided sufficient clinical or translational detail on task definition, endpoint validity, interpretability, validation, calibration, decision-curve analysis, missing-data handling, fairness, workflow fit, or deployment boundaries. Supplementary Table 1 presents this focused clinical evidence matrix and summarizes 18 clinically relevant evidence records selected for detailed synthesis, including 17 model or deployment-oriented records and one recent systematic review/meta-analysis used for field-level context.

Evidence Appraisal and Scope Boundaries

Evidence appraisal followed a framework-informed narrative approach. Interpretation drew on MI-CLAIM for clinical AI modeling transparency.⁵³ It also used TRIPOD and TRIPOD+AI for prediction-model reporting.^16,56 Risk-of-bias and applicability judgments were informed by PROBAST and PROBAST+AI.^57,58 Clinical evaluation and trial-reporting considerations drew on DECIDE-AI, CONSORT-AI, and SPIRIT-AI.^59–61 Implementation governance also drew on FUTURE-AI.⁶² These domains included task definition, endpoint validity, sample and event adequacy, predictor timing, missingness, calibration, threshold utility, external validation, explainability, fairness, workflow evidence, monitoring, and governance. These frameworks structured the clinical interpretation of task validity, reporting quality, translational readiness, and evidence boundaries across the selected records.

To keep interpretation conservative, performance claims were limited to studies with sufficient detail on performance, validation, calibration, and explanation methods. Preliminary or incompletely reported studies were used for context only. DKD-specific clinical claims were anchored to diabetes and kidney evidence, and synthesis-level interpretations indicate translational plausibility rather than established clinical utility.

Conceptual Foundations: From Prediction to Explainable Risk Profiling

Machine-learning prediction and clinical risk profiling are related but distinct. Prediction estimates the probability of an outcome. Risk profiling places that estimate in a clinical setting where the result can be interpreted and, when appropriate, acted on. Prediction-model literature emphasizes that performance, calibration, and intended use should be evaluated together rather than inferred from a single discrimination metric.^16,17 In early DKD care, risk profiling should help decide who may need confirmatory UACR/eGFR testing, closer follow-up, medication review, or referral, and why the model suggests that action. Established kidney-failure risk equations show how prediction becomes more clinically useful when endpoints, time horizons, and external validation are explicit.^63,64

Most tabular DKD-XAI studies use SHAP or feature importance to identify variables influential for the model. SHAP can provide global explanations across a cohort and local explanations for an individual prediction.⁵² Partial dependence plots and local explanation methods can show or approximate how features contribute to predictions.²³ These tools improve transparency, but they explain model behavior rather than disease mechanism. That distinction is important when highly ranked features are kidney-status markers, diabetes duration, or treatment-related variables.

Retinal and multimodal studies add image-based explanation. Grad-CAM and related saliency methods can highlight retinal regions that influenced model predictions, while SHAP can be used for clinical variables in multimodal systems.^36,39 Retinal deep-learning studies outside DKD-specific XAI show that fundus photographs can support detection or prediction of diabetic retinopathy, CKD, T2D, or cardiovascular-risk factors, making retinal kidney-risk profiling biologically and informatically plausible.^65,66 Grad-CAM was developed for visual explanation of deep networks, but medical-XAI reviews caution that saliency requires task-specific interpretation.^24,50 Measured retinal vascular parameters may offer an intermediate path between opaque image embeddings and routine tabular features because they are more directly interpretable as vascular measurements.⁴⁹ Image saliency and vascular-feature importance still need biological and clinical validation.

Calibration and decision curve analysis connect prediction to clinical use more directly than discrimination alone. Calibration asks whether predicted probabilities match observed risk, while DCA estimates whether a model offers net benefit across thresholds. Several higher-readiness studies report calibration and DCA, but these analyses remain intermediate evidence unless paired with prospective workflow or outcome evaluation.^19,22 The Diabetes Care multimodal model reported calibration and DCA alongside SHAP and Grad-CAM, strengthening appraisal without proving deployment readiness.³⁶

XAI appraisal should separate explanation from calibration, utility, and fairness. Current DKD-XAI studies often address interpretability, sometimes calibration or DCA, and less often fairness, uncertainty, workflow effect, or post-deployment monitoring.^67,68 Table 2 aligns explanation and appraisal methods with the clinical questions they can and cannot answer.

Table 2.

XAI Methods and Clinical Questions in DKD Risk Profiling

XAI or Appraisal Method	Typical Output	Clinical Question It Can Help Answer	Main Limitation
SHAP	Global and local feature attributions	Which variables drove model risk estimates?	Explains model behavior, not causality
Partial dependence plots	Average feature-response curves	How does predicted risk change across a feature range?	Can obscure interactions and subgroup heterogeneity
Feature importance	Ranked predictors	Which variables dominate prediction?	Method-dependent and correlation-sensitive
Grad-CAM/saliency	Image regions influencing predictions	Which retinal areas affected the output?	Not pathologic proof and may be device/population sensitive
Calibration analysis	Agreement between predicted and observed risk	Are predicted probabilities trustworthy?	Requires local validation and recalibration
Decision curve analysis	Net benefit across thresholds	Is there potential utility at plausible thresholds?	Retrospective DCA does not prove workflow benefit
Web calculators/apps	Clinician-accessible risk interface	Can the model be practically inspected?	Access does not equal deployed decision support
Fairness/subgroup analysis	Performance differences across groups	Are errors distributed equitably?	Requires meaningful subgroup and outcome definitions
Uncertainty estimation	Confidence, abstention, or out-of-distribution warning	When should clinicians distrust the output?	Rarely reported in current DKD-XAI studies

Evidence by Clinical Task

The evidence base is clearer when organized by clinical task. XAI models have been used to identify occult or early kidney involvement, classify current DKD or DN, predict incident DKD or CKD, forecast progression or referral need, analyze retinal or multimodal signals, and incorporate SDOH into kidney-risk profiling. These applications differ in timing, endpoint validity, input availability, and actionability. A task-based synthesis therefore gives a more realistic picture of readiness than a model-by-model inventory. Key characteristics of the clinically relevant studies–including population, endpoint definition, input modality, model type, explainability method, validation status, calibration or decision-curve reporting, missing-data handling, and fairness or deployment signals–are summarized in Supplementary Table 1.

Detecting Occult, Early, or Existing Kidney Disease

The closest evidence for early DKD risk profiling comes from models that explicitly target occult DKD, early DKD, albuminuria-based risk, or current DKD/DN classification. The occult-DKD clinlabomics model is the most conceptually aligned example: it targeted early microalbuminuria that may be missed when routine creatinine or eGFR appears normal, used SHAP-guided simplification, externally validated an 8-feature logistic model, and reported training and external AUCs of 0.824 and 0.786.¹⁹

Endpoint clarity is the main strength of the routine-variable early-DKD model, which defined early DKD as UACR 30–300 mg/g with preserved eGFR. Its modest AUC and low sensitivity are a useful warning: a clinically well-aligned endpoint is necessary, but it does not guarantee a clinically adequate screening tool.³⁴

High-performing tabular DKD/DN classifiers show how routine clinical data can be converted into transparent risk tools, but they also expose the circularity problem. When influential predictors include microalbuminuria, eGFR, creatinine, proteinuria, or related renal markers, high discrimination may partly reflect recognition of existing kidney disease rather than prediction beyond standard assessment.^44,45

Two additional studies broaden this cluster in different directions. A metabolite-enriched SHAP model supports biomarker-oriented XAI, although internal validation and renal-status variables limit its clinical interpretation.⁶⁹ A low-burden SHAP/eCO model deliberately used non-laboratory inputs and achieved moderate albuminuria-risk discrimination; its better role is to prompt confirmatory UACR/eGFR testing rather than to replace laboratory screening.⁴⁸

Taken together, this group shows that XAI can make early or current kidney-risk classification more transparent. The strongest clinical signal comes from clear early or occult endpoints, but sensitivity, retrospective design, and predictor circularity still limit translation.

Predicting Incident DKD or CKD After T2D Diagnosis

Incident-risk prediction is closer to prevention because predictors are measured before future kidney events. The SuperLearner model at initial T2D diagnosis is important for that reason, even though its holdout AUC was modest and validation remained internal.²¹

Multimodal incident-risk work adds a stronger translational signal. The Diabetes Care study combined fundus images with clinical variables for five-year CKD prediction in T2D and reported SHAP, Grad-CAM, calibration, DCA, and external validation. Its main limitation for this review is endpoint scope: the outcome was broad CKD rather than DKD-specific disease, and external performance dropped in UK Biobank.³⁶

NephraRisk is the clearest current example of deployment-oriented reporting in this cluster. It predicted three-year incident or progressive DKD/DN and reported AUROC 0.852, calibration, Brier score, DCA, fairness checks, SHAP explanation, temporal validation, and a Streamlit application.²²

NephraRisk also shows why readiness claims need independent testing: literature-informed imputation and kidney-status predictors require audit in other populations.²² The Korean multicenter EHR model adds a complementary lesson. It achieved external validation across hospitals, but logistic regression matched or exceeded XGBoost externally, showing that model complexity is not the same as clinical maturity.³⁵ Recent Chinese EHR data extend the evidence geographically: in a single-center cohort of 12,190 adults with T2D, LightGBM achieved the highest AUROC for DKD prediction, but the study used retrospective data and requires external validation before clinical use.⁷⁰

Overall, incident-risk evidence is promising but uneven. Models that preserve temporal separation and include external validation are closer to clinical use than cross-sectional classifiers, but endpoint review, local calibration, threshold evaluation, and workflow testing remain necessary before they guide monitoring, medication review, or referral.

Forecasting Progression, Hard Renal Outcomes, and Referral Need

Progression and hard-outcome models sit later in the kidney-care pathway. They do not solve early DKD detection, but they can support escalation, monitoring intensity, and specialist referral. The cardiometabolic index/WRF study is better interpreted as metabolic trajectory profiling for progression risk than as early DKD diagnosis.⁴⁶

ESRD and referral models are clinically useful downstream. An ESRD model after newly diagnosed T2D demonstrates long-horizon hard-outcome prediction, while referral models address a practical workflow problem; neither directly answers whether a patient has silent early DKD.^32,33

The one-year CKD progression model in T2D with established CKD is useful for readiness discussion because it combines a short horizon, SHAP-guided simplification, temporal validation, and a web tool. Its population already had CKD, so it should not be used as evidence for early DKD detection.⁴⁷

These studies define the downstream end of XAI kidney-risk profiling. Their contribution is to show how explainable models might support the kidney-care continuum while making clear why endpoint timing changes the clinical meaning of a result.

Retinal and Multimodal Risk Profiling

Retinal and multimodal models are important because they may capture microvascular and systemic risk signals not fully represented by routine laboratory values. The retinopathy-nephropathy literature and kidney-eye disease literature support an association between retinal and renal microvascular disease, but they do not prove that image-based AI identifies DKD pathology.^71,72 DeepDKD provides the strongest retinal evidence, with large-scale training, external validation, and biopsy-linked differentiation of DKD from non-DKD.³⁹

The Diabetes Care multimodal model supports retinal-plus-clinical fusion, while the retinal vascular-parameter study offers a more interpretable route through measured vascular features.^36,49 Retinal CKD-screening, retinal CKD/T2D prediction, retinal vascular-sign, and prospective microvascular studies support this pathway, although they are not DKD-specific proof.^73–75

Retinal and multimodal AI should be framed as a non-invasive risk-profiling route, especially where retinal imaging already exists through diabetic retinopathy screening programs.⁷⁶ The boundary is practical: image explanations are not pathology proof. Grad-CAM heatmaps show model attention, not confirmed diabetic kidney pathology, and retinal CKD prediction in T2D should not be overread as diabetes-attributable kidney disease unless the endpoint and validation support that claim.

Equity and SDOH-Aware Kidney-Risk Profiling

Most DKD-XAI studies focus on biomarkers, routine laboratory variables, EHR features, or imaging. The BRFSS SDOH model is therefore important because it uses clinical, behavioral, and social variables to profile self-reported CKD risk among adults with T2D, with SHAP explanations at global and local levels.²⁰

Its limitations are substantial. Self-reported CKD is not DKD confirmed by UACR/eGFR or adjudication, and cross-sectional survey data cannot establish causality. SDOH variables are better used to guide equity-aware support and resource allocation than to assign blame to patients.²⁰

Equity-aware risk profiling should therefore be part of DKD-XAI rather than an afterthought. If these tools prioritize monitoring or referral, developers and health systems need to monitor subgroup calibration, false-negative rates, alert burden, and access consequences because kidney health is shaped by social determinants and socioeconomic gradients before any model is deployed.^77,78

Section Synthesis

Across tasks, XAI models can combine heterogeneous data, identify model-influential drivers, and support more transparent risk communication. The strongest current signals come from occult/early models and deployment-oriented tools such as NephraRisk.^19,22 Retinal and multimodal studies add a complementary non-invasive signal.^36,39

The limitations are consistent enough to shape the field’s next steps. Endpoints vary widely, kidney-status predictors can inflate apparent performance, external validation is uneven, prospective workflow evidence is rare, and explanation methods usually clarify model behavior rather than disease mechanism. The CMRO systematic review/meta-analysis reinforces this caution: AI/ML studies for DKD show promising pooled discrimination but substantial heterogeneity and external validation in only a minority of studies.²⁷

Missing-data handling was uneven across the studies reviewed. Some studies reported imputation or design choices that improved feasibility, such as literature-informed Bayesian MICE in NephraRisk or exclusion of high-missingness variables in the Korean multicenter model, whereas many reports provided limited detail on missingness patterns, complete-case proportions, sensitivity analyses, or whether missingness would trigger warnings at deployment.^22,35,53 Clinically, this matters because unavailable UACR, medication, retinal, duration-of-diabetes, or social-risk data can change both the risk estimate and the population to which the model applies.

For clinical practice, the evidence supports XAI as a risk-profiling, triage, audit, and communication aid when aligned with a clear task. It does not support replacing UACR/eGFR assessment, diagnostic evaluation, or clinician judgment. The decisive question is whether a model is ready for the specific decision it claims to support.

Translational Readiness Framework

The evidence is moving beyond proof-of-concept development, but routine clinical use remains uncommon. Translational readiness is best treated as a sequence of clinical questions rather than as a property inferred from AUROC, model novelty, or the presence of SHAP plots. Clinical AI implementation literature makes the same point: technical performance has to connect to workflow, documentation, monitoring, and impact.^53,54 In this review, a model moves up the readiness ladder when it defines a usable task, uses clinically valid endpoints, measures predictors before the decision, handles missingness transparently, reports calibration and threshold behavior, explains outputs in a way users can inspect, validates across settings, monitors fairness and uncertainty, and is tested in workflow. Table 3 provides a translational readiness checklist that operationalizes these requirements and highlights common failure modes in the current evidence. Figure 2 summarizes this progression.

Table 3.

Translational Readiness Checklist for XAI Kidney-Risk Models

Readiness Domain	Minimum Expectation Before Clinical Use	Common Failure Mode in Current Evidence
Clinical task and endpoint	Define the user, decision point, action, endpoint, and time horizon; distinguish DKD, CKD, WRF, ESRD, referral, and progression.	DKD, CKD, WRF, ESRD, referral, and progression are pooled.
Predictor timing and availability	Ensure inputs are available before the decision and before the outcome; state whether kidney-status variables are predictors, endpoints, or both.	Concurrent kidney markers classify existing disease.
Missing data	Report variable-level missingness, complete-case proportions, imputation method, sensitivity analyses, and whether missingness warnings are shown at use.	Missing UACR, medication, duration, retinal, or SDOH data are ignored.
Discrimination	Report AUROC or C-statistic with confidence intervals and threshold-specific sensitivity and specificity where relevant.	High AUROC is treated as sufficient.
Calibration	Report calibration plots and at least calibration-in-the-large, calibration slope, Brier score, or observed/expected risk by strata.	Predicted probabilities are not checked.
Threshold utility	Report sensitivity, specificity, PPV, NPV, decision-curve analysis, and expected alert burden at clinically plausible thresholds.	No decision threshold or clinical action is defined.
Explainability	Specify global and local explanation methods, intended users, stability or plausibility checks, and warnings that attribution is not causation.	SHAP or Grad-CAM is displayed without user interpretation.
External or temporal validation	Demonstrate validation in at least one external, temporal, device-diverse, or population-diverse cohort and report any recalibration.	Random split validation is treated as deployment evidence.
Workflow, fairness, uncertainty, and governance	Before clinical use, test EHR/pathway integration, usability, decision impact, subgroup calibration and errors, uncertainty warnings, drift monitoring, recalibration triggers, and accountability.	Web calculator or app is treated as implementation.

Figure 2.

Translational readiness ladder for XAI kidney-risk models. Readiness progresses from clinical task definition to endpoint validity, predictor timing, missing-data handling, discrimination, calibration, threshold utility, explainability, external validation, workflow testing, fairness and uncertainty assessment, and governance. The ladder emphasizes measurable pre-use criteria, including calibration reporting, threshold-specific operating characteristics, decision-curve analysis, external or temporal validation, workflow evaluation, subgroup fairness assessment, uncertainty warnings, drift monitoring, recalibration triggers, and accountability. Current XAI kidney-risk models are closest to clinical use when they function as complementary risk-stratification and audit tools within guideline-based care.

Task, Endpoint, and Intended User

From a translational perspective, readiness begins with a task that can be acted on. Occult DKD detection, early-DKD classification, incident DKD prediction, broad CKD prediction in T2D, progression forecasting, ESRD prediction, referral triage, retinal AI, and SDOH-aware CKD profiling sit on the same care continuum, but they serve different users and decisions. A model designed to prioritize UACR testing in newly diagnosed T2D does different work from a model that separates advanced DKD from earlier DKD among patients already labeled with kidney disease.

The intended user and action should be explicit. A risk score is more clinically meaningful when it points toward repeat UACR/eGFR testing, medication eligibility review, intensified risk-factor management, closer monitoring, referral, or evaluation for non-diabetic kidney disease, all of which remain anchored to guideline-based care.^6–8 Without this action link, even a transparent explanation remains informative rather than translationally useful.

Predictor Availability, Temporality, and Missingness

A model is hard to translate if its predictors are unavailable when the decision is made. Routine laboratory and clinical variables may be accessible through EHR systems or diabetes visits; retinal images are feasible where screening infrastructure exists; and SDOH variables depend on survey, EHR, or community-data capture.^20,39 Timing then determines meaning. When eGFR, creatinine, UACR, proteinuria, or microalbuminuria are measured at the same time as the outcome, the model may be classifying current kidney status rather than predicting future risk.^19,22

Missing data are not a minor technical detail in DKD. Reporting guidance asks authors to describe inputs, missingness, preprocessing, and evaluation context.⁵³ NephraRisk used Bayesian MICE with literature-informed external priors, which improved clinical completeness but still requires independent testing across populations and care settings.²² The Korean multicenter model took a different route by excluding UACR and other high-missingness variables, improving feasibility while leaving albuminuria-related risk less directly represented.³⁵

Performance, Calibration, and Explanation Quality

Discrimination is necessary, but it cannot carry the clinical argument. AUROC describes ranking performance; it does not show whether predicted probabilities are accurate, whether a threshold is useful, or whether a decision improves care.^17,79,80 Higher-readiness studies therefore report more than AUROC. The occult-DKD model, NephraRisk, and related deployment-oriented work include calibration or decision curve analysis, making them more informative for translational appraisal than studies limited to internal discrimination.^19,22

Threshold behavior is equally practical. Sensitivity, specificity, PPV, NPV, and alert burden need to be reported at cutoffs a clinic could plausibly use. The routine-variable early-DKD model paired a clear early-DKD endpoint with low sensitivity, whereas the CMPB non-invasive model showed higher sensitivity but modest overall discrimination; both fit better as triage evidence than as diagnostic evidence.^34,48 Explanation quality should be judged in the same pragmatic way. SHAP, feature importance, Grad-CAM, or partial dependence plots are clinically useful only if the explanation helps the intended user understand the risk estimate without overstating mechanism or causality.^23,24,52

External Validation, Workflow Fit, and Prospective Evidence

External validation is one of the clearest readiness dividers. Kidney-failure risk-equation work shows how strongly transportability can shape clinical use.^63,64 Several DKD-XAI studies include independent or external cohorts, including the occult-DKD model, DeepDKD, and the Diabetes Care multimodal model.^19,39 NephraRisk and the Korean multicenter EHR model add more recent readiness-oriented examples.^22,35 Even then, external validation is not a final guarantee: the Diabetes Care model performed well in South Korea but dropped in UK Biobank, illustrating the importance of device, population, and care-system differences.³⁶

Workflow evidence remains limited. Web calculators, apps, DCA, prospective proof-of-concept screening, and threshold analyses are useful signals, but an interface is not the same as deployed clinical decision support.^19,22 A working clinical system still needs EHR or pathway integration, usability testing, reliable data pipelines, alert governance, clinician training, accountability, and evidence that decisions improve.^53,54 Model cards can document intended use and limits, but they cannot substitute for local implementation evidence.⁸¹

Fairness, Uncertainty, Monitoring, and Governance

Fairness belongs in readiness assessment because kidney risk is shaped before any algorithm is introduced. Socioeconomic status, race and ethnicity, geography, access to care, medication access, diet, comorbidities, and documentation patterns all influence T2D and CKD trajectories. CKD disparity literature shows how social determinants and racial inequities affect kidney health upstream of prediction models.^77,82 A model with strong overall AUROC may still perform poorly in groups with delayed diagnosis, incomplete albuminuria testing, different retinal-imaging access, or different treatment exposure.^83,84

The SDOH model illustrates one way to incorporate structural variables, although its self-reported CKD endpoint and internal validation limit direct clinical claims.²⁰ NephraRisk raised the bar by including fairness assessment in deployment-oriented reporting.²² For clinical use, fairness metrics should be tied to false negatives, false positives, referral access, treatment eligibility, and monitoring intensity. In practice, fairness for kidney-risk models can be evaluated through calibration within clinically relevant subgroups, subgroup-specific sensitivity and specificity, false-negative disparities that could delay confirmatory testing, and false-positive burden that could drive unnecessary referral. Nephrology SDOH literature frames unmet social needs and kidney disparities as care-system problems rather than individual deficiencies.^77,85 Race-correction debates in kidney function estimation further show why prediction models require careful handling of race, ethnicity, and structural variables.⁸⁶

Uncertainty is another underdeveloped part of the evidence. Most DKD-XAI studies explain point predictions but give clinicians little help in recognizing when a model is uncertain or outside its training domain. In practice, uncertainty estimates, out-of-distribution detection, missing-data warnings, and drift monitoring may be as important as SHAP plots.⁸⁷ Governance then has to define monitoring, recalibration, update approval, and accountability when model output conflicts with guideline-based care or clinical judgment.^53,54,62

Readiness Synthesis

Taken together, these findings suggest that a practical readiness ladder should start with task definition and endpoint validity, then move through predictor timing, calibration, utility, explanation, external validation, fairness, uncertainty, and prospective workflow evaluation. Most available studies remain on the retrospective development or validation rungs. The strongest examples add external validation, multimodal data, or deployment-oriented reporting.^22,36 Retinal and occult-DKD studies further illustrate translational signals, but they still need prospective evidence before they can be treated as routine early-DKD decision support.^19,39

The near-term role of XAI is therefore narrower but clinically plausible. These models may help stratify risk, make risk drivers visible, and support decisions about closer monitoring or confirmatory testing. Moving beyond that role will require evidence that the tools improve decisions and reduce missed risk in real workflows.^53,59 It will also require trial-reporting and implementation guidance for clinical AI evaluation.^60–62 Transparent prediction-model reporting should remain aligned with TRIPOD and TRIPOD+AI.^16,56 Appraisal and local clinical use should also address PROBAST/PROBAST+AI domains and clinical-impact challenges in deployed settings.^54,57,58

Clinical Boundaries

XAI can make kidney-risk predictions easier to inspect, but it does not resolve the core clinical problems around early DKD: ambiguous endpoints, circular predictors, transportability limits, and possible workflow harm. Box 1 summarizes practical clinical-action boundaries. The following boundaries describe the current clinical envelope for XAI-enabled kidney-risk profiling in T2D.

Box 1.

Clinical Actionability Boundaries

Clinical Actionability Boundary
Use XAI output as a risk discussion and audit aid, not a standalone DKD diagnosis.
Confirm high-risk predictions with guideline-based UACR/eGFR assessment.
Do not treat SHAP, feature importance, or Grad-CAM as causal or pathologic proof.
Link model output to repeat testing, monitoring, treatment review, or referral.
Validate locally before workflow use.
Monitor calibration, subgroup performance, false negatives, false positives, and alert burden.
Treat SDOH variables as support and resource-allocation signals, not patient blame.

Prediction is Not Diagnosis

A risk model estimates probability under its chosen endpoint; it does not make a clinical diagnosis. Diagnosis still depends on clinical assessment, repeat or confirmatory testing when appropriate, and evaluation for alternative causes. In T2D, CKD, DKD, DN, albuminuria, eGFR decline, and ESRD should be kept separate, as supported by nomenclature guidance and DKD consensus literature.^1,37 Broad CKD prediction in T2D can be valuable, but it should not be described as DKD diagnosis unless diabetic kidney disease is distinguished from other causes.^20,36 A high-risk output is best used to prompt guideline-concordant UACR/eGFR confirmation, risk-factor review, medication eligibility assessment, closer monitoring, or referral consideration.^6,7 The CMPB SHAP/eCO model illustrates this position: non-invasive predictors may enrich for albuminuria risk, but modest discrimination and incomplete external validation make it a triage prompt rather than a replacement for laboratory assessment.⁴⁸

Explanation is Not Causation

SHAP, feature importance, partial dependence, and Grad-CAM describe how a model reached its output; they do not establish causality, mechanism, or treatment effect.^19,44 Healthcare-XAI commentaries make this distinction explicit.^25,26 Manuscripts should therefore state that an attribution method identified variables influential for the model, not that the model discovered causal risk factors. This distinction matters when top features are eGFR, albuminuria, creatinine, diabetes duration, HbA1c, blood pressure, retinal findings, or SDOH: each may be clinically meaningful, but none becomes causal because a model ranks it highly.

CKD in T2D is Not Synonymous with DKD

CKD endpoints in T2D are clinically important but etiologically broad. Kidney impairment may reflect diabetic kidney disease, hypertensive nephrosclerosis, vascular disease, medication-related injury, obstruction, age-related decline, primary renal disease, or mixed causes. Population data and non-proteinuric pathway literature support this heterogeneous view.^10,38 Nonalbuminuric renal-impairment studies reinforce the same boundary.^11,12 Models trained on CKD codes or eGFR thresholds may still be useful for monitoring or referral, but they should be labeled as CKD-risk models unless diabetic kidney etiology is explicit.

Outcome Timing Defines Clinical Meaning

Outcome timing changes what a model can claim. Early DKD, ESRD, advanced DKD, WRF, and referral are not interchangeable. ESRD and referral models can help with escalation, capacity planning, and specialist triage, but they are not evidence that XAI detects early or preclinical DKD.^32,46 Wu et al provide useful evidence for one-year CKD progression in patients with established CKD, which is a different clinical setting from early screening.⁴⁷ Nonalbuminuric progression literature also shows why albuminuria and eGFR trajectories should not be treated as identical.¹⁴

Performance and Interfaces are Not Clinical Utility

A high AUROC is not clinical utility. Prediction-model literature separates discrimination from calibration, thresholds, and clinical consequences.¹⁷ A model may rank patients well and still be miscalibrated, biased in subgroups, inconvenient in workflow, or driven by variables that encode current disease.^32,44 Calibration and DCA make models more informative, as shown in stronger studies such as the occult-DKD model, NephraRisk, and the Diabetes Care multimodal model, but they remain intermediate evidence until linked to decisions, workflow, and outcomes.^19,22,36 Web calculators or Streamlit tools can improve access to model output, but they do not by themselves establish clinical benefit.^22,81

Retinal, Fairness, and Treatment Boundaries

Retinal attention maps should be interpreted as model explanations, not pathology. Grad-CAM and saliency maps identify image regions that influenced the output; they do not prove histologic truth, and retinal signals may reflect vascular injury, diabetic retinopathy, systemic microvascular disease, device differences, or population correlations.^36,39 Broader saliency literature supports the same caution.^24,50 Retinal information may still support non-invasive kidney-risk profiling or triage when the signal is grounded in retinal-kidney evidence and interpretable vascular parameters.^49,74 The same restraint applies to treatment. A high SHAP contribution from HbA1c, blood pressure, eGFR, UACR, or SDOH can support communication and audit, but management remains anchored to guidelines, eligibility, contraindications, preferences, and clinician judgment.^6,7 Fairness also cannot be inferred from overall AUROC. CKD disparity evidence makes subgroup calibration, false-negative rates, false-positive burden, referral access, and actionability clinically relevant.^82,88 Health-algorithm bias literature reinforces the deployment risk.^67,68 NephraRisk and the SDOH model point in useful directions, but social variables should guide support and resource allocation rather than blame or rationing.^20,22

Future Research and Implementation Agenda

Future work should start with cleaner clinical tasks. Studies should state whether they target occult DKD, early albuminuria, incident DKD, broad CKD in T2D, progression, ESRD, referral, or DKD vs non-DKD differentiation, and whether the intended use is screening enrichment, risk communication, medication review, monitoring, referral triage, or population-health outreach. Recent DKD-AI reviews show that model-development activity is increasing, making task clarity more important rather than optional.^28,55

Endpoint and predictor timing deserve equal attention. Future incident-risk models should ensure that predictors are measured before the outcome and should state whether kidney-status variables are included. When eGFR, creatinine, UACR, proteinuria, or albuminuria are used, authors should explain whether the model predicts future disease, classifies current disease, or stratifies severity. Diabetes-specific nomenclature work reinforces this distinction.⁵¹

Validation should move beyond internal splitting toward multicenter, cross-country, device-diverse retinal, primary-care, community, and prospective temporal validation. Models should also be recalibrated and tested across clinically meaningful subgroups, including age, sex, race and ethnicity where available, kidney function stage, albuminuria status, obesity, cardiovascular disease, medication exposure, and access-related variables. Kidney-failure risk-equation validation provides a useful precedent for transportability expectations.^63,64

Reporting should make clinical use assessable. AUROC should be accompanied by calibration, threshold-specific sensitivity and specificity, PPV, NPV, DCA, missing-data handling, predictor availability, and alert burden. Prediction-model reporting and sample-size guidance support transparent reporting of candidate predictors, outcome events, and development-sample adequacy.^16,89 Early screening models need sensitivity and false-negative analysis; referral and progression models need false-positive burden and capacity assessment.

XAI methods should be tested for usefulness rather than displayed alone. Studies should examine whether explanations improve clinician understanding, appropriate testing, treatment review, referral decisions, patient communication, and trust calibration. Counterfactual explanations, uncertainty estimates, and out-of-distribution warnings should be explored where they map to real decisions. Explanations should be designed for specific users rather than shown as generic technical outputs.^25,26

Biomarker-driven and multi-omics approaches may also reshape future explainable-AI kidney-risk profiling, particularly if metabolomic, proteomic, genomic, retinal, and routine clinical signals can be integrated without losing endpoint clarity or transportability. The metabolite-enriched diabetic-nephropathy model illustrates this potential, but internal validation and renal-status circularity mean that biomarker-rich explanations should remain hypothesis-generating until they are externally validated and linked to clinical decisions.^53,69

Fairness and implementation monitoring should become standard. Future studies should report subgroup performance, calibration, false-negative and false-positive rates, and consequences of errors. SDOH variables should guide support and resource allocation, not blame or rationing.^68,77 Socioeconomic disparities in CKD make subgroup and access monitoring clinically relevant.⁷⁸ Deployed systems should monitor model drift, data shifts, medication-era changes, and practice changes.^53,54

Implementation studies should use reporting and appraisal frameworks as guardrails, not as a final checklist. TRIPOD+AI and PROBAST+AI support transparent reporting and appraisal of prediction models.^56,57 PROBAST remains the base risk-of-bias tool for prediction-model appraisal.⁵⁸ DECIDE-AI is relevant for early clinical evaluation, while CONSORT-AI and SPIRIT-AI become important when systems move toward trials.^59,60 Protocol-level AI trial reporting should also align with SPIRIT-AI, and broader trustworthy-AI implementation should be judged against FUTURE-AI principles.^61,62

Conclusion

Explainable AI for kidney-risk profiling in T2D has moved beyond isolated model development. Recent studies increasingly report external validation, calibration, decision curve analysis, explanation methods, web tools, or prospective proof-of-concept evaluation. The strongest evidence suggests that XAI may help identify occult or early kidney involvement, support deployment-oriented DKD risk estimation, and add retinal or multimodal signals as complementary risk-profiling routes.

The evidence is not yet strong enough for broad routine deployment. Endpoints remain heterogeneous, many models include predictors close to kidney disease definitions, external validation is uneven, prospective workflow evidence is rare, and explanations have not consistently been shown to improve decisions or outcomes. The most defensible current role is as a complementary risk-stratification and audit layer within guideline-based care. High-risk outputs should lead to confirmatory assessment, clinical review, and shared decision-making rather than automated diagnosis or treatment.

The next phase should prioritize models that are prospectively validated, tested in workflow, fair across clinically relevant subgroups, and explicit about calibration, thresholds, clinical utility, uncertainty, monitoring, and maintenance. XAI will matter clinically only if it helps diabetes and kidney-care teams make safer, more equitable, and more timely decisions in real practice.

Funding Statement

This work was supported by the Science and Technology Project of Xuzhou Municipal Health Commission (XWKYHT20250099). The funder had no role in the design, literature interpretation, writing of the paper, or decision to submit the article for publication.

Declaration of Generative AI and AI-Assisted Technologies in the Writing Process

During the preparation of this paper, the authors used ChatGPT (GPT-5.5 Thinking, OpenAI) and OpenAI Codex to assist with language editing, paper organization, formatting checks, and refinement of non-substantive wording. These tools were used to improve clarity, readability, and consistency of presentation. Generative AI tools were not used to generate original research data, perform statistical analyses, make clinical decisions, or replace author-led interpretation of the literature. The authors reviewed, edited, and verified all AI-assisted outputs and take full responsibility for the content of the final version.

Abbreviations

AI, artificial intelligence; AUROC, area under the receiver operating characteristic curve; CKD, chronic kidney disease; DCA, decision curve analysis; DKD, diabetic kidney disease; DN, diabetic nephropathy; eGFR, estimated glomerular filtration rate; EHR, electronic health record; ESRD, end-stage renal disease; ESKD, end-stage kidney disease; PPV, positive predictive value; NPV, negative predictive value; SDOH, social determinants of health; SHAP, Shapley additive explanations; T2D, type 2 diabetes; UACR, urinary albumin-to-creatinine ratio; WRF, worsening renal function; XAI, explainable artificial intelligence.

Data Sharing Statement

No new datasets were generated or analyzed for this article. The review is based on previously published literature.

Author Contributions

Tianqiang Wu: Conceptualization, Methodology, Investigation, Data Curation, Writing – Original Draft, Visualization, Writing – Review and Editing. Wenpin Cai: Methodology, Supervision, Validation, Writing – Review and Editing. Zihan Guan: Investigation, Data Curation, Validation, Project Administration, Writing – Review and Editing. Zhixiang Li: Conceptualization, Methodology, Validation, Supervision, Project Administration, Funding Acquisition, Writing – Review and Editing. All authors made a significant contribution to the work reported, whether in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising, or critically reviewing the article; gave final approval of the version to be published; agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

Zihan Guan reports support for the paper from Wenzhou Hospital of Traditional Chinese Medicine and Zhixiang Li reports support for the paper from Science and Technology Project of Xuzhou Municipal Health Commission. The authors report no other conflicts of interest in this work.