Artificial intelligence in chronic kidney disease: Early detection, risk prediction, and personalized treatment strategies

Abstract

Chronic kidney disease (CKD) is a heterogeneous and frequently underdiagnosed condition in which conventional diagnostic and prognostic approaches based on static biomarkers and linear risk models offer limited sensitivity for early detection and individualized risk stratification. Artificial intelligence (AI) enables integration of longitudinal, high-dimensional, and multimodal clinical data, providing opportunities to address key limitations in current CKD care. This narrative review synthesizes human clinical studies published between 2019 and 2025 evaluating AI applications across the CKD continuum, including early detection, risk prediction, disease progression, treatment personalization, complication management, dialysis, and kidney transplantation. Evidence was analyzed thematically with emphasis on model performance, validation strategies, explainability, and translational readiness. Across multiple domains, AI-based models generally demonstrated improved discrimination and risk stratification compared with traditional statistical approaches, particularly when incorporating longitudinal trajectories and multimodal inputs. However, most models were exploratory or internally validated, with limited external validation and minimal prospective evaluation. Reported applications included earlier CKD identification, improved prediction of progression and acute kidney injury, optimization of pharmacologic and dialysis strategies, and enhanced prognostication in kidney transplantation. Overall, AI holds promise for more precise CKD detection and management, but real-world impact will require prospective validation, standardized reporting, and rigorous assessment of clinical utility.

Key Words: Machine learning; Deep learning; Precision nephrology; Clinical decision support; Multimodal data integration

Core Tip: Artificial intelligence (AI) holds significant promise for transforming chronic kidney disease (CKD) care by integrating longitudinal, high-dimensional, and multimodal clinical data to support earlier detection, improved risk stratification, and personalized management. This review synthesizes recent evidence on AI applications across the CKD continuum, including early detection, risk prediction, personalized treatment, complication management, dialysis, and kidney transplantation. Beyond reporting model performance, it critically examines validation quality, calibration, explainability, equity, and real-world implementation barriers, highlighting the persistent gap between methodological innovation and routine clinical deployment. The review outlines key priorities required to translate AI tools into safe, equitable, and clinically meaningful CKD care.

INTRODUCTION

Chronic kidney disease (CKD) is a pervasive and progressively debilitating condition characterized by persistent abnormalities in kidney structure and function, including reduced glomerular filtration rate (GFR) and/or markers of kidney damage such as albuminuria. It represents a major global public health challenge, affecting more than 650 million individuals worldwide, with prevalence and incidence continuing to rise over the past three decades; in 2021 alone, nearly 674 million people were living with CKD, and its associated disability-adjusted life years continue to increase in most regions of the world[1,2]. CKD contributes substantially to global mortality and disproportionately affects older adults and populations with high prevalences of diabetes mellitus, hypertension, and obesity[2]. Beyond its direct clinical burden, CKD amplifies cardiovascular risk and imposes significant socioeconomic strain on healthcare systems, particularly in low- and middle-income countries where access to dialysis and transplantation remains limited. Importantly, early CKD is frequently asymptomatic, resulting in delayed recognition and missed opportunities for timely intervention[3].

From a clinical perspective, CKD is not a static diagnosis but a dynamic, trajectory-based condition characterized by heterogeneous etiologies and highly variable rates of progression. However, traditional approaches to CKD diagnosis, prognostication, and management continue to rely largely on periodic laboratory measurements, including estimated GFR, urine albumin-to-creatinine ratios, and conventional risk scoring systems. Although these tools provide essential clinical information, they exhibit important limitations in early detection, individualized risk stratification, and adaptability to longitudinal patient trajectories[4]. Standard prediction models often fail to capture the complex, non-linear interactions among demographic characteristics, comorbidities, biochemical markers, and temporal trends, leading to limited sensitivity for early-stage disease and suboptimal identification of rapid progressors[5]. As a result, therapeutic decisions are frequently guided by generalized protocols that inadequately reflect inter-individual variability in disease biology and treatment response, constraining the realization of precision nephrology[6].

Artificial intelligence (AI), encompassing machine learning (ML) and deep learning (DL) methodologies, has emerged as a promising analytical framework for addressing these limitations. AI techniques are capable of learning complex patterns from high-dimensional and longitudinal data derived from electronic health records (EHRs), laboratory testing, medical imaging, and omics platforms, enabling more granular risk prediction and earlier disease detection than traditional statistical approaches[7,8]. A range of AI paradigms has been explored in CKD research, including supervised ML algorithms for risk stratification, DL models for imaging-based phenotyping and temporal modeling, natural language processing for extraction of information from unstructured clinical text, and multimodal frameworks that integrate clinical, imaging, and molecular data to characterize disease heterogeneity[9-12]. Collectively, these approaches aim to generate clinically actionable and interpretable insights that align with individualized care pathways and real-world clinical workflows[8] (Figure 1).

Figure 1 Artificial intelligence across the chronic kidney disease continuum. Schematic showing how artificial intelligence supports clinical decision-making across all chronic kidney disease stages, integrating diverse data modalities to enable longitudinal, multimodal, and stage-specific care from early detection to transplantation. AI: Artificial intelligence; CKD: Chronic kidney disease; DL: Deep learning; NLP: Natural language processing; EHR: Electronic health record.

Several prior reviews have summarized AI applications in nephrology or specific CKD subdomains, often focusing on algorithmic performance or technical model development. In contrast, this work is intentionally designed as a structured narrative review, allowing integration of heterogeneous study designs, data sources, and AI methodologies that are not readily amenable to formal meta-analysis or systematic synthesis. A narrative approach is particularly appropriate in the rapidly evolving field of AI in CKD, where models vary widely in objectives, inputs, validation maturity, and clinical readiness, and where contextual interpretation is essential to assess translational relevance. The present review synthesizes evidence across the full CKD continuum-from early detection and progression prediction to complication management, dialysis, and kidney transplantation—while explicitly emphasizing model validation, explainability, equity considerations, and translational readiness. By integrating methodological appraisal with clinical context, this review seeks to identify not only areas of promise but also the key barriers that must be addressed to enable responsible and effective implementation of AI-driven tools in routine CKD care.

METHODOLOGY

This review was designed as a structured narrative literature review to synthesize contemporary evidence on applications of AI in CKD, with emphasis on early detection, risk prediction and disease progression modeling, personalized treatment strategies, complication management, dialysis, and kidney transplantation. A narrative (non-systematic) approach was deliberately chosen because AI research in CKD is characterized by substantial heterogeneity in model objectives, data sources, outcome definitions, validation strategies, and clinical maturity, which precludes meaningful quantitative pooling or formal meta-analysis. This approach allowed thematic integration of diverse study designs while preserving clinical interpretability and translational relevance.

Literature search strategy

A comprehensive literature search was conducted across major biomedical and interdisciplinary databases, including PubMed/MEDLINE, EMBASE, Web of Science Core Collection, and Scopus. The search covered articles published between January 2019 and September 2025, reflecting the contemporary era of ML, DL, and explainable AI (XAI) applications in nephrology. Only peer-reviewed journal articles were considered; preprints and non-peer-reviewed manuscripts were excluded to ensure evidentiary robustness.

To enhance completeness, reference lists of key reviews and highly cited primary studies were manually screened.

Search strategies combined controlled vocabulary terms (MeSH and Emtree) with free-text keywords related to AI and kidney disease. Core AI-related terms included “artificial intelligence”, “machine learning”, “deep learning”, “neural networks”, “explainable AI”, “natural language processing”, “radiomics”, “reinforcement learning”, “digital twin” and “multimodal models”. CKD-related terms included “chronic kidney disease”, “CKD”, “kidney failure”, “end-stage renal disease”, “ESRD”, “dialysis”, “acute kidney injury”, “AKI”, and “kidney transplantation”. Database-specific adaptations were applied, and a representative PubMed search strategy is provided in the Supplementary Table 1 to support reproducibility.

Study selection and eligibility criteria

Retrieved records were de-duplicated using reference-management software, followed by a three-stage screening process consisting of title screening, abstract screening, and full-text eligibility assessment for relevance to AI applications in CKD.

Studies were eligible for inclusion if they: Involved human subjects with CKD at any stage, individuals at risk of CKD, patients with CKD-related complications, or kidney transplant donors or recipients; applied AI-based model types, including ML, DL, natural language processing, radiomics-based models, reinforcement learning, or multimodal AI frameworks; and addressed at least one clinically relevant domain, including early CKD detection or screening, risk stratification and progression prediction, acute kidney injury prediction, personalized treatment or dosing strategies, dialysis management, complication prediction [e.g., cardiovascular disease (CVD), anemia, CKD-mineral and bone disorder (CKD-MBD)], or kidney transplantation (allocation, rejection, graft survival, or immunosuppression management).

Studies were excluded if they were purely methodological, simulation-only, or algorithm-development investigations without direct human clinical data or clear clinical applicability. Additional exclusions included case reports, narrative editorials, commentaries, conference abstracts without full-text availability, and preclinical or animal-only investigations. When duplicate datasets were identified across multiple publications, the most comprehensive or methodologically mature report was retained.

Data extraction and thematic synthesis

For eligible studies, data were extracted on study design, population characteristics, data sources (e.g., EHR, registries, imaging, omics), AI methodology, degree of model validation (exploratory development, internal validation, external validation, or prospective evaluation), performance metrics [e.g., area under the receiver operating characteristic (AUROC), C-index, calibration measures], explainability approaches, and key clinical findings. Given the diversity of study designs and outcomes, quantitative pooling was not undertaken. Instead, findings were synthesized thematically in alignment with the structure of the review.

Studies were grouped into the following thematic domains: (1) AI for early detection of CKD; (2) AI-based risk prediction and disease progression modeling; (3) AI in personalized treatment and pharmacotherapy optimization; (4) AI for prediction and management of CKD complications; (5) AI applications in kidney transplantation; and (6) Explainability, bias, ethical considerations, and real-world implementation barriers. This framework facilitated comparison across AI methods and clinical contexts while emphasizing translational readiness.

Quality assessment and risk of bias considerations

Owing to heterogeneity in study designs and reporting standards, a formal meta-analytic risk-of-bias assessment was not performed. Methodological quality was instead appraised qualitatively using established reporting and validation principles for clinical prediction models and AI-based decision-support systems. Diagnostic and prognostic studies were evaluated with reference to TRIPOD-AI and PROBAST-AI domains, focusing on population selection, outcome definition, predictor handling, validation strategy, and calibration reporting. Interventional and decision-support studies were assessed for clarity of clinical integration, comparator selection, and outcome ascertainment.

Particular attention was given to the presence of external validation, subgroup performance reporting, use of explainability techniques, and bias mitigation strategies, as these factors critically influence generalizability, equity, and clinical adoption. Common limitations, including single-center development, data leakage, class imbalance, and incomplete calibration assessment, were explicitly considered during synthesis.

Transparency and reporting

The review process followed principles consistent with narrative and scoping review guidance, with explicit description of database sources, eligibility criteria, screening steps, and thematic synthesis methods to enhance transparency and reproducibility. Given the narrative (non-systematic) design of the review, study selection is described textually rather than as a formal systematic review workflow. Emphasis was placed on clinical relevance, validation strength, and translational applicability of AI models in CKD care.

CURRENT DIAGNOSTIC CHALLENGES IN CKD: GAPS EXPLOITABLE BY AI

Accurate assessment of kidney function remains a central diagnostic challenge in CKD. Estimated GFR (eGFR), most commonly derived from serum creatinine, underpins CKD staging, yet widely used creatinine-based equations exhibit clinically meaningful variability and systematic bias across populations and care settings[13]. The 2021 race-free CKD-epidemiology collaboration (EPI) equation, while addressing equity concerns, demonstrates altered bias characteristics compared with prior formulations; multicohort analyses report a larger negative bias relative to measured GFR in some non-Black subgroups (approximately -3.9 mL/minute/1.73 m² vs -0.5 mL/minute/1.73 m²), with downstream implications for CKD classification, referral timing, and treatment eligibility[14]. In contrast, cystatin C-based estimation has repeatedly shown reduced bias and improved classification accuracy in specific clinical contexts, particularly where creatinine is confounded by muscle mass, diet, or non-renal factors[15,16]. Collectively, inter-equation and inter-assay discrepancies lead to misclassification around key eGFR thresholds used for staging, medication dosing, and transplant evaluation, creating diagnostic uncertainty that is amenable to analytic approaches capable of integrating heterogeneous biomarkers and estimating latent kidney function. Most AI models developed to address eGFR estimation or CKD detection at this stage remain exploratory or internally validated, relying on retrospective datasets and surrogate reference standards rather than prospective measurement of true GFR.

A second major challenge is the frequent late recognition of CKD due to its asymptomatic early course and inconsistent testing practices. Large registry and population-based studies demonstrate persistently high rates of unrecognized disease. Multisite analyses of stage-3 CKD report unawareness exceeding 70%-90% in several high-income settings, with reported values of 95.5% in France, 84.3% in Germany, and 77.0% in other cohorts; concordantly, registry data indicate that more than 70% of individuals meeting biochemical criteria for CKD lack a contemporaneous diagnostic code[17,18]. Stage-specific analyses further show unawareness rates of approximately 80% for stages 1-2, about 71% for stage 3a, about 49% for stage 3b, and > 30% for stage 4 in some cohorts[19-21] (Table 1). These diagnostic gaps delay preventive interventions, reduce opportunities to slow disease progression, and increase downstream costs when patients present with advanced CKD or require renal replacement therapy. AI-based screening models proposed to address this gap are predominantly internally validated and evaluated using discrimination metrics such as AUROC, which may overestimate real-world utility in low-prevalence early-stage CKD populations where positive predictive value remains modest despite high area under the curve (AUC) values.

Table 1 Current diagnostic limitations in chronic kidney disease and key quantitative evidence.

Diagnostic domain	Limitation	Key quantitative findings	Ref.
eGFR estimation	Bias and variability across equations	CKD-EPI 2021 mean bias -3.9 mL/minute/1.73 m² vs -0.5 mL/minute/1.73 m² (prior equation)	[14]
CKD awareness	High rates of undiagnosed disease	Stage-3 unawareness 70%-90%; > 70% lack diagnostic code	[17,18]
Stage-specific awareness	Late recognition in early stages	Approximately 80% (stages 1-2), approximately 71% (3a), approximately 49% (3b), >30% (stage 4)	[19]
Data integration	Fragmented multimodal data	Incomplete lab/imaging capture; limited NLP use	[20,21]

Bias values are expressed relative to measured glomerular filtration rate and reported in mL/minute/1.73 m². eGFR: Estimated glomerular filtration rate; CKD-EPI: Chronic kidney disease epidemiology collaboration; CKD: Chronic kidney disease; NLP: Natural language processing.

Fragmentation of clinical data across laboratories, imaging systems, and EHRs further limits timely and accurate CKD detection. Relevant signals-including longitudinal serum chemistry, albuminuria, renal imaging, histopathology, and clinician documentation-are often distributed across disconnected data silos and stored in heterogeneous formats, impeding automated surveillance and longitudinal phenotyping[20]. Informatics studies highlight persistent barriers to multimodal data integration: Incomplete capture of external laboratory or imaging results, underutilization of free-text clinical notes without natural language processing, and limited data standardization that constrains model portability across institutions[21,22]. Although pragmatic trials show that structured EHR-based algorithms and care-management workflows can improve CKD identification and process-of-care metrics, such gains depend on interoperable data streams and robust phenotype definitions that remain unevenly implemented across health systems[22]. AI models addressing data fragmentation often report improved discrimination but inconsistently assess calibration or decision-curve performance, limiting insight into whether earlier detection would translate into meaningful net clinical benefit or reduced overtreatment.

Finally, CKD is characterized by substantial etiologic and prognostic heterogeneity that challenges uniform diagnostic and management strategies. While diabetes and hypertension account for much of the global CKD burden, glomerular diseases, inherited conditions, and environmental or drug-induced nephropathies produce diverse phenotypes with highly variable progression trajectories[23]. Many patients experience slow functional decline, whereas a clinically important minority progress rapidly to kidney failure-patterns that conventional summary measures often fail to distinguish. Treatment response is similarly heterogeneous; variability in outcomes following renin-angiotensin-aldosterone system (RAAS) blockade, SGLT2 inhibition, or immunomodulatory therapy suggests that stratified or individualized prediction could improve therapeutic selection and risk-benefit assessment. Exploratory and internally validated multimodal AI models integrating longitudinal laboratory data, imaging features, and molecular or pharmacologic information have shown promise in capturing this heterogeneity, but few have undergone external validation across ethnically or socioeconomically diverse cohorts, raising concerns regarding bias, fairness, and generalizability.

Across these diagnostic domains, common methodological limitations include class imbalance, enrichment of training datasets with advanced disease, and reliance on retrospective labeling, all of which can inflate apparent model performance while limiting applicability to real-world early CKD detection. Addressing these issues will require prospective evaluation, robust calibration assessment, and explicit consideration of clinical decision thresholds rather than sole reliance on discrimination metrics.

AI FOR EARLY DETECTION OF CKD

ML models leveraging routine clinical, laboratory, and administrative data have demonstrated strong ability to identify early or preclinical CKD before conventional diagnostic thresholds are met. Using claims- or EHR-derived features-including demographics, comorbidities, medications, and longitudinal laboratory trends-these models can detect subtle deviations in creatinine or eGFR trajectories that precede formal diagnosis[24]. Large-scale studies show that when longitudinal data are available, ML approaches outperform static rule-based criteria by capturing gradual kinetic changes rather than relying on single time-point thresholds[25,26]. The majority of these studies represent exploratory or internally validated models developed retrospectively, with limited assessment of performance stability across healthcare systems or population subgroups.

At the point-of-care level, ML-based early-warning systems are designed to flag individuals already engaged with healthcare services who are at imminent risk of developing CKD. A representative example is a convolutional neural network (CNN) trained on Taiwan’s National Health Insurance Research Database (about 18000 CKD cases and 72000 controls), which achieved AUROCs of 0.957 and 0.954 for 6- and 12-month incident CKD prediction using claim-level features such as age, diabetes, and medication exposure[25]. Comparable gradient-boosting and ensemble models applied to EHR data typically report AUROCs between 0.80 and 0.92 when laboratory variables are included, with explicit modeling of longitudinal creatinine or eGFR trajectories consistently improving sensitivity for early disease detection[25,26]. However, discrimination metrics alone do not capture clinical utility; few studies report calibration performance or decision curve analysis, and high AUROC values may overestimate benefit in early CKD where disease prevalence is low and false-positive alerts can dominate clinical workflows. Clinically, these tools function as surveillance mechanisms that prompt confirmatory testing or nephrology referral before irreversible decline occurs, but most remain internally validated and pre-deployment rather than prospectively evaluated in routine care (Table 2)[27-33].

Table 2 Major artificial intelligence model categories for chronic kidney disease detection and early risk stratification.

Model category	Data source	Typical task	Key performance range	Primary clinical use	Ref.
Clinical early-warning	EHR, claims, labs	6-12 months incident CKD prediction	AUROC 0.80-0.95	Trigger confirmatory testing/referral	[25,26]
Multimarker ML	Creatinine + cystatin C + labs	Threshold reclassification	↓ False negatives near cutoffs	Reduce misclassification	[16,29]
Imaging-based DL	CT, ultrasound	Structural injury detection	Dice 81%-94%; Acc 86%-90%	Noninvasive fibrosis/triage	[30,31]
Population screening	Registries, insurance	Risk stratification	AUROC 0.80-0.95	Targeted screening	[33]

Performance ranges reflect external validation when reported. CKD: Chronic kidney disease; CT: Computed tomography; AUROC: Area under the receiver operating characteristic curve; ML: Machine learning; DL: Deep learning.

To address known limitations of creatinine-based estimation, several ML frameworks incorporate alternative filtration markers such as cystatin C. In populations where creatinine is confounded by muscle mass, body composition, or pharmacologic interference, cystatin C-based estimation improves early detection and risk stratification[27]. Meta-analytic and cohort evidence demonstrates that adding cystatin C reduces systematic bias and improves reclassification accuracy, particularly near eGFR thresholds that guide medication dosing and referral decisions[16,28]. ML models that fuse creatinine, cystatin C, albuminuria, and electrolyte data can dynamically weight biomarkers according to patient context and have been shown to reduce false-negative rates for early CKD compared with creatinine-only rules in external validation cohorts[29]. These externally validated findings strengthen translational relevance, although calibration drift across ethnic, socioeconomic, and care-access strata remains a concern when models are applied beyond their derivation populations.

DL methods applied to renal imaging provide complementary structural markers of early chronic injury. CNN-based segmentation models achieve high concordance with expert annotations; for example, a 3D nnU-Net applied to contrast-enhanced computed tomography (CT) achieved dice similarity coefficients of 93.5% for renal parenchyma and 81.5% for renal cortex in a 190-patient cohort, enabling volumetric measures that differed significantly across CKD stages[30]. Ultrasound-based DL radiomics has similarly demonstrated clinical relevance: A UNet-based pipeline combined with CNN feature extraction achieved 86%-90% accuracy and F1 scores around 0.84 for noninvasive grading of interstitial fibrosis and tubular atrophy, with added prognostic value beyond baseline eGFR and proteinuria[31]. Automated extraction of cortical thickness, echogenicity, and texture features further correlates with histopathologic fibrosis and longitudinal functional decline, supporting imaging-derived biomarkers as early indicators of CKD when integrated with laboratory trajectories[32]. Most imaging-based models remain exploratory or internally validated, often trained on single-center datasets with limited evaluation of generalizability across imaging platforms or operator-dependent acquisition protocols.

At the population scale, registry and insurance databases enable AI-driven screening tools that identify high-risk but undiagnosed individuals, particularly among patients with diabetes or hypertension who account for most CKD burden. National claims analyses, including the Taiwan National Health Insurance Research Database study, demonstrate that models trained solely on administrative features can achieve AUROCs exceeding 0.95 for short-term CKD onset prediction, supporting targeted laboratory follow-up rather than universal screening[25]. Other multisite studies focusing on diabetic cohorts report externally validated AUROCs of 0.80-0.88 for 1- to 3-year CKD risk prediction, enabling risk-stratified outreach where screening resources are constrained[26,33]. Despite external validation, few studies evaluate downstream clinical outcomes or net benefit using decision curve analysis, and model performance may be inflated by class imbalance and enrichment of datasets with high-risk individuals. By estimating absolute risk, these tools facilitate prioritization of individuals most likely to benefit from confirmatory testing, although careful calibration is required to avoid disproportionate false positives in low-risk or underserved subgroups.

Across these applications, the primary clinical value of AI lies in earlier identification and targeted intervention rather than marginal improvements in discrimination metrics. Deployment risks include miscalibration across age, comorbidity, ethnic, or socioeconomic strata, alert fatigue from false positives, and reduced generalizability when models are transferred across health systems. XAI approaches-most commonly SHapley Additive exPlanations (SHAP) and local interpretable model-agnostic explanations (LIME)-have been applied to nephrology datasets to identify clinically plausible drivers (e.g., longitudinal creatinine trends, diabetes duration, albuminuria) and to reveal subgroup-specific performance disparities, thereby supporting clinician trust and safer implementation[5,10,34,35]. However, explainability is inconsistently reported and rarely integrated into real-time EHR interfaces, limiting its impact on clinical decision-making and acceptance.

From an implementation perspective, early CKD detection models face practical barriers including clinician acceptance of algorithmic alerts, integration into heterogeneous EHR systems, and unresolved medico-legal accountability for missed or erroneous predictions. These factors contribute to a persistent deployment gap, whereby high-performing models remain confined to retrospective evaluation rather than prospective, real-world use.

AI-BASED RISK PREDICTION AND DISEASE PROGRESSION MODELING

ML survival models have expanded risk prediction for progression from CKD to end-stage renal disease (ESRD) by incorporating longitudinal biomarker trajectories and flexible hazard estimation. Approaches span penalized Cox models (e.g., CoxNet), ensemble survival learners (random survival forests), and neural time-to-event architectures such as DeepSurv, often implemented using landmarking or joint-modeling frameworks to accommodate time-varying covariates[36,37]. Compared with traditional single-baseline Cox regression and widely used clinical tools such as the kidney failure risk equation (KFRE), these ML models consistently demonstrate improved discrimination and calibration when longitudinal eGFR slopes, albuminuria trajectories, and time-dependent laboratory features are included[38,39]. Most published CKD progression models are exploratory or internally validated and derived from retrospective cohorts, with a smaller subset undergoing external validation and very few evaluated prospectively in real-world clinical settings.

A representative multicohort study applying a landmarking random-survival-forest pipeline to longitudinal clinicopathological data (n = 4950 derivation; n = 8729 external validation) reported a median concordance index of 0.848 (84.84%) for ESKD prediction, explicitly modeling death as a competing risk, with median integrated Brier scores of approximately 0.03 across landmark times[40]. These results compare favorably with static baseline Cox models and KFRE-based estimates, which do not fully exploit temporal information and may misestimate risk when competing mortality is substantial. Across studies, sustained eGFR decline, rising albuminuria, recurrent hemoglobin drops, and progressive hypoalbuminemia consistently rank among the most influential predictors and contribute to reduced miscalibration relative to single-timepoint risk scores in external validation[41]. However, even externally validated models may exhibit calibration drift when applied across healthcare systems with differing case-mix, referral patterns, or access to nephrology care, underscoring the need for local recalibration and clinical impact assessment.

Prediction of acute kidney injury superimposed on CKD represents a complementary risk-prediction task with immediate clinical relevance. ML models trained on intensive-care datasets such as MIMIC-IV and eICU use high-frequency data-including vital signs, urine output, serial creatinine, hemodynamic support, and comorbidity indices-to enable near-real-time risk stratification[42]. In a MIMIC-IV cohort of 1457 patients with diabetes and heart failure, a LightGBM model achieved an AUC of 0.973 in training and 0.804 in validation for 30-day AKI prediction, in a population with a very high AKI incidence (about 83%), underscoring the importance of population selection and external generalizability testing[36]. This example illustrates how class imbalance and disease enrichment can inflate apparent performance, particularly when models are evaluated in high-acuity settings that differ substantially from general CKD populations. Operational implementations emphasize continuous risk updating and calibrated alert thresholds; explainability layers such as SHAP identify proximal drivers (e.g., recent creatinine rise, hypotension, diuretic exposure) and support clinician trust[43]. While SHAP-based explanations enhance transparency, their clinical usefulness depends on consistent performance across demographic and socioeconomic subgroups, which remains insufficiently reported. Strategies to mitigate alarm fatigue-including sustained-risk confirmation, tiered alerts, and clinician verification-reduce false positives but may trade sensitivity for specificity, requiring context-specific calibration[44,45].

Multimodal progression models integrating longitudinal EHR data with imaging-derived biomarkers and molecular information further improve prognostic stratification, particularly for identifying rapid progressors. The addition of radiomic features reflecting cortical thickness, echogenicity, and texture from CT or ultrasound improves C-indices by approximately 0.03-0.07 over laboratory-only models, with imaging signatures correlating with histologic fibrosis and subsequent eGFR decline in validation cohorts[46]. Genomic and polygenic risk scores provide modest incremental reclassification for selected patients, such as those with genetic susceptibility to rapid decline despite otherwise low-risk clinical profiles, although large sample sizes are required for stable integration[47]. Consequently, most implementations favor ensemble or hierarchical fusion strategies and report time-dependent discrimination, calibration, and decision-curve analyses to quantify clinical utility[48-50] (Table 3). Despite these advances, most multimodal models remain internally validated, with limited evaluation of fairness across ethnic groups, socioeconomic strata, or healthcare access contexts.

Table 3 Survival modeling approaches for kidney failure prediction and comparative performance.

Model type	Data inputs	Competing risks handled	Typical performance	Comparator	Key advantage	Ref.
Traditional Cox	Baseline labs	Rarely	C-index 0.65-0.75	KFRE	Simplicity	[39]
KFRE	Static clinical variables	No	C-index 0.70-0.80		Widely adopted	[39]
ML survival (RSF, DeepSurv)	Longitudinal labs	Yes	C-index 0.75-0.88	Cox, KFRE	Dynamic risk	[40]
Multimodal ML	Labs + imaging ± genomics	Yes	+0.03-0.07 C-index gain	Lab-only ML	Rapid progressor ID	[46]

Reported performance ranges are study-level estimates derived from independent validation cohorts. KFRE: Kidney failure risk equation; ML: Machine learning; RSF: Random survival forest.

Beyond statistical performance, AI-based risk models influence several critical clinical decisions. High predicted short- to intermediate-term ESRD risk can guide earlier nephrology referral, prioritization for vascular access creation, and timing of modality education. Competing-risk-aware predictions help distinguish patients likely to die before reaching ESRD from those who may benefit from aggressive CKD management or transplant evaluation. Risk stratification also supports clinical trial eligibility, enabling enrichment of studies with rapid progressors who are more likely to experience endpoints within feasible follow-up periods (Figure 2). From an implementation perspective, integration of CKD risk models into clinical workflows is constrained by clinician acceptance of algorithmic risk scores, challenges embedding time-updated predictions within EHR systems, and unresolved medico-legal accountability for decisions influenced by AI outputs. In this context, AI models function as decision-support tools that refine-but do not replace-clinical judgment and established guidelines.

Figure 2 Dynamic prediction of chronic kidney disease progression using longitudinal data. Illustrative comparison of static baseline risk estimation vs artificial intelligence based models that update risk over time using longitudinal estimated glomerular filtration rate trajectories, enabling earlier identification of rapid progressors and accounting for competing risks such as death vs progression to end-stage kidney disease. AI: Artificial intelligence; eGFR: Estimated glomerular filtration rate; ESRD: End-stage renal disease.

AI IN PERSONALIZED CKD TREATMENT STRATEGIES

AI applications in CKD increasingly focus on treatment decision support, translating individualized risk estimates into actionable clinical choices rather than merely predicting progression. By integrating longitudinal laboratory trajectories, medication exposure, and physiologic data, these systems inform when and how to intervene while preserving clinician oversight (Figure 3). Most treatment-focused AI tools remain exploratory or internally validated and are designed to augment, rather than replace, clinician judgment.

Figure 3 Artificial intelligence-supported personalized treatment decision framework in chronic kidney disease. Schematic illustrating how artificial intelligence-generated risk predictions inform individualized clinical actions across chronic kidney disease care, with clinician oversight maintained at all stages of decision-making. CKD: Chronic kidney disease; RAAS: Renin-angiotensin-aldosterone system.

A major area of application is prediction of drug-related nephrotoxicity and optimization of pharmacotherapy in CKD. ML models have been developed to anticipate nephrotoxic events for specific agents, including cisplatin, colistin, and calcineurin inhibitors, as well as electrolyte disturbances associated with RAAS blockade[51]. For cisplatin, multicenter ML studies report AUROCs of 0.78-0.86 and sensitivity/specificity tradeoffs around 0.80-0.88 using baseline renal function, cumulative dose, and concomitant nephrotoxins[52]. Colistin nephrotoxicity models similarly achieved AUROCs > 0.80 and identified cumulative exposure, baseline GFR, concurrent vancomycin use, and age as dominant predictors[53] (Figure 3). Although discrimination performance is consistently reported, calibration metrics and decision curve analyses are inconsistently assessed, limiting understanding of whether predicted risk thresholds translate into net clinical benefit or avoidance of unnecessary treatment modification. Class imbalance and enrichment for high-risk patients may further inflate apparent performance in retrospective evaluations.

In transplant and advanced CKD populations, tacrolimus-associated nephrotoxicity prediction has leveraged joint modeling of trough levels and creatinine kinetics to provide early warning signals within narrow dose-adjustment windows[54]. These models are predominantly internally validated and derived from specialized transplant centers, raising concerns regarding generalizability across ethnically diverse populations and healthcare systems with differing access to therapeutic drug monitoring. AI has also been applied to RAAS inhibitor management: Models predicting individualized hyperkalemia risk after RAAS initiation combine baseline potassium, eGFR, comorbidities, and concomitant medications to forecast 30-90-day risk with AUROCs of approximately 0.72-0.86, supporting tailored monitoring strategies rather than indiscriminate discontinuation[55,56]. XAI techniques, including SHAP-based feature attribution, have been used to highlight clinically intuitive drivers such as baseline potassium, diabetes status, and diuretic exposure, supporting clinician trust while revealing subgroup-specific risk gradients. These tools function primarily as decision-support systems, recommending monitoring cadence and alerting clinicians to elevated risk rather than automating prescribing decisions.

AI-based decision support has also been applied to dialysis planning, particularly in identifying patients likely to require renal replacement therapy within defined time horizons. ML models trained on longitudinal clinical data outperform clinician judgment in retrospective comparisons for predicting dialysis initiation within 1-3 years, facilitating earlier modality education and access planning. An XGBoost model trained on routine clinic data demonstrated improved early identification of patients initiating maintenance dialysis within 12 months, as reflected by favorable time-dependent C-indices and precision-recall performance[57]. Most such models are internally or externally validated retrospectively, with limited prospective or real-world deployment data assessing downstream outcomes such as timely access creation or patient-centered decision-making. These systems support referral timing and vascular access creation, rather than automating initiation decisions.

For intradialytic management, ML models predict intradialytic hypotension (IDH) using real-time hemodynamic data, ultrafiltration rates, and interdialytic weight changes. Models capable of forecasting IDH 15-75 minutes in advance report sensitivities of approximately 0.78-0.88 and specificities of 0.72-0.85, enabling preemptive clinical adjustments[58,59]. Performance is typically evaluated using discrimination metrics, while calibration and decision curve analyses are less frequently reported despite their importance for balancing hypotension prevention against unnecessary intervention. Reinforcement-learning (RL)-style approaches have been explored to optimize ultrafiltration profiling and dry-weight estimation, showing reductions in hypotension events and improved fluid-management metrics in simulations and small pilot deployments[60]. These studies represent early exploratory or limited real-world implementations, with clinician-in-the-loop oversight required to mitigate safety and liability concerns.

In peritoneal dialysis (PD), ML models predict PD adequacy and peritonitis risk using nutritional indices, prior infection history, and catheter-related factors. Recent PD-specific models achieved AUROCs of 0.75-0.88 for early peritonitis prediction, supporting individualized surveillance intensity rather than uniform follow-up schedules[61] (Table 4). Multimodal AI models that integrate structured EHR data with imaging-derived biomarkers and molecular information further support treatment personalization by identifying high-risk subphenotypes. Radiomic features from CT or ultrasound-such as cortical thickness, parenchymal volume, and texture-based fibrosis surrogates-improve prognostic discrimination by approximately 0.03-0.07 in C-index when added to laboratory-only models and correlate with histologic fibrosis grade[32,62]. Molecular inputs, including polygenic risk scores, provide modest but meaningful reclassification for subsets of patients, with net reclassification improvements of about 4%-6% reported for identifying rapid progressors in some cohorts[63,64]. Despite these advances, most multimodal personalization models remain internally validated, with limited evaluation of bias related to ethnicity, socioeconomic status, or access to dialysis modalities. In practice, these models inform treatment intensity and monitoring frequency rather than directly dictating therapeutic choices. Consistent external validation and harmonization across centers remain prerequisites for routine deployment[65].

Table 4 Clinical decision-support applications of artificial intelligence across chronic kidney disease treatment domains.

Treatment domain	Clinical task	AI role	Performance range	Decision-support vs automation	Ref.
Pharmacotherapy	Nephrotoxicity risk	Monitoring and dose support	AUROC 0.72-0.86	Decision support	[52,53]
RAAS therapy	Hyperkalemia prediction	Monitoring intensity	AUROC 0.72-0.86	Decision support	[56]
Dialysis planning	Timing and access planning	Risk stratification	Improved time-dependent C-index	Decision support	[57]
Intradialytic care	IDH prediction	Preemptive adjustment	Sens. 0.78-0.88	Conditional automation	[58]
Peritoneal dialysis	Peritonitis risk	Surveillance tailoring	AUROC 0.75-0.88	Decision support	[61]

“Conditional automation” denotes algorithm-initiated actions requiring clinician confirmation. AI: Artificial intelligence; RAAS: Renin-angiotensin-aldosterone system; IDH: Intradialytic hypotension; AUROC: Area under the receiver operating characteristic curve.

From an implementation standpoint, personalized treatment models face substantial barriers, including clinician acceptance of algorithm-guided recommendations, integration into heterogeneous EHR and dialysis information systems, and unresolved medico-legal responsibility for adverse outcomes influenced by AI-supported decisions. These factors contribute to a persistent gap between promising retrospective performance and sustained real-world adoption.

AI FOR EARLY DETECTION AND MANAGEMENT OF CKD COMPLICATIONS

CVD remains the leading cause of morbidity and mortality in CKD, motivating AI-based approaches to detect upstream structural and clinical complications such as left ventricular hypertrophy (LVH), heart failure (HF)/congestive HF (CHF), and atrial fibrillation (AF) at earlier stages[66]. Most cardiovascular AI models in CKD are exploratory or internally validated, with relatively few undergoing rigorous external or prospective evaluation. In a single-center cohort of 8894 CKD patients with a composite CVD incidence of 25.9%, an XGBoost model integrating routine laboratory, demographic, and treatment variables achieved an AUC of 0.89 for predicting major cardiovascular events, with age, hypertension history, 24-hour urinary protein, and eGFR emerging as dominant predictors via SHAP analysis[67]. While discrimination was strong, calibration and decision curve analyses were not consistently reported, limiting interpretation of whether predicted risk thresholds translate into actionable net clinical benefit across care settings.

For specific arrhythmic outcomes, discrimination has been more modest: ML models predicting incident AF in CKD populations report C-indices around 0.67 when relying on standard clinical variables, indicating the need for enriched feature sets[67]. These modest values underscore the clinical limitation of relying on baseline discrimination alone, particularly in outcomes with relatively low prevalence, where class imbalance may inflate apparent performance in derivation cohorts. Exploratory studies incorporating radiomics and deep-learning features from echocardiography or cardiac magnetic resonance imaging show incremental gains in discrimination (absolute AUC/C-index improvements of about 0.02-0.07) for detecting LVH and subclinical systolic or diastolic dysfunction, but these findings are largely derived from single-center cohorts with limited external validation[67]. Bias related to differential access to advanced imaging, ethnicity, and socioeconomic status may further constrain generalizability, as imaging-enriched datasets often underrepresent resource-limited CKD populations. Collectively, cardiovascular AI models support dynamic risk monitoring and referral prioritization, though their clinical impact depends on integration into longitudinal care pathways.

Anemia management in CKD is well suited to AI-driven decision support because erythropoiesis-stimulating agent (ESA) dosing requires repeated titration in the setting of variable hemoglobin responses. Supervised ML and RL approaches have been developed to predict hemoglobin trajectories and recommend ESA dose adjustments[68]. Most anemia-focused models are internally validated using retrospective dialysis datasets, although anemia management represents one of the few CKD complication domains with emerging prospective evidence. Temporal aggregation of laboratory data markedly improves performance; for example, in a retrospective multicenter hemodialysis dataset (n = 1351), an XGBoost model using three months of continuous features achieved a weighted AUROC of 0.922 for classifying intact parathyroid hormone (iPTH) categories, illustrating the value of longitudinal features in closely related dosing tasks[69]. However, high AUROC values in such settings may partially reflect enriched high-risk populations and frequent laboratory sampling, which may not generalize to earlier-stage CKD or low-resource environments.

Importantly, clinical trial evidence is emerging. A recent randomized controlled trial comparing AI-assisted ESA dosing with standard clinician-managed dosing in maintenance hemodialysis reported non-inferior hemoglobin control with reduced ESA dose variability in the AI arm, suggesting that algorithmic support can stabilize anemia management without compromising safety[70]. This trial represents one of the few prospective, real-world evaluations of AI decision support in CKD complications, although it was conducted in a tightly controlled dialysis setting with structured workflows. Across cohort studies and early trials, AI-supported ESA strategies are associated with improved hemoglobin stability, fewer overshoot and undershoot events, and potential reductions in cumulative ESA exposure, though effect sizes vary by data completeness (e.g., iron indices, inflammatory markers) and calibration quality.

CKD-MBD presents a nonlinear control problem involving calcium, phosphate, PTH, and multiple interacting therapies. ML classifiers have been developed to identify CKD-MBD phenotypes and to predict PTH dynamics using routine biochemical panels[71]. In a derivation cohort of 116 subjects with independent prospective validation in 114 samples, ML models achieved training AUCs up to 0.91 for CKD-MBD classification, but validation AUCs declined to approximately 0.74 overall and to 0.58-0.63 for certain subtypes, highlighting overfitting risk and the importance of external testing[72]. These performance drops emphasize the need for calibration assessment and decision curve analysis, particularly when models are intended to guide therapy escalation or de-escalation.

More focused tasks show stronger performance. In a Taiwanese multicenter hemodialysis cohort (n = 1351), an XGBoost model using three months of longitudinal laboratory and medication data achieved a weighted AUROC of 0.922 for classifying iPTH categories (< 150, 150-599, ≥ 600 pg/mL)[69]. RL frameworks have been tested in silico to optimize dosing of phosphate binders, vitamin D analogs, and calcimimetics, demonstrating more consistent attainment of guideline targets compared with fixed heuristic protocols[73] (Table 5). Most RL-based CKD-MBD studies remain exploratory or simulation-based, with limited real-world deployment and unresolved questions regarding clinician acceptance and medico-legal accountability. Interpretable ML models have also been applied to predict hungry bone syndrome following parathyroidectomy with promising discrimination[74]. Explainability methods such as SHAP and attention-weight analysis have been used to identify dominant contributors (e.g., baseline PTH, alkaline phosphatase, dialysis vintage), supporting clinical interpretability and trust. Together, these approaches exemplify closed-loop monitoring and adjustment, wherein predictions are iteratively updated as biochemical responses evolve.

Table 5 Artificial intelligence approaches for prediction and management of major chronic kidney disease complications.

Complication domain	Clinical task	AI approach	Key performance	Care model	Ref.
Cardiovascular	Major CVD event prediction	XGBoost, radiomics	AUC 0.89; ΔAUC 0.02-0.07	Dynamic monitoring	[67]
Anemia	ESA dose support	ML, RL	Non-inferior Hb control; ↓ variability	Closed-loop support	[70]
CKD-MBD	PTH classification	XGBoost	AUROC 0.92	Closed-loop support	[69]
CKD-MBD	Multitarget dosing	RL (simulation)	Improved target attainment	Simulation-based	[73]

CKD-MBD: Chronic kidney disease-mineral and bone disorder; AUROC: Area under the receiver operating characteristic curve; ΔAUC: Change in area under the curve; PTH: Parathyroid hormone; ESA: Erythropoiesis-stimulating agent; ML: Machine learning; RL: Reinforcement learning; CVD: Cardiovascular disease; AI: Artificial intelligence; AUC: Area under the curve.

Across cardiovascular, anemia, and CKD-MBD applications, AI systems increasingly support dynamic monitoring and closed-loop care, combining prediction, recommendation, and feedback. However, important evidence gaps remain. Many studies rely on single-center or retrospective datasets, exhibit performance attenuation on external validation, or focus on surrogate endpoints rather than patient-centered outcomes. Bias related to ethnicity, socioeconomic status, and access to specialty care remains insufficiently examined, raising concerns about equitable deployment across CKD populations. Randomized controlled trials are still sparse outside anemia management, and few studies directly demonstrate reductions in hard outcomes such as cardiovascular events, hospitalization, or mortality. Additional challenges include assay heterogeneity (notably for PTH), variability in treatment protocols, incomplete reporting of calibration and decision-curve analyses, and practical barriers to clinical integration-including EHR interoperability, clinician acceptance, and medico-legal responsibility for AI-informed decisions. Multicenter prospective trials, standardized outcome definitions, and evaluation of workflow integration are therefore essential to establish the real-world effectiveness of AI-driven complication management in CKD[10,75] (Figure 4).

Figure 4 Closed-loop artificial intelligence-enabled management of chronic kidney disease complications. Conceptual feedback loop illustrating continuous data acquisition, prediction, intervention, and response updating to support dynamic and adaptive management of chronic kidney disease-related complications using artificial intelligence. AI: Artificial intelligence; IDH: Intradialytic hypotension; PTH: Parathyroid hormone; CKD-MBD: CKD-mineral and bone disorder; ESA: Erythropoiesis-stimulating agent; UF: Ultrafiltration.

AI INTEGRATION IN KIDNEY TRANSPLANTATION

AI applications across the kidney transplant continuum increasingly function as clinical decision-support tools, augmenting-rather than replacing-established allocation frameworks, pathology workflows, and post-transplant management. Most transplant-focused AI studies to date are registry-based and internally or externally validated retrospective analyses, with relatively few prospective or real-world implementation trials. ML models for donor-recipient matching and graft-outcome prediction synthesize high-dimensional donor and recipient features to estimate short- and long-term transplant risk[76]. Registry-based studies show that ML approaches (gradient boosting, random forests, neural survival models) improve discrimination compared with conventional risk tools. In a United Kingdom cohort of 36653 accepted transplant offers, a neural survival model achieved concordance indices of 0.63 for graft failure and 0.79 for patient death, with AUROCs improving at longer follow-up horizons[77]. While these metrics indicate improved ranking performance, calibration and decision-curve analyses are variably reported, limiting assessment of clinical net benefit when risk thresholds are used to influence acceptance decisions. ML-based ranking and acceptance algorithms outperform simple similarity metrics in identifying high-risk offers and can simulate acceptance trade-offs before allocation decisions[76].

Incorporation of immunologic mismatch metrics-such as eplet- or hydrophobic-mismatch scores-further refines risk estimation beyond traditional antigen matching. National registry analyses show that integrating eplet-level metrics meaningfully reclassifies immunologic risk for a substantial proportion of offers and improves prediction of alloimmune events and graft failure[78,79]. These studies are largely externally validated using large registries, but remain observational and retrospective in nature. However, these gains raise ethical and equity considerations: Algorithms that optimize projected graft survival may inadvertently disadvantage certain demographic groups unless explicitly constrained to preserve equitable access. Bias related to ethnicity, socioeconomic status, and differential access to transplantation evaluation may be encoded in registry data, necessitating fairness-aware modeling and subgroup calibration analyses. The literature increasingly emphasizes the need for transparent simulation of trade-offs between immunologic risk and equitable allocation[80].

During the peri-transplant period, AI has been applied to assess donor organ quality and early rejection risk, particularly when rapid or standardized interpretation is required. A large international study developed a “virtual day-zero biopsy” ensemble using routinely collected donor parameters and 14032 biopsies from 17 centers, successfully predicting Banff lesion scores and providing a practical surrogate when physical biopsies are unavailable[81]. This approach represents an externally validated, real-world dataset, although outcome impact on graft survival has not yet been prospectively tested. DL systems applied to whole-slide images of renal allograft biopsies further support peri- and early post-transplant decision-making. A multi-instance learning model achieved an AUC of 0.798 for three-way rejection classification (T cell-mediated rejection, antibody-mediated rejection other) and an AUC of 0.936 for predicting graft loss within one year after rejection in independent testing[82]. Earlier multicenter studies similarly demonstrate that DL-based pathology tools approach or exceed inter-observer pathologist agreement for selected tasks[83,84]. Attention maps and patch-level attribution methods are increasingly used to localize histologic features driving predictions, supporting interpretability and clinical trust. Clinically, these systems aim to standardize interpretation and accelerate therapeutic decisions, rather than automate diagnosis.

Post-transplant pharmacologic management-particularly tacrolimus dosing-represents one of the most mature AI applications in transplantation. Time-series ML models, including LSTM networks and gradient-boosted regressors, trained on sequential dose-concentration data and laboratory covariates outperform linear regression and many population pharmacokinetic models in external validation[85]. Most tacrolimus models are externally validated but remain retrospective, with limited prospective deployment. In a multicenter cohort of 443 transplant recipients (6264 tacrolimus samples), an LSTM model achieved a median performance error of 8.8% and median absolute performance error of 22.3%, outperforming PK-based approaches; patients receiving doses outside model-suggested ranges experienced longer intensive care unit stays[86]. Comparable performance gains have been reported in renal transplant cohorts, with improved prediction of therapeutic-range trough concentrations[87,88]. Despite favorable accuracy metrics, clinical readiness depends on calibration stability across centers and avoidance of overconfidence in settings with sparse or delayed drug-level sampling. These systems function as dose-adjustment decision support, requiring clinician review rather than autonomous dosing.

AI-enabled remote monitoring represents an emerging adjunct to post-transplant care. Wearable sensors and smartphone-based platforms paired with ML analytics can detect deviations in physiologic signals, surrogate markers of renal function, and adherence behaviors between clinic visits[89]. Most studies in this domain are exploratory or pilot-level, with limited randomized or prospective outcome data. Reviews of telemonitoring interventions report improved short-term adherence and engagement, though randomized evidence for reductions in rejection or graft loss remains limited[90] (Table 6). Access-related bias is a key concern, as digital monitoring tools may preferentially benefit patients with higher socioeconomic resources or digital literacy. These tools support early detection and closed-loop follow-up, but require careful validation, data-security safeguards, and integration into clinical workflows[90,91].

Table 6 Applications of artificial intelligence across the kidney transplantation continuum.

Phase	Clinical task	AI approach	Key performance	Clinical role	Ref.
Pre-transplant	Allocation and offer ranking	ML survival, ranking models	C-index 0.63-0.79	Decision support	[77]
Pre-transplant	Immunologic risk	Eplet mismatch ML	Risk reclassification	Equity-aware support	[78]
Peri-transplant	Biopsy assessment	DL (WSI)	AUC up to 0.94	Diagnostic support	[82]
Post-transplant	Tacrolimus dosing	LSTM, PK-ML	↓ Dosing error	Dose support	[86]
Post-transplant	Remote monitoring	Wearables + ML	Improved adherence	Surveillance support	[90]

LSTM: Long short-term memory; PK-ML: Pharmacokinetic-machine learning; DL (WSI): Deep learning applied to whole-slide images; AI: Artificial intelligence; ML: Machine learning; AUC: Area under the curve.

Despite promising performance, translational challenges persist. Performance decay across centers highlights the need for standardized data curation, assay harmonization (notably histology staining and tacrolimus assays), and prospective implementation studies[92]. Ethical considerations are particularly salient in allocation modeling: Immunogenicity-aware algorithms must be explicitly designed to avoid amplifying disparities while maintaining transparency and clinician oversight[80]. From an implementation standpoint, barriers include clinician acceptance, EHR interoperability across transplant centers, medico-legal responsibility for AI-informed allocation or dosing decisions, and the absence of clear regulatory guidance for adaptive models. Explainability methods (e.g., SHAP, attention maps) and hybrid clinician-AI workflows are widely recommended to support trust and safe deployment[93,94]. Overall, AI shows substantial promise across pre-, peri-, and post-transplant phases, but widespread clinical benefit will depend on rigorous prospective validation, equity-aware design, and carefully managed real-world integration.

EXPLAINABILITY, BIAS, AND ETHICAL CONSIDERATIONS IN AI FOR CKD

Safe use of AI in CKD requires adherence to three interlinked principles: Equity-aware design, calibration and transportability, and transparent, accountable deployment. Across the CKD literature, most bias and explainability analyses remain exploratory or internally validated, with relatively few studies incorporating external validation or prospective assessment of equity impacts in real-world workflows. Violations of any principle have CKD-specific consequences because algorithmic outputs directly inform referral timing, monitoring intensity, medication dosing, dialysis planning, and transplant eligibility. Algorithmic bias arises when models encode historical inequities or rely on proxies (e.g., health-care utilization) that correlate imperfectly with physiologic need. The seminal demonstration by Obermeyer et al[95] showed that a widely deployed risk algorithm prioritized patients based on predicted costs rather than illness severity; at identical risk scores, Black patients were substantially sicker than White patients, and correcting the proxy increased identification of Black patients for additional care from 17.7% to 46.5%[95]. Most nephrology AI studies addressing bias operate at the exploratory or retrospective stage, underscoring the gap between conceptual fairness frameworks and validated clinical deployment. In nephrology, analogous mechanisms are plausible when CKD risk tools or operational rules depend on utilization-linked variables, potentially delaying referral, access creation, or intensified monitoring for disadvantaged groups.

Subgroup error imbalance further compounds these risks. Analyses of diagnostic AI consistently show higher false-negative rates for under-represented populations; Seyyed-Kalantari et al[96] documented systematic disparities across sex and skin-tone strata, illustrating how aggregate performance metrics can mask clinically meaningful subgroup harm[96]. In CKD datasets, similar imbalances may arise from differential access to laboratory testing, delayed presentation, and socioeconomic gradients in care utilization, all of which contribute to class imbalance and early-disease underrepresentation. In CKD, elevated false-negative rates translate directly into missed opportunities for early nephrology referral, anemia or CKD-MBD intervention, and timely transplant evaluation.

The debate surrounding race in kidney function estimation exemplifies how technical modeling decisions affect equity. Removal of the race coefficient from the CKD-EPI creatinine equation (2021 refit) resulted in reclassification of many patients’ eGFR, altering eligibility around clinically salient thresholds[97]. Comparative evaluations report mixed effects on mean bias but consistent evidence of threshold reclassification with downstream implications for medication dosing, nephrology referral, and transplant listing[98,99]. This experience highlights that variable selection decisions-whether inclusion of race, socioeconomic proxies, or access-related features-must be evaluated using clinical impact analyses (e.g., decision-curve analysis) rather than discrimination metrics alone. This case underscores that inclusion or exclusion of demographic variables is not value-neutral and must be evaluated in terms of clinical consequences, not solely statistical fit.

Because CKD predictors and care pathways vary across populations and health systems, calibration and external validation are as critical as discrimination. Multiple CKD models demonstrate strong internal performance yet miscalibrate when transported. External validations of the KFRE repeatedly show preserved discrimination but variable calibration; in a national cohort, Bravo-Zúñiga et al[100] reported poor calibration requiring recalibration to yield clinically usable absolute risks. By contrast, recent ensemble ML models that explicitly report calibration metrics achieved near-ideal calibration slopes (0.96, 95%CI: 0.94-0.98) after careful subgroup assessment[5]. Few studies, however, report decision-curve analyses to quantify net clinical benefit across risk thresholds, limiting interpretation of whether improved calibration translates into safer or more equitable care decisions. These differences are consequential: Inaccurate absolute risk estimates can misallocate surveillance resources, delay dialysis preparation, or inappropriately prioritize patients for advanced therapies.

Explainability is a practical mechanism for bias detection and clinician oversight. Systematic reviews of XAI show that post-hoc methods (SHAP, LIME, attention maps) and interpretable models improve transparency and clinician trust, and can reveal reliance on non-physiologic proxies or subgroup miscalibration[101,102]. Quantitatively, recent surveys report that 30%-60% of clinical decision-support studies now include explainability analyses, with higher clinician acceptability and greater likelihood of prospective evaluation when XAI is present[103]. In nephrology-specific applications, SHAP analyses frequently identify longitudinal eGFR slope, albuminuria burden, and anemia markers as dominant contributors, providing face-valid explanations that support clinician review prior to action. In CKD, explainability supports case-level review before automated action (e.g., referral triggers, dose recommendations), aligning AI use with shared decision-making.

Equitable AI deployment also depends on secure, interoperable infrastructure. Privacy-preserving approaches-including federated learning, secure multiparty computation, and differential privacy-reduce centralized data sharing while approaching centralized performance; reported accuracy gaps typically range from 0%-5% depending on data heterogeneity and aggregation strategy[104,105] (Table 7). Most CKD federated-learning studies remain proof-of-concept or internally validated, with limited demonstration of sustained performance or calibration stability in prospective, multi-institutional deployments. However, AI systems introduce cybersecurity risks (model inversion, membership inference, supply-chain vulnerabilities) that require continuous governance[106]. Interoperability standards such as Substitutable Medical Applications, Reusable Technologies on Fast Healthcare Interoperability Resources (FHIR) and CDS Hooks enable EHR integration, yet reviews estimate that reliable FHIR API availability exists in only a minority of health systems globally, with custom integration accounting for 40%-60% of deployment cost and time[107,108]. Beyond technical barriers, real-world implementation is constrained by clinician acceptance, medico-legal liability concerns surrounding algorithm-informed decisions, and uncertainty regarding accountability when AI recommendations conflict with standard practice. These constraints are material determinants of whether AI can be deployed safely and equitably in routine CKD care.

Table 7 Ethical, methodological, and implementation principles for artificial intelligence deployment in chronic kidney disease care.

Principle	CKD-specific risk	Quantitative evidence	Mitigation
Equity-aware design	Delayed referral, under-monitoring	17.7%→46.5% care reclassification after bias correction	Fairness audits, proxy review
Calibration	Misallocation of resources	KFRE miscalibration on external validation	Recalibration, subgroup testing
Explainability	Hidden proxy reliance	30%-60% studies include XAI	SHAP/LIME, clinician oversight
Privacy and integration	Limited deployment	0%-5% FL accuracy gap; 40%-60% integration cost	Federated learning, FHIR/CDS hooks

KFRE: Kidney failure risk equation; SHAP/LIME: SHapley Additive exPlanations/Local interpretable model-agnostic explanations; FHIR/CDS: Fast healthcare interoperability resource/clinical decision support; FL: Federated learning; CKD: Chronic kidney disease; XAI: Explainable artificial intelligence.

IMPLEMENTATION BARRIERS AND REAL-WORLD DEPLOYMENT

Despite strong retrospective performance, most AI models in nephrology fail to translate into sustained clinical impact. Across the CKD literature, the majority of deployed systems remain exploratory or internally validated, with only a small minority progressing to externally validated, prospective, or real-world implementations. Evidence from implementation science indicates that failure is rarely due to algorithm design alone; rather, it reflects interacting technical, organizational, and regulatory barriers that degrade performance, delay adoption, or prevent scaling.

The dominant technical failure mode is inadequate validation. Systematic reviews show that fewer than 10%-20% of published medical AI models undergo external validation, and fewer than 5%-10% are tested prospectively or in randomized comparisons with clinicians[109,110]. Most CKD-focused AI studies report internal AUROC values without parallel reporting of calibration metrics or decision-curve analyses, limiting interpretation of clinical utility beyond discrimination. Even when internal discrimination is high, transport to new settings commonly results in material performance decay: External validations report median AUROC reductions of about 0.05-0.15, with some models falling from > 0.90 in derivation to < 0.80 externally[111,112]. In nephrology, such degradation is amplified by site-level heterogeneity in laboratory assays, eGFR estimation practices, case-mix, and data completeness, leading to miscalibration and unreliable absolute risk estimates unless recalibration or refitting is performed. Class imbalance and low prevalence of early-stage CKD in routine care further inflate apparent performance in derivation cohorts, contributing to systematic overestimation of readiness for deployment. These technical limitations explain why many high-performing CKD models fail to influence outcomes when deployed beyond their development environment.

Even technically sound models frequently fail because they are poorly embedded in clinical workflows. Implementation studies consistently show that EHR integration, user-interface customization, workflow redesign, and clinician training dominate deployment effort; 40%-60% of total project time and cost is attributable to interoperability engineering and change management rather than model development[107,108,113]. Limited adoption of interoperability standards (e.g., incomplete FHIR APIs, fragmented laboratory and imaging systems, legacy EHR constraints) further restricts real-time data flow and reduces coverage across patient populations[114,115]. From a clinical perspective, this deployment gap manifests as delayed or inconsistent alert delivery, limited coverage of high-risk populations, and erosion of clinician trust when AI outputs are perceived as unreliable or poorly contextualized. Clinically, these organizational frictions result in delayed rollouts, inconsistent alert delivery, and low clinician engagement-undermining the potential benefits of early detection or dynamic monitoring in CKD. Human factors compound these challenges. Surveys indicate that many clinicians defer reliance on AI outputs until locally validated evidence, explainability, and governance structures are in place[116]. Concerns regarding medico-legal liability-particularly accountability when AI-informed recommendations conflict with clinician judgment or guidelines-remain a major barrier to adoption. Conversely, implementation programs that incorporate clinician co-design, explicit accountability, and operational KPIs (e.g., alert response rates, proportion of patients rescored within 24 hours, referral-rate changes) report measurable improvements in process metrics during rollout[117].

Regulatory requirements impose additional constraints that shape real-world feasibility. In the United States, the Food and Drug Administration’s framework for AI/ML-based Software as a Medical Device mandates predefined change-control plans, transparency around model updates, and post-market real-world performance monitoring, substantially increasing surveillance and reporting burdens[118,119]. Few nephrology AI tools have yet demonstrated sustained compliance with these requirements in prospective clinical use. In the European Union, the AI Act (in force August 2024) classifies many medical AI tools as high-risk systems, requiring enhanced data governance, conformity assessment, and technical documentation, with phased applicability extending over several years[120]. Regulatory readiness analyses suggest that compliance can add months to years of development time and materially increase costs, factors that must be incorporated into translational planning and cost-effectiveness evaluation.

Implementation barriers are magnified in low-and middle-income countries (LMICs). Economic evaluations of clinical AI show mixed cost-effectiveness, with favorable results more likely in high-prevalence, low-cost screening contexts and less consistent benefit for specialized diagnostics requiring substantial infrastructure[121,122]. World Health Organization and multi-stakeholder assessments identify infrastructure (power reliability, broadband, cloud access), governance, and workforce capacity as critical enablers, estimating that significant upfront investment is required before AI can be reliably scaled[123]. Socioeconomic gradients in access to care and laboratory testing further complicate deployment, increasing the risk that AI tools preferentially benefit populations already well served by health systems. Models developed in data-rich high-income settings often perform poorly when transplanted without retraining on local data or adaptation to constrained laboratory and EHR environments, increasing the risk of performance degradation and inequitable benefit distribution. For CKD care in LMICs, practical deployment therefore favors simpler, robust models aligned with available assays and staged implementation coupled with local validation (Table 8).

Table 8 Key barriers to clinical implementation of artificial intelligence in chronic kidney disease and proposed mitigation strategies.

Barrier category	Failure mechanism	Quantitative evidence	Mitigation
Technical	Overfitting, drift, miscalibration	AUROC drop 0.05-0.15 on external validation	Multisite validation, recalibration
Organizational	Poor EHR/workflow integration	40%-60% of deployment cost/time	FHIR-based integration, co-design
Regulatory	Extended approval timelines	Months-years added to development	Early regulatory planning
LMIC context	Infrastructure and data gaps	Limited cost-effectiveness evidence	Local retraining, staged rollout

AUROC: Area under the receiver operating characteristic curve; LMIC: Low- and middle-income countries; EHR: Electronic health records; FHIR: Fast healthcare interoperability resources.

Across settings, the evidence indicates that AI models fail in practice not because prediction is impossible, but because external validation, workflow integration, regulatory compliance, and local context adaptation are insufficiently addressed. Bridging the deployment gap will require a shift from performance-centric reporting toward evidence of prospective effectiveness, calibration stability, decision-curve-informed clinical benefit, and sustained clinician adoption in real-world CKD workflows. Overcoming these barriers is feasible but requires explicit planning for multisite validation, interoperability engineering, regulatory evidence generation, and context-specific cost-effectiveness-particularly when extending AI-enabled CKD care beyond well-resourced health systems (Figure 5).

Figure 5 Barriers to translating artificial intelligence models into clinical chronic kidney disease care. Overview of technical, organizational, regulatory, equity, and infrastructure barriers that limit the progression of artificial intelligence models from development to real-world clinical impact in chronic kidney disease care. AUROC: Area under the receiver operating characteristic; AI: Artificial intelligence; FDA: United States Food and Drug Administration; EU: European Union; LMIC: Low-and middle-income countries.

FUTURE DIRECTIONS

The trajectory of AI in nephrology is best framed around a small number of high-leverage directions with clear time horizons, grounded in current empirical progress rather than speculative capability. Most work informing these directions remains exploratory or internally validated, with only limited externally validated or prospective evidence, underscoring the importance of cautious interpretation. In the near term, the most plausible advance is adaptation of medical foundation models to kidney care. Large multimodal architectures trained via self-supervision on structured EHR data, laboratory time series, imaging, and clinical text can provide transferable representations for downstream nephrology tasks without extensive task-specific labeling[124]. Systematic reviews document rapid growth in multimodal medical AI research, with over 1100 screened studies in medical imaging alone, supporting feasibility of such approaches for segmentation, risk prediction, and decision support[125]. To date, reported performance gains are largely expressed as improvements in discrimination (e.g., AUC or C-index), while calibration and decision-curve analyses are inconsistently reported, limiting assessment of clinical utility across risk thresholds. Applied to nephrology, these models are likely to first enhance consistency and calibration across CKD risk stratification, dialysis planning, and treatment response prediction, rather than replace existing task-specific models.

Another near-term direction is integration of AI with wearables and internet of things platforms for higher-frequency physiologic monitoring. Digital health reviews in CKD highlight real-time tracking of blood pressure, fluid status, and related physiologic signals as immediately actionable data streams[126]. Most studies in this domain are exploratory or based on short-term internal validation, with sparse prospective evaluation of downstream outcomes such as hospitalization or progression rates. When combined with validated predictive models, these systems can support earlier detection of clinically meaningful deviations between clinic visits, addressing a well-recognized gap in episodic CKD care. However, low event prevalence outside high-risk subgroups and class imbalance in continuous monitoring data may inflate apparent discrimination unless calibration and decision-curve performance are explicitly assessed.

Natural language processing and large language models are increasingly applied to patient narratives and clinical communication. Scoping reviews show that approximately 41% (17/41) of recent AI studies in CKD address unstructured text or patient-centered communication domains[7]. Most reported applications remain at the internally validated or pilot stage, with performance typically summarized using accuracy or AUC rather than clinically interpretable utility metrics. In the near term, incorporating patient-reported outcomes such as symptom burden and quality of life alongside biologic metrics is likely to improve clinical relevance and shared decision-making without requiring fundamentally new model architectures. Explainable natural language processing methods, including attention visualization and SHAP-style token attribution, may be particularly important to support clinician trust and mitigate bias related to language, literacy, or socioeconomic context.

Over longer horizons, AI-driven digital twins represent a more transformative but technically demanding paradigm. Early healthcare prototypes demonstrate patient-specific computational replicas capable of simulating disease trajectories and virtual interventions; one CKD-focused implementation used a generalized metabolic flux digital twin to stratify 3-year risk in simulated scenarios[7]. Current evidence for digital twins in CKD is exploratory and simulation-based, with no prospective clinical validation to date. If validated prospectively, such systems could support in silico evaluation of therapeutic strategies including dietary modification, pharmacologic intensification, and timing of dialysis initiation, shifting care from reactive adjustment toward anticipatory planning. Robust calibration, transparency of assumptions, and clear delineation of medico-legal responsibility will be prerequisites for any clinical deployment.

Long-term progress is also expected from integration of multi-omics data with clinical phenotypes to refine disease mechanisms and therapeutic targeting. Machine-learning-based multi-omics integration has been proposed as a mean to identify biomarkers of progression, stratify patients for targeted therapies, and uncover modifiable pathways in complex kidney disease biology[127]. Most nephrology-focused multi-omics studies remain exploratory, often limited by small sample sizes, population homogeneity, and incomplete external validation, which constrain generalizability. While quantitative evidence of clinical impact in nephrology remains limited, advances in data scale and harmonization suggest that such approaches may contribute to precision therapeutics over the next decade (Table 9).

Table 9 Anticipated future directions of artificial intelligence in chronic kidney disease care across time horizons.

Time horizon	Direction	Evidence base	Primary clinical impact
5 years	Multimodal foundation models	Rapid growth in medical multimodal AI[125]	Improved calibration and consistency
	Wearables and continuous monitoring	Digital CKD health reviews[126]	Earlier detection between visits
	PRO and text integration	41% of CKD AI studies use unstructured data[7]	Patient-centered decision support
≥ 10 years	Digital twins	Early CKD prototype models[7]	In silico therapy planning
	Multi-omics AI	Precision medicine frameworks[127]	Target discovery and stratification

PRO: Patient-reported outcomes; CKD: Chronic kidney disease; AI: Artificial intelligence.

Across time horizons, the unifying trend is a shift from isolated, task-specific models toward integrated systems that combine multimodal data, longitudinal monitoring, and patient input. Near-term gains are most likely in calibration, consistency, and patient-centered decision support; longer-term advances depend on prospective validation of digital twins and multi-omics-driven therapeutic insights. Bridging the deployment gap will require explicit attention to evidence maturity, calibration across diverse populations, explainability to support clinician acceptance, and integration into existing clinical and regulatory frameworks. Realizing these futures will require rigorous external validation, attention to regulatory and ethical constraints, and demonstration of clinical benefit across diverse populations.

To translate these directions into practice, future research should prioritize multicenter prospective trials with prespecified clinical endpoints, coordinated data harmonization initiatives to support external validation, and adherence to emerging reporting standards for AI in healthcare to enable reproducibility and regulatory assessment.

CONCLUSION

AI has moved beyond proof of concept in nephrology and is already demonstrating tangible value across the CKD continuum. Across large registries, longitudinal EHRs, and imaging datasets, AI models consistently outperform conventional rule-based approaches in early CKD detection, short- and long-term risk stratification, complication monitoring, and decision support for pharmacotherapy, dialysis planning, and transplantation. These systems are particularly effective at integrating longitudinal trajectories and multimodal data, enabling earlier identification of high-risk patients and more individualized, data-driven clinical decisions than static thresholds or single-timepoint risk scores.

However, routine clinical adoption requires several conditions to be met. Models must undergo rigorous external and prospective validation across diverse populations and care settings, with transparent reporting of calibration, subgroup performance, and false-positive trade-offs. Equity considerations must be addressed explicitly to avoid reinforcing disparities in referral, access to advanced therapies, or transplantation. In parallel, AI tools must be embedded within interoperable, clinician-centered workflows that preserve clinical oversight, support interpretability, and demonstrate measurable benefit on patient-centered outcomes, not only predictive accuracy. Regulatory clarity, implementation science, and cost-effectiveness evaluations will be essential to ensure safe, scalable deployment.

The central take-home message for clinicians and policymakers is clear: AI is already capable of enhancing nephrology care, but its value will only be realized if development shifts from isolated performance optimization toward validated, equitable, and workflow-integrated clinical systems. Progress now depends on deliberate investment in prospective, multicenter evaluations; shared data infrastructures that enable external validation; and standardized reporting frameworks that align technical performance with clinical decision-making and regulatory expectations. Strategic investment in prospective trials, data infrastructure, and governance frameworks will determine whether AI becomes a routine and trusted component of CKD care or remains confined to experimental use.