Risk stratification in diabetic kidney disease: a review of prediction models for methodological advances and clinical application

Abstract

Background

Diabetic kidney disease (DKD) represents the leading cause of end-stage renal disease (ESRD) worldwide, characterized by a complex pathophysiology and heterogeneous progression. Accurate prediction of the onset, progression, and adverse outcomes of DKD is critical for early intervention and personalized management.

Main body

This review systematically summarizes the current research on prediction models in DKD, encompassing both diagnostic and prognostic models. It discusses key methodological considerations in model development and validation, with a specific focus on the application of machine learning (ML) techniques in model construction. Furthermore, this article also evaluates the performance of prediction models based on routine clinical parameters and multimodal models integrating multi-omics, imaging, retinal parameters, and renal pathological features. The primary challenges in clinical translation are analyzed, and future directions for optimizing DKD prediction are proposed.

Conclusions

In summary, advancing the optimization and clinical translation of DKD prediction models holds significant potential to improve patient care. Future research should focus on addressing the existing challenges, aiming to advance risk-stratified and personalized management and inform future precision medicine approaches in nephrology.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-026-07746-6.

Introduction

As one of the most common microvascular complications of diabetes, diabetic kidney disease (DKD) exhibits epidemiological characteristics marked by both high prevalence and insidious onset. It is reported that approximately 40% of diabetic patients will develop DKD, and a considerable proportion of these individuals—particularly those without effective glycemic control—will progress to end-stage renal disease (ESRD), imposing a substantial healthcare burden on families and society [1]. In current clinical practice, the diagnosis and management of DKD primarily rely on indicators such as albuminuria and glomerular filtration rate. However, these traditional biomarkers suffer from limitations including non-specificity and hysteresis, making it difficult to achieve individualized risk stratification and early intervention for DKD [2].

Precision medicine is a modern medical model centered on individualized health care, which aims to tailor diagnosis, treatment, and prevention strategies to match patients based on individual differences such as genetics, environment, and lifestyle [3]. Among its priorities, enhancing the accuracy of predicting diabetes-related outcomes is a key focus of precision medicine by facilitating risk-stratified and personalized management, and also provides a theoretical foundation for the subsequent construction of prediction models [4]. The pathophysiological landscape of DKD evolves over time [5]. Early stages are primarily driven by hemodynamic alterations and direct metabolic injuries. As the disease advances, these initial insults instigate a state of chronic inflammation and activate profibrotic signaling pathways, which subsequently become the principal mechanisms responsible for the irreversible renal functional decline and structural damage characteristic of late-stage DKD. This mechanistic shift necessitates prediction tools capable of capturing distinct biological processes across the disease continuum [6]. Specifically in the context of DKD, prediction models can be categorized into diagnostic and prognostic types according to their application scenarios and objectives. Diagnostic prediction models are dedicated to the early identification of subclinical DKD in asymptomatic populations, providing a solid basis for early detection and timely intervention of the disease [7]. In contrast, prognostic prediction models aim to assess the disease progression in patients already diagnosed with DKD, thereby enabling more intensive medical interventions to control risk factors and prevent related complications [8].

The review aims to summarize the methodological approaches in the development of DKD prediction models, including the application of artificial intelligence (AI) technologies such as machine learning (ML) in model construction. Simultaneously, it comprehensively evaluates the performance of existing diagnostic and prognostic models and analyzes the utility of multimodal models integrating multi-omics, imaging, retinal parameters, and renal pathology. Furthermore, this article delves into the challenges faced in the clinical translation of these models and proposes key directions for future research.

Methodological approaches for development and validation of DKD prediction models

Compared to non-diabetic populations, patients with diabetes have a higher risk of developing chronic kidney disease (CKD) [9]. Although DKD patients share common features of impaired renal structure and function, there is significant heterogeneity in clinical manifestations and disease progression trajectories among individuals. This is because the onset and progression of DKD involve multiple pathophysiological mechanisms, such as hemodynamic abnormalities in the kidneys caused by dyslipidemia and hypertension, inflammation-mediated renal tissue damage, and renal parenchymal remodeling triggered by fibrosis [10]. By developing diagnostic and prognostic prediction models, patients can obtain accurate risk assessments regarding whether they have DKD and its potential future progression. DKD diagnostic models can utilize current clinical indicators, lifestyle factors, environmental exposures, background information, or novel biomarkers to evaluate patients suspected of having DKD [11]. In contrast, DKD prognostic models rely on baseline or historical indicators such as demographic data, comorbidities, or treatment history to predict the risk of specific clinical outcomes in patients already diagnosed with DKD [8]. Both types of models collectively aim to achieve individualized risk stratification and decision-making, thereby improving patient outcomes (Table 1 and Fig. 1). In real-world settings, integrating DKD diagnostic and prognostic prediction models into clinical practice involves stages such as model development, performance evaluation, translation into clinical tools, and assessment of application impact [12].

Table 1.

Comparison of diagnostic and prognostic models for DKD

Feature dimension	Diagnostic models	Prognostic models
Core question	Determine whether diabetic patients currently have DKD	Predict the risk of future disease progression or complications in patients with DKD
Temporal rientation	Cross-sectional: A one-time assessment of the current or future status	Longitudinal: Tracks outcomes over time starting from a specific point
Study design	Cross-sectional study	Cohort study (prospective or retrospective)
Study population	Patients suspected of having DKD	Patients with confirmed DKD
Model types	ML algorithms such as LR, RF, SVM, DT, KNN, and XGB	CPH model, ML survival models, competing risk models, etc
Model output	Probability of having DKD	Risk of specific future outcomes such as ESRD, CVD, or death
Predictors	Current indicators: • Clinical indicators • Lifestyle • Environmental exposures • Background information • Biomarkers	Baseline or historical indicators: • Baseline clinical indicators • Demographic indicators (e.g., sex, age) • BP and glycemic control status • Complications (e.g., CVD) • History of medication use
Study outcome	• Pathological: Kidney biopsy • Clinical: UACR > 30 mg/g and/or decreased eGFR	• Progression to ESRD • 40% or greater decline in eGFR from baseline • Cardiovascular events • All-cause mortality
Model validation	• Discrimination: ROC curves • Calibration: Calibration plot• Clinical utility: DCA	• Discrimination: C-index • Calibration: Time-dependent calibration plot• Clinical utility: DCA
Advantages	• Non-invasive screening • Improves early diagnosis • Facilitates immediate decision-making	• Individualized risk stratification • Guides intensive therapy • Predicts outcome events
Limitations	• Difficulty obtaining gold standard reference • Fluctuation of indicators	• High cost and long duration of long-term follow-up • Competing risks: DKD patients often die from CVD rather than ESRD • Changes in treatment regimen: Widespread use of drugs alters the natural history of the disease

Abbreviation: DKD, diabetic kidney disease; LR, logistic regression; RF, random forest; SVM, support vector machine; DT, decision tree; KNN, k-nearest neighbors; XGB, extreme gradient boosting; CPH, cox proportional hazards; ESRD, end-stage renal disease; CVD, cardiovascular disease; UACR, urine albumin-to-creatinine ratio; eGFR, estimated glomerular filtration rate; ROC, receiver operating characteristic; DCA, decision curve analysis

Fig. 1.

Precision medicine workflow. Through prediction models that integrate individual demographic information and biological factors—including routine clinical data, laboratory tests, genetic information, and non-routine omics biomarkers—along with lifestyle, environmental, and contextual information, a risk assessment for diabetic kidney disease is performed. This supports subsequent clinical decision-making and personalized intervention planning (by Figdraw, ID = AYUOPa4444)

Model development

During the model development phase, the selection and determination of the study population, predictors, clinical outcomes, and prediction time horizon for both DKD diagnostic and prognostic models must align with the model’s intended use. This ensures that the study population and setting are consistent with the characteristics of the target population for model application [13]. The identification of candidate predictors should balance predictive capability with model simplicity, while also considering their availability in the intended application scenario. The prediction time horizon should be logically connected to potential interventions, such as timely treatment following diagnosis or intensified management after prognostic stratification [13].

Diagnostic prediction of DKD involves a one-time assessment at a specific time point. Utilizing a cross-sectional design and targeting patients suspected of having DKD, it employs current indicators such as urine albumin-to-creatinine ratio (UACR), eGFR, and duration of diabetes. By applying various ML algorithms, it outputs the probability of having DKD, enabling non-invasive screening and immediate decision-making to provide timely intervention evidence [7, 14]. In contrast, prognostic models for DKD focus on longitudinally tracking clinical outcomes. Based on prospective or retrospective cohort studies and targeting confirmed DKD patients, they integrate baseline or historical indicators such as clinical parameters, complications, or medication history. Using cox proportional hazards (CPH) models, machine learning survival models, and competing risk models, they calculate the risk of future events such as ESRD, cardiovascular events, or all-cause mortality, providing theoretical support for individualized risk stratification, predicting clinical outcomes, and guiding intensive therapy [15–17].

Precision diagnosis and prognosis prediction of DKD differ from approaches that simply classify patients based solely on physiological variables. In recent years, some studies have attempted to subtype newly diagnosed diabetic patients, categorizing individuals with similar phenotypes and potentially observing differences in DKD incidence across subtypes to identify high-risk subgroups. Nevertheless, this approach remains essentially a classification strategy at the population level [18]. In comparison, DKD diagnostic and prognostic models can directly calculate an individual’s current probability of having the disease or future risk of progression based on multidimensional indicators, generally demonstrating superior predictive performance compared to such grouping methods [19].

Model validation and evaluation

The evaluation of discrimination, calibration, and clinical utility constitutes critical steps in assessing model performance after development. For diagnostic prediction models, discrimination reflects the ability to distinguish between patients who currently “have DKD” and those who “do not have DKD”. The performance of a model needs to be evaluated by a set of metrics: sensitivity (true positive rate), specificity (true negative rate), precision (positive predictive value, i.e., the proportion of cases diagnosed as positive by the model that are truly diseased), and accuracy (the proportion of all correctly predicted cases among the total cases). The receiver operating characteristic (ROC) curve intuitively demonstrates a model’s trade-off between capturing true positives and avoiding false positives. It plots the relationship between sensitivity and (1-specificity) across all possible decision thresholds. The area under the curve (ROC-AUC) provides a single summary metric for the model’s overall discrimination ability. However, the core of applying the ROC curve to clinical practice lies in the selection of a specific decision threshold, which must be closely aligned with the specific clinical objectives of DKD management. In different application scenarios, the tolerance for false positives (misdiagnosis) and false negatives (missed diagnosis) varies drastically [20]. For instance, in scenarios of early screening or preliminary screening of high-risk populations, the primary goal is often to avoid missing potential patients as much as possible. In such cases, clinical strategies tend to prioritize ensuring high sensitivity. This implies that a relatively high false positive rate is acceptable, since positive results can be verified through more precise examinations in subsequent steps. Conversely, in decision-making scenarios for definitive diagnosis or initiation of treatment, the goal shifts to ensuring that treatment is administered only to patients who are highly likely to benefit, thereby avoiding unnecessary medication use and potential adverse effects. In such cases, clinical strategies tend to prioritize ensuring high specificity or high precision. This means that a more stringent threshold is required to improve diagnostic certainty, but this will result in a decrease in sensitivity [20]. Notably, in datasets with class imbalance, the ROC-AUC may yield overly optimistic results. This is due to its insensitivity to the majority of negative samples. In contrast, the precision-recall (PR) curve and its corresponding area under the curve (PR-AUC) can more accurately evaluate the model’s ability to identify minority classes. This is because the PR curve focuses on the trade-off between precision and recall (i.e., sensitivity) and is more sensitive to false positives. This is particularly applicable in scenarios where positive cases are rare, such as the screening for early DKD [21]. Therefore, when reporting the performance of DKD diagnostic models, it is recommended to present both ROC-AUC and PR-AUC simultaneously, and in combination with the aforementioned clinical objectives, report specific metrics such as sensitivity, specificity, and precision at key decision thresholds, so as to comprehensively evaluate their clinical applicability. In prognostic prediction models, discrimination refers to the ability to differentiate between patients who will and will not experience a specific future outcome, quantified using the concordance index (C-index). The values for both metrics range from 0.5 (indicating random chance) to 1.0 (representing perfect discrimination) [22, 23].

Regarding calibration, for diagnostic models, it represents the agreement between predicted probabilities and actual diagnostic outcomes at a specific assessment time point. This is most intuitively assessed using a calibration plot, where predicted probabilities are plotted against the observed event frequencies. Perfect calibration is indicated by points aligning with the 45-degree line. This visual assessment is often supplemented by statistical tests such as the Hosmer-Lemeshow goodness-of-fit test, where a non-significant P-value (p > 0.05) suggests no major miscalibration. For prognostic models, calibration reflects the consistency between predicted probabilities and observed frequencies of the target outcome over a defined period and is best evaluated using a time-dependent calibration plot at clinically relevant time horizons (e.g., 1, 3 or 5 years) [24]. When two models exhibit similar discrimination, their calibration may differ significantly. Persistent overestimation or underestimation of risk by a model can severely impact clinical decision-making; thus, both discrimination and calibration must be reported together [25]. Clinical utility can be quantified using decision curve analysis (DCA), which evaluates the net benefit of a model across various threshold probabilities, balancing benefits against harms. By comparing the model with other risk stratification methods, DCA comprehensively assesses its value and generalizability in real-world applications [26]. The evaluation of these three aspects is essential for determining whether adding new predictors (such as novel kidney injury biomarkers to a model) can effectively enhance its predictive performance.

When C-statistics—including ROC and C-index—fail to capture clinically meaningful yet modest improvements in model performance, the net reclassification improvement (NRI) and integrated discrimination improvement (IDI) may be introduced as supplementary metrics to evaluate the value of newly added predictors [27]. Furthermore, to mitigate overfitting and avoid overoptimistic performance estimates derived from internal validation, external validation of diagnostic models across diverse suspected populations and prognostic models across varied patient cohorts (with temporal, geographic, and ethnic diversity) is necessary to ensure reliable application in clinical practice [28].

Modeling algorithms involved in the model

For the development of most current prediction models, ML is the core tool. ML is a key subfield of AI, which is not constrained by hard-coded rules but instead enables computers to analyze large amounts of data, automatically learn patterns from it, and use these patterns to make accurate predictions for new or unknown situations. ML can be categorized by learning paradigms into supervised learning, unsupervised learning, semi-supervised learning, and reinforcement learning. The processes of all these paradigms rely on various algorithms, and algorithms are a set of rules and statistical techniques capable of learning meaningful patterns from data [29]. As shown in Fig. 2 and Table 2, this paper lists some commonly used algorithms in the medical field. From linear regression to deep learning (DL) algorithms, these algorithms significantly differ in their ability to handle complex data, modeling performance, and model interpretability and explainability. Existing evidence indicates that ML shows great potential in DKD research [41]. It is not limited to cross-sectional diagnosis or short-term outcome prediction using methods like logistic regression (LR) [7]，but can also extend to long-term dynamic risk prediction through techniques such as survival analysis [42], which is of great significance for the clinical diagnosis and treatment of DKD.

Fig. 2.

Diagram of ML algorithms. A. Linear regression; B. Logistic regression; C. K-nearest neighbors; D. K-means clustering; E. Support vector machine; F. Naive bayes; G. Principal component analysis; H. Extreme gradient boosting; I. Multilayer perceptron; J. Decision tree; K. Convolutional neural network; L. Random forest

Table 2.

ML algorithm summary

Algorithm name	Learning paradigm	Category	Core idea	Advantages	Disadvantages
Linear regression [30]	Supervised learning	Regression	Find a straight line that minimizes the sum of squared errors	Simple, strong interpretability, computationally efficient	Poor at fitting nonlinear relationships, sensitive to outliers
Logistic regression [31]	Supervised learning	Classification	Uses a Sigmoid function to map output to a probability	Output has probabilistic meaning, model is interpretable	Can only handle linear decision boundaries
Decision tree [32]	Supervised learning	Classification/Regression	Recursively splits the data using a series of questions	Intuitive, easy to interpret, no need for feature scaling	Prone to overfitting, sensitive to small data changes
Random forest [33]	Supervised learning	Ensemble/Classification/Regression	Builds multiple trees and combines results	Powerful performance, strong resistance to overfitting	Computationally expensive, less intuitive
Gradient boosting machines (XGBoost, LightGBM) [34]	Supervised learning	Ensemble/Classification/Regression	Trains trees sequentially to correct errors	High accuracy, good flexibility	Training is slow, prone to overfitting
Support vector machine [35]	Supervised learning	Classification/Regression	Finds a hyperplane that maximizes the margin	Effective in high dimensions, handles nonlinearity	Slow on large datasets, sensitive to parameters
Naive bayes [36]	Supervised learning	Classification	Based on Bayes’ theorem with feature independence assumption	Fast, good for small data and multi-class	Independence assumption rarely holds, hurting accuracy
K-nearest neighbors [37]	Supervised learning	Classification/Regression	Class is based on the majority vote of nearest neighbors	Simple, no training, insensitive to outliers	High prediction cost, sensitive to high-dimensional data
K-means clustering [38]	Unsupervised learning	Clustering	Partitions data into K clusters by minimizing squared distances	Simple, efficient, suitable for large datasets	Must specify K, sensitive to initialization and outliers
Principal component analysis [38]	Unsupervised learning	Dimensionality reduction	Projects data to lower dimensions retaining maximum variance	Reduces cost, eliminates feature correlation	Linear method, cannot capture complex nonlinearities
Multilayer perceptron [39]	Deep learning	Neural network	A fully connected feedforward network with hidden layers	Can approximate complex nonlinear functions	Many parameters, prone to overfitting
Convolutional neural network [40]	Deep learning	Neural network	Uses convolutional kernels to extract spatial features	Parameter sharing, translation invariant, ideal for images	Many hyperparameters, requires lots of data and compute

However, the selection of specific algorithms is not fixed, but rather requires careful consideration based on the specific context of the research. No single model type is universally optimal; model selection should be based on a detailed assessment of the particular research setting (Table 3). The dimensionality and complexity of the data are the primary determining factors. ML models excel at handling high-dimensional, non-linear data (such as genomics or radiomics), capturing complex interactions that traditional linear models cannot identify. However, for low-dimensional data, such as routine clinical variables, this complexity advantage may disappear or even lead to performance degradation due to overfitting. In such cases, simpler and more robust models like the CPH or LR models are often more suitable [45]. Sample size is equally crucial. The ability of ML models to learn complex patterns requires substantial data for reliable training and validation. In cohorts with limited sample sizes, traditional statistical models, with their solid theoretical foundations and fewer parameters, typically demonstrate better generalization capability [45]. Furthermore, the clinical demand for model interpretability profoundly influences the choice. Risk ratios or odds ratios directly provided by traditional models are indispensable for mechanistic exploration (e.g., validating the independent effect of a biomarker) and for building clinical trust. While many high-performance machine learning models, often regarded as “black boxes,” can offer insights through post hoc explanation tools, their internal logic is inherently less transparent than that of traditional models [47]. Ultimately, the trade-off in model selection also depends on the research objectives of the clinical question. If the primary goal is to achieve accurate risk stratification to identify high-risk patients and the sample size is sufficient, optimizing the discriminative performance of machine learning models may be prioritized. Conversely, if the research aims to understand disease pathogenesis or evaluate the independent effect of interventions, traditional models that provide clear and interpretable parameters should be chosen [48]. Therefore, future research reports should not focus solely on performance metrics. Instead, they should comprehensively evaluate models by integrating data characteristics, sample size, interpretability needs, and clinical objectives. This approach will promote the development of prediction tools that are both precise and practically feasible.

Table 3.

Comparative analysis of common model types for DKD prediction

Feature dimension	Logistic regression [31, 43]	Cox proportional hazards model [44]	Machine learning models [45, 46]
Core principle	Models the relationship between a linear combination of predictors and the probability of an event via a logistic function	Models the log hazard ratio as a linear combination of covariates, based on the proportional hazards assumption	Learns complex patterns (linear/non-linear) from data automatically through algorithms
Main advantages	• Highly interpretable results (odds ratios) • Computationally efficient, less prone to overfitting • Stable with small sample sizes	• Handles censored data perfectly, the gold standard for survival analysis • Interpretable results (hazard ratios) • Model robustness	• High discriminative performance: often achieves higher AUC/C-index when handling high-dimensional, non-linear, interactive data • Automated feature selection: can handle a large number of predictors and identify key features automatically • High flexibility: applicable to diverse data types (images, sequences, etc.)
Main limitations	• Assumes linearity, ineffective at capturing complex relationships • Sensitive to multicollinearity • Requires manual feature selection	• Similarly constrained by the proportional hazards and linearity assumptions • Difficult to handle very high-dimensional data directly	• “Black-box” nature: poor interpretability, reducing clinical trust • Overfitting risk: especially with small sample sizes or noisy data • Computational and tuning complexity: requires more computational resources and expertise
Application scenarios in DKD Prediction	• Cross-sectional diagnostic models (e.g., predicting current disease probability based on clinical indicators) • Short-term risk prediction	• Long-term prognostic prediction (e.g., predicting time to ESRD, death). • Studies requiring clear HR values for risk factors.	• Diagnostic/prognostic models integrating high-dimensional data like multi-omics or imaging • Large-scale risk stratification based on Electronic Health Records • When traditional model performance plateaus
Selection recommendations	Use as a baseline model, suitable for preliminary studies with few variables, clear relationships, and where interpretability is prioritized	The preferred choice for survival data analysis, especially when the research aim is to clarify the independent impact of factors on a time-to-event outcome	Use when data is complex, sample size is large, and prediction accuracy is the primary goal. Must be accompanied by rigorous validation (to prevent overfitting) and interpretability tools (e.g., SHAP)

Considerations in modeling

During the development of DKD models, several core issues may arise that significantly impact model interpretation, validation, and performance evaluation, necessitating careful consideration. Firstly, the lack of standardization in defining diagnostic and prognostic outcomes hinders cross-study comparisons of different models and their performance. Current guidelines or studies on DKD exhibit discrepancies in defining diagnostic thresholds. For example, the positive criteria for UACR are set as ≥ 30 mg/g [49] or ≥ 300 mg/g [50], leading to inconsistent diagnostic benchmarks. The variation in defining prognostic outcomes is even more pronounced, including composite endpoints such as a sustained decline in eGFR ≥30%, ESRD, or ESRD combined with renal disease-related death [51]. Such heterogeneity in definitions not only impedes direct comparison between models developed using different criteria, but also substantially hinders the benchmarking of model performance. Furthermore, the lack of standardized definitions limits the effective validation and generalizability of models across studies and diverse populations, creating practical challenges for clinicians in selecting appropriate prediction tools [52]. Therefore, there is a pressing need to foster international consensus on unified core outcome definitions and reporting standards. This will enhance model comparability and reproducibility, facilitate clinical translation, and ultimately lay a solid foundation for robust validation, reliable risk stratification, and broader clinical adoption.

Secondly, the complex pathological mechanisms and temporal heterogeneity of DKD substantially increase the difficulty of model development. DKD and cardiovascular disease (CVD) exhibit a bidirectional promoting relationship, and overlooking this synergistic effect may lead to underestimation of overall risks. Concurrently, the disease progression demonstrates distinct stages, each characterized by different dominant risk factors. Although some models have attempted to incorporate longitudinal data to capture dynamic disease evolution, their feasibility is constrained by high dependency on follow-up data and significant issues with data missingness. In contrast, single-time-point models prove inadequate for effectively identifying rapid progressors [53]. Future model development should focus on constructing dynamic prediction frameworks capable of integrating longitudinal data (e.g., employing time-dependent covariate models or recurrent neural networks to process sequential data) and, on this basis, incorporating mechanisms for periodic recalibration and competing risk analysis [54]. This approach will systematically address the challenges posed by the temporal evolution of risk profiles, enabling more precise capture and prediction of disease progression trajectories.

Third, research samples often exhibit significant bias, which undermines the generalizability of the models. This bias is primarily reflected in three levels of heterogeneity that have not been adequately modeled. Most existing models are developed based on type 2 diabetes (T2D) populations and fail to sufficiently incorporate diverse genetic backgrounds and environmental variations, resulting in limited external validity [55]. Type 1 diabetes (T1D) and T2D differ fundamentally in their pathogenesis, metabolic environment, and complication profiles. For instance, DKD progression in T1D patients may be more closely associated with long-term autoimmunity and glycemic fluctuations, whereas T2D patients are predominantly driven by metabolic syndrome [56]. Consequently, directly applying models developed from T2D populations to T1D may lead to distorted risk weighting or the omission of T1D-specific key predictors. Moreover, heterogeneity in disease stages is another critical issue. The dominant biological processes in early-stage and advanced-stage DKD differ significantly. Metabolic indicators relied upon by early-stage models may lose their predictive efficacy in advanced stages, while fibrosis markers incorporated into advanced-stage models are often unavailable in early stages. This “temporal heterogeneity” makes it difficult for static models based on single time points to accurately delineate individualized disease trajectories, particularly in reliably identifying “rapid progressors” [6]. In terms of population and clinical context heterogeneity, there is generally insufficient representation of elderly patients, individuals with comorbidities, and specific medication subgroups, leading to suboptimal predictive performance of the models in these populations [57]. Additionally, the pathological differences between type 1 and type 2 diabetic kidney disease have not been adequately addressed. Combined modeling often distorts risk weighting, while subtype-specific modeling faces challenges of insufficient sample sizes and compromised performance. Furthermore, differences in genetic background, social environment, and medical practices contribute to unstable model performance across different racial and geographical populations [58, 59]. For example, models developed based on single-ethnicity cohorts may exhibit significantly reduced predictive efficacy when their incorporated genetic risk scores or specific biomarkers are applied to another ethnic group [59]. These issues collectively constrain the accuracy and practical utility of the models. Consequently, while predictive models (including ML models) show promising performance in controlled development settings, their robustness across diverse populations and clinical environments remains a critical challenge that must be addressed through rigorous external validation and the development of more adaptable systems.

DKD models with classic risk factors for diagnosis or prognosis

Currently, various statistical models have been employed to predict CKD. Some of these models were specifically developed for diabetic populations, while others, such as the Kidney Failure Risk Equation (KFRE) [60] and the CKD Prognosis Consortium Equations [61], were derived from general populations. Research indicates that hyperglycemia can directly damage the kidneys through mechanisms like advanced glycation end products (AGEs) accumulation, oxidative stress, and glomerular hyperfiltration, while also exhibiting synergistic effects with risk factors such as hypertension and dyslipidemia to accelerate pathological progression. Due to the insidious onset and rapid progression of DKD, coupled with significant differences in the absolute risk distribution of CKD between diabetic and non-diabetic populations, models based on general populations may demonstrate suboptimal predictive accuracy when applied to diabetic patients [62].

It is important to note that the design and feature selection of DKD prediction models typically focus on its unique pathophysiological mechanisms (such as hyperglycemia-induced accumulation of advanced glycation end products, renal hemodynamic abnormalities, activation of inflammatory and fibrotic pathways, etc.) and clinical phenotypes (such as diabetes duration, HbA1c, UACR, and eGFR, etc.) [63]. Most of these models were developed and validated using diabetic patient cohorts, and their predictive efficacy relies on DKD-specific combinations of risk factors [14, 64]. Furthermore, because the core predictors of these models (such as HbA1c, UACR, eGFR, and retinal features, etc.) may have weaker pathological relevance in non-diabetic CKD (such as IgA nephropathy, hypertensive nephropathy, and lupus nephritis, etc.), models specifically developed for DKD may likewise be poorly generalizable to CKD caused by other etiologies [65]. Although some common clinical variables (such as eGFR, UACR, age, and blood pressure) hold predictive value across different CKD types, their weight, interactions, and synergistic patterns with etiology-specific factors may differ significantly [65]. Therefore, the effectiveness and applicability of DKD models in non-diabetic CKD populations must be rigorously evaluated through external validation studies involving cohorts with different CKD etiologies.

In this section, we will focus specifically on diagnostic and prognostic models developed for diabetic patients utilizing conventional risk factors (A detailed assessment of each prediction model study according to the TRIPOD standards is provided in ESM Table 1).

Conventional prediction models for diagnosing DKD onset

Numerous LR models have been employed for personalized assessment of CKD risk in diabetic patients. Since some of these models provide predictions for fixed time points (e.g., 1, 3, or 5 years in the future), they essentially represent a cross-sectional evaluation of the risk of developing the disease at a specific time point based on the current status; therefore, this paper also categorizes them as diagnostic models. Dunkler et al. [7] developed and validated a multinomial logistic regression model using data from the ONTARGET and ORIGIN clinical trials. This model incorporated only four baseline indicators—albuminuria, eGFR, sex, and age—to predict the risk of developing CKD within 5.5 years in individuals aged 55 and older, achieving an AUC of 0.69 with good calibration. For the 3-year prediction time point, Sun et al. [64] constructed a nomogram prediction model for individuals aged 18–75 with T2D based on age, UACR, eGFR, and neutrophil percentage, achieving an AUC of 0.864, indicating high predictive accuracy. Another Chinese retrospective study developed a tool capable of effectively identifying high-risk DKD populations among T2D patients complicated with diabetic retinopathy (DR), with an AUC of 0.739 [66]. Furthermore, research by Low et al. [14] demonstrated that a conventional clinical diagnostic model constructed using stepwise multivariate LR could achieve an AUC of 0.832.

Meta-analysis has confirmed that ML can also achieve superior predictive performance and generalizability in the diagnosis of DKD [67]. Researchers are progressively exploring the use of ML to analyze both structured and unstructured data to improve targeted screening. Allen et al. [68] developed an ML algorithm using large-scale Electronic Health Records (EHR), which achieved AUCs above 0.75 for predicting the progression stage of DKD within 5 years in newly diagnosed T2D patients, and even AUCs above 0.82 for predicting more severe endpoints. A study by Sarkhosh et al. [69] indicated that the optimal model for predicting the onset of DKD within 5 years was LR, with an AUC of 0.758. Furthermore, two retrospective analyses based on electronic medical records of Chinese T2D patients showed that the light gradient boosting machine (LightGBM) demonstrated the best performance in predicting the 3-year risk of onset, with both AUC and accuracy above 0.80 [70, 71]. Lee et al. [72] constructed seven ML models using multicenter clinical data from T2D patients, with external validation revealing that the extreme gradient boosting (XGB) model, with an AUC of 0.811, provided the most robust prediction for DKD. Meanwhile, a clinical study by Cho et al. [73] on T2D patients confirmed that the support vector machine (SVM) could robustly identify predictors of DKD, achieving an AUC of 0.969. The variations in performance among these models may be attributable to heterogeneity in the size and ethnic diversity of the study populations.

Currently, research on diagnostic models for T1D remains relatively limited. Among the reported large-scale clinical studies, the XGB algorithm has consistently demonstrated superior predictive performance, with externally validated AUC values of 0.718 [74] and 0.760 [75], respectively. These findings hold promise for facilitating early and precise identification of DKD patients.

Conventional prediction models for forecasting ESRD in DKD

To further enable dynamic prediction of ESRD in diabetic patients, the academic community has developed various prognostic models [15, 51, 76]. These models aim to forecast the disease trajectory of DKD and assess its progression towards ESRD. It should be noted that research on CKD prognostic models specifically for T1D patients remains relatively scarce, with existing models primarily focused on T2D populations. Goubar et al. [77] applied the KFRE to predict ESRD in a T2D population, achieving C-statistics above 0.80 over both 2-year and 5-year follow-up periods. Elley et al. [42] established a CPH model based on a large New Zealand cohort of T2D patients, using fatal or non-fatal ESRD as the endpoint. The model demonstrated C-statistics of 0.89–0.92 in both training and validation cohorts over a 5-year follow-up. Similarly, Jardine et al. [78] developed a CPH model using data from T2D patients in the ADVANCE trial, achieving C-statistics of 0.847 for ESRD and 0.647 for new-onset proteinuria, significantly outperforming traditional models based solely on eGFR and/or ACR. Studies by Eric et al. [79] and Dong et al. [80], each involving over 140,000 Chinese T2D patients with average follow-up periods of 5 and 10 years respectively, both developed ESRD prediction models showing excellent discrimination (C-statistic above 0.85), which will facilitate risk stratification in primary diabetes care. Beyond traditional CPH models, ML has also been utilized for ESRD prognosis prediction. Zou et al. [81] followed T2D patients with biopsy-confirmed diagnosis for at least one year, identifying random forest (RF) as the optimal model for ESRD screening, with a C-statistic of 0.90 and an accuracy of 82.65%. A team from Peking University further integrated CPH with ML methods (XGB and survival SVM) to develop the PKU-CKD model. Validation in a T2D population showed comparable predictive accuracy between the CPH model and XGB [82] (C-index ≈0.83).

Prediction models for CVD and mortality risk in DKD

Given the significantly increased risks of CVD and mortality faced by DKD patients during disease progression [83], several studies have consequently developed specialized prediction models for this population. In the realm of cardiovascular risk prediction, Ren et al. [16] utilized a feedforward neural network to construct a DeepSurv deep learning model for predicting the risk of composite cardiovascular events in DKD patients. This model achieved a C-index of 0.767, outperforming traditional CPH and random survival forest (RSF) models. Another team, Januzzi et al., developed a risk tool for DKD patients capable of comprehensively assessing both kidney outcomes and major adverse cardiovascular events, achieving a C-index of 0.80, which has been validated using clinical trial data from canagliflozin studies [84]. Regarding mortality risk, Jin et al. [17] built a prognostic model using data from DKD patients in the MIMIC database, with all-cause mortality as the endpoint, achieving a C-index of 0.795. Within the NHANES study, a model based on CPH also demonstrated excellent predictive capability for assessing long-term all-cause mortality, with C-indices of 0.773, 0.788, and 0.817 for 3-year, 5-year, and 10-year predictions, respectively [85]. Another study focusing on a T2D population with new-onset DKD developed a cardiovascular mortality risk score model with a C-index of 0.747, whose predictive efficacy was significantly superior to the traditional UKPDS risk engine [86]. Furthermore, a study by Galsgaard et al. [87] on patients with T1D and kidney disease found that a prognostic model for all-cause mortality developed using data from this specific population achieved a maximum C-index of 0.868.

Prediction models for treatment response in DKD

Prediction of treatment response is a crucial subfield of prognostic modeling. It aims to assess improvements in key clinical outcomes—such as reduced albuminuria or slowed eGFR decline—following specific treatments like renin-angiotensin system (RAS) inhibitors or sodium-glucose cotransporter 2 (SGLT2) inhibitors. These models provide direct support for individualized treatment decisions [88]. Although research on DKD treatment-response prediction remains limited, its importance is growing. For example, Siriyotha et al. [89] used logistic regression on a large real-world cohort to develop a model predicting the absolute reduction in CKD risk with SGLT2 inhibitor therapy in T2D. The model estimated a median absolute risk reduction of −4.49% per patient and identified high-benefit subgroups, such as older males with hypertension. Internal validation showed a discrimination of 0.668 and good calibration (p = 0.059). In another study, Jaimes et al. [90] analyzed urinary peptide profiles from diabetic patients on RAS inhibitors and built the DKDp189 model using a support vector machine based on 189 peptides. This model significantly predicted non-response or DKD progression in two independent validation cohorts, with discrimination values of 0.60–0.63. Thus, treatment-response prediction models are valuable tools for DKD management. They help identify patients most likely to benefit from specific therapies and provide a quantitative basis for prioritizing treatment in clinical practice.

Based on the aforementioned research, it can be observed that existing models show a high degree of overlap in the selection of potential predictors, as illustrated in Fig. 3 and ESM Table 2. Most models incorporate basic demographic information, lifestyle indicators, anthropometric measurements, and laboratory parameters. While the widespread inclusion of these variables reflects their recognized importance in prediction, feature redundancy may also limit further improvement in the models’ discriminative ability. The primary reasons are: (1) Information overlap and diminishing marginal returns: Adding new clinical indicators that are highly correlated with existing variables contributes little to the overall information gain of the model, making it difficult to effectively enhance metrics such as AUC or C-index; (2) Multicollinearity: In traditional models, strongly correlated predictors can weaken the stability of coefficient estimates, impairing the model’s generalization ability; (3) Overfitting risk: In ML, redundant features increase model complexity, making it more prone to capturing noise rather than genuine disease signals, thereby reducing external validation performance [31, 91].

Fig. 3.

Current status of research on routine diagnosis and prognostic models for DKD. Predictors are summarized into categories. The numbers in the color scheme indicate the number of individual predictors included in the respective predictor category. For full table see ESM Table 2. a includes age, sex, race, marital status. b includes smoking status, alcohol consumption status, physical activity, diet. c includes systolic blood pressure, diastolic blood pressure, BMI, weight, waist-to-hip ratio, waist circumference. d includes total cholesterol, high-density lipoprotein cholesterol, low-density lipoprotein cholesterol, triglycerides, lipid-lowering medication. e includes glycated hemoglobin, fasting blood glucose, 2-hour postprandial blood glucose, serum insulin, diabetes duration, family history of diabetes, oral hypoglycemic agents, insulin use. f includes coronary heart disease, hypertension, congestive heart failure, myocardial infarction, cerebrovascular disease, antihypertensive medication. g includes renal insufficiency, kidney disease, eGFR, 24-hour urinary protein, uric acid, serum creatinine, UACR, blood urea nitrogen, cystatin C, urinalysis parameters (e.g., pH, clarity, color). h includes alanine aminotransferase, aspartate aminotransferase, gamma-glutamyl transferase, alkaline phosphatase, total protein, albumin, total bilirubin, direct bilirubin. i includes white blood cell count, hemoglobin, platelet count, absolute neutrophil count or percentage, absolute lymphocyte count or percentage, absolute monocyte count or percentage, red blood cell count. j includes complications (retinopathy, peripheral neuropathy, peripheral vascular disease), comorbidities (tuberculosis, pneumonia, sepsis), thyroid function (thyroid-stimulating hormone, free triiodothyronine, free thyroxine), inflammatory markers (C-reactive protein, procalcitonin, erythrocyte sedimentation rate), vital signs (pulse, respiratory rate, body temperature), serum electrolytes (potassium, sodium, chloride), coagulation parameters (activated partial thromboplastin time, prothrombin time, fibrinogen), retinol-binding protein, composite parameters (triglyceride-glucose index, atherogenic index, neutrophil-to-lymphocyte ratio). Abbreviation: LR, logistic regression; LSTM, long short-term memory; XGB, extreme gradient boosting; RF, random forest; SVM, support vector machine; DT, decision tree; KNN, k-nearest neighbors; ANN, artificial neural network; NB, naive bayes; LASSO, least absolute shrinkage and selection operator; CPH, cox proportional hazards; RRT, renal replacement therapy; KRT, kidney replacement therapy; CHD, coronary heart disease; PAD, peripheral arterial disease

In this context, exploring more novel predictors, such as omics, imaging, fundus photography, and digital pathology indicators, may help break through the current performance bottleneck of models and achieve more accurate risk stratification. These emerging biomarkers capture unique and complementary pathophysiological aspects of DKD that are not reflected by routine clinical indicators, providing incremental predictive value. Emerging biomarkers can capture earlier pathophysiological events. For example, metabolomics can detect metabolic pathway dysregulation before a decline in eGFR, and retinal microvascular parameters can serve as early synchronous markers of systemic microvascular disease [92, 93]. Simultaneously, they can provide mechanism-specific insights. For instance, specific proteomic or transcriptomic signatures can differentiate between inflammation-dominant or fibrosis-dominant disease stages, thereby enabling mechanism-oriented prognostic assessment [94]. Furthermore, these markers can directly quantify structural damage. AI-based radiomics or digital pathology can precisely quantify the extent of renal fibrosis or retinal lesions, with their predictive value surpassing that of functional surrogate indicators such as proteinuria [95]. Most importantly, by integrating multi-modal data such as genomic risk, clinical phenotypes, and imaging features, a more comprehensive individual risk profile can be constructed, overcoming the limitations of single data dimensions posed by the high heterogeneity of DKD [96].

Landscape of DKD prediction models based on multimodal data integration

As mentioned earlier, traditional models for predicting the onset, progression, and mortality of DKD typically incorporate only a limited set of clinical features and biomarkers that are readily available in routine clinical practice. However, single-modal data may be insufficient to fully characterize complex biological processes or disease states. Presently, more researchers are turning their attention to unconventional predictors such as multi-omics data, medical imaging, fundus images, and pathological images. These multidimensional indicators can not only capture disease-related signals from molecular, structural, and microscopic perspectives, compensating for the limitations of traditional variables in terms of informational breadth and depth, but some can also provide “incremental predictive information beyond established risk factors,” thereby offering new directions for developing novel models with superior performance [97]. At present, such research is primarily focused on the diagnostic aspect of DKD.

Molecular profiling

The search for biomarkers at the molecular level represents a critical strategy for significantly improving the prediction of DKD. Current molecular-level prediction models are primarily constructed using four categories of biological molecules: genes, RNA, proteins, and metabolites, with data largely derived from in-depth mining and validation through various omics technologies (Table 4). ML is widely employed to screen predictors and build models aimed at early identification, progression prediction, mechanism exploration, and therapeutic target discovery for DKD [41]. Research at the genetic level is the most extensive. Given that most gene sets originate from the Gene Ontology (GO) database and are closely linked to DKD pathophysiological mechanisms such as immunity [101], pyroptosis [102], oxidative stress [105], mitochondrial dysfunction [108], glycolysis [104], ferroptosis [99], renal fibrosis [103], and abnormal amino acid metabolism [98], this may significantly contribute to the exceptional predictive performance of these models (AUC = 0.8–1.0). Although studies on RNA are relatively limited, existing reports indicate that models using renal-protective miRNAs from urinary exosomes achieve nearly 90% accuracy [110], and mRNA models can reach an AUC of 0.950 [112], demonstrating considerable predictive potential. Additionally, proteins used for constructing DKD diagnostic or prognostic models are often associated with complement activation, inflammation, and cellular injury, sourced from urinary proteomics or prior validation studies. These models show strong predictive performance, with AUC or C-index values ranging from 0.740 to 0.880 [120–124]. Finally, metabolites—serving as end-effect molecules of gene function and direct phenotypic carriers—derived from metabolomics and closely related to energy metabolism regulation are also vital for building DKD prediction models, achieving AUC or C-index values between 0.8 and 1.0 [113–118]. In summary, biomarkers across different molecular levels hold substantial value in DKD prediction, and integrating multi-omics data will be essential for developing high-accuracy predictive models in the future.

Table 4.

Molecular profiles used for building prediction models

Specific marker(s)	Discovery source	Research level	Marker function	Algorithm(s)	Performance	Objective	Ref.
AOC1, HAAO, STAT1, OGDHL, TDO2	GEO and MSigDB databases	Gene	Tryptophan metabolism-related	LR	AUC = 0.996	Diagnosing DKD tubular injury	[98]
SKIL, RASA1, YTHDC2, SON, MRPL11, HSD17B14, DUSP1, FOS	GEO database	Gene	Ferroptosis-related	LASSO	AUC = 0.818	Diagnosing DKD	[99]
FOXD1, LOX, GJA1, BTG3	GEO and CellAge databases	Gene	Cellular senescence	LASSO,RF	AUC = 0.996–1.000	Diagnosing glomerulus-associated DKD	[100]
FSTL1, CX3CR1, AGR2	GEO database	Gene	Immunity and cuproptosis	LASSO,SVM-RFE	AUC = 0.946	Predicting DKD progression	[101]
CASP1, CITED2, HTRA1, PTGS2, S100A12	GEO database	Gene	Regulating pyroptosis	RF,LASSO,LR	Good diagnostic capability	Early diagnosis of DKD	[102]
FibrosisScore model	GEO database	Gene	Renal fibrosis-related	RF,LASSO	AUC = 0.803–0.992	Identifying DKD patients at higher fibrosis risk	[103]
GScore model	GEO database	Gene	Glycolysis-related	108 combinations of 10 ML algorithms	AUC = 0.743–0.975	Early diagnosis and treatment of DKD	[104]
DUSP1, PRDX6, S100A8	GEO database	Gene	Oxidative stress	LR,SVM-RFE,LASSO	AUC = 0.946	Developing a diagnostic model for DKD	[105]
EFHD1, CASP3, AASS, MPC1, NT5DC2, BCL2A1	GEO database	Gene	Mitochondria-related genes	LR	AUC = 0.948–0.983	Exploring DKD pathogenesis	[106]
COL4A1, COL4A2, COL6A2, COL6A3, FN1, ITGA4, LAMB1	GEO database	Gene	Basement membrane-related	LR	AUC = 0.876	Identifying therapeutic targets for DKD	[107]
CLDN1, TYROBP, HDAC9, CASP3, RCN1	GEO database	Gene	Mitochondrial dysfunction and age-related inflammation	LASSO,SVM,RF,XGB,GLM	SVM performed best (AUC > 0.8)	Identifying potential therapeutic targets for DKD	[108]
MRS+Clinical variables	Epigenetics	Gene	Associated with CKD or its risk factors	LR	C-index = 0.870	Predicting CKD risk in T2D patients	[109]
miR-126,-146,-155	Urinary exosomes	miRNA	Maintaining homeostasis and regulating disease	SIMPER, PCO, CAP	Classification ACC = 89.01%	Building a DKD prediction model combining exosomal miRs and clinical variables	[110]
miR-133b,miR-342,miR-30a	Urinary exosomes	miRNA	Associated with diabetes or kidney disease	LR	ACC = 0.876	Identifying a miRNA combination for recognizing DKD	[111]
UMOD mRNA	Urinary exosomes	mRNA	Promoting kidney protection	LR	AUC = 0.950	Predicting early-stage DKD	[112]
Serum C0, C10:2, C8:1 and urinary C12:1+Clinical variables	Targeted metabolomics	Metabolite	Regulating energy metabolism	Linear regression and LR	AUC = 0.913	Improving early-stage DKD identification	[113]
g_Prevotella,g_Faecalibacterium, g_Klebsiella,Imidazolepropionic acid,Adipoylcarnitine,1-Methylhistidine	Untargeted metabolomics and 16S rRNA gene sequencing	Metabolite and gut microbiota	Involved in metabolic disease regulation	LASSO,LR	AUC = 0.939 and AUPR = 0.940	Identifing biomarkers for predicting DKD	[114]
trigonelline, hippurate, phenylalanine, glycolate, dimethylamine, alanine, 2-hydroxybutyrate, lactate, and citrate	Urinary metabolomics	Metabolite	Associated with CKD progression	PCA,PLS-DA	AUC = 0.912	Predicting T2D-associated CKD risk	[115]
Alanine, choline, N-phenylacetylglycine, and trigonelline+Clinical variables	Urinary metabolomics	Metabolite	Regulating energy metabolism, gut microbiota, oxidative stress	PLS-DA	AUC = 1.000	Predicting DKD	[116]
Citrulline + 9 Acylcarnitines	Plasma metabolomics	Metabolite	Regulating energy metabolism	PLS-DA,LR,SVM,RF,XGB	AUC = 0.970	Identifying renal function decline in T2D	[117]
13 Metabolites+Clinical variables	Urinary metabolomics	Metabolite	Associated with energy and amino acid metabolism	LASSO,CPH	C-index = 0.846–0.854	Predicting DKD progression	[118]
TNFR1, TNFR2, KIM1 + Clinical variables	BioMe and PMBB	Protein	Systemic inflammation and kidney injury	RF	C-index = 0.770	Predicting DKD progression	[119]
pTNF-R1 + Clinical variables	-	Protein	Inducing inflammation and apoptosis	LR	AUC = 0.740	Predicting DKD progression in Asian patients with early-onset T2D	[120]
Plasma CXCL-16, Angiopoietin-2, TGF-β1 + Clinical variables	-	Protein	Indicating vascular inflammation and endothelial dysfunction	Mixed LR	AUC = 0.760	Improving predictive performance of kidney risk models	[121]
ApoA4, CD5L, IGFBP3 + Clinical variables	PromarkerD	Protein	Associated with metabolism, inflammation, and cell fate	LR	AUC = 0.880	Predicting renal function decline in T2D	[122]
Urinary complement proteins + Clinical variables	Urinary untargeted proteomics	Protein	Associated with complement and coagulation cascades	LR	AUC = 0.767	Predicting kidney outcomes in DKD	[123]
17 Urinary protein markers	Urinary proteomics	Protein	Associated with diabetes and kidney disease	SVM and Linear Combination Model	Sensitivity = 81%, Specificity = 91%	Differentiating DKD from other chronic kidney diseases	[124]

Abbreviation: LR, logistic regression; LASSO: least absolute shrinkage and selection operator; RF: random forest; SVM-RFE: support vector machine - recursive feature elimination; GLM: generalized linear model; CPH: cox proportional hazards; XGB: extreme gradient boosting; PCA: principal component analysis; PLS-DA: partial least squares discriminant analysis; ACC: accuracy

Imaging indicators

Imaging techniques based on ultrasound and magnetic resonance imaging (MRI) can transform macroscopic structural and functional changes in the kidneys into quantifiable high-order features. This non-invasively reflects the complex biological processes underlying DKD and enhances the biological interpretability and clinical utility of models [125]. Studies have shown that integrating radiomics features extracted from renal ultrasound images, such as renal volume index, shear wave elastography, and resistive index, with routine clinical parameters, and using ML algorithms like LR, K-nearest neighbors (KNN), and SVM to build DKD models yields excellent performance (AUC 0.850–0.993). This provides significant auxiliary value for identifying early-stage DKD [126–129]. Regarding MRI, this technology also demonstrates strong potential. Based on imaging sequences of different kidney structures and using ML algorithms, researchers can distinguish different progression stages of DKD (early [130, 131], progressive [132], and end-stage [133]), with the maximum discrimination reaching 0.980 (Table 5).

Table 5.

Other specific parameters for predictive model construction

Auxiliary examination	Specific markers (and/or clinical indicators)	Data source	Algorithm(s)	Performance	Objective	Ref.
Imaging parameters
Ultrasound	Clinical indicators + Ultrasonographic parameters	Renal volume data/images	LR	AUC = 0.993	Predict early DKD	[126]
	Clinical indicators + Radiomics	2D ultrasound images	LASSO,LR	AUC = 0.850	Assess glomerular status in DKD patients	[129]
	Clinical indicators + Radiomics	2D ultrasound images	KNN,SVM,LR	KNN performed best (AUC = 0.910)	Predict early DKD	[127]
	Clinical indicators + Rad-score	2D ultrasound images	SVM-RFE,LASSO,LR	AUC = 0.878	Identify DKD in T2D patients	[128]
MRI	Radiomics	ASL images of both kidneys	DT,GBM,LR,RF,KNN,SVM,NB	ASL: AUC = 0.865; ASL+biological indicators: AUC = 0.734	Predict early renal injury and progression in diabetic patients	[130]
	ΔADC	Renal imaging	LR	Discriminate DKD: AUC = 0.707; Stratify ESKD risk categories: AUC = 0.823	Identify T2D patients diagnosed with DKD and those at higher ESKD risk	[133]
	Radiomics based on DKI and DTI	DSI sequences of renal parenchyma	LASSO,LR	AUC = 0.918	Differentiate T2D and DKD	[131]
	MRI-based texture features	Renal cortex images	LASSO,LR	AUC = 0.980	Early detection of Stage III DKD	[132]
Fundus parameters	Clinical indicators + OCTA parameters	Fundus imaging	LR	AUC = 0.681	Diagnostic tool for DKD	[134]
	Fundus images + Clinical metadata	Fundus images	Visual geometry group 16,DNN	AUC = 0.722	Predict CKD in T2D patients	[135]
	AI deep learning system	Fundus images	LDA, SBGT, MARS, SVM, NB, RPCT, RF, RDA, DT, KNN, NN.	AUC = 0.791–0.826	Detect DKD and differentiate pure DKD	[136]
Renal pathology	Three pathological indicators	Total glomerular excision images	LR,GBDT,Adaboost,LightGBM,XGB	XGB performed best (Precision = 0.921,F1 = 0.917)	Quantitative analysis of pathological glomerular tissues and changes in DKD	[137]
	Clinical indicators + Pathological parameters	Renal pathology slides	LR,LASSO	AUC = 0.940	Predict tubulointerstitial lesions in DKD	[138]
	34 Baseline clinical-pathological parameters	Pathology slides of renal biopsy specimens	CPH	C-index = 0.790	Assess renal prognosis in diabetic patients	[139]
	Clinical indicators + D-score	Pathology slides of renal biopsy specimens	CPH	C-index = 0.932	Predict renal outcomes in DKD patients	[140]
	Clinical indicators + J-score	Pathological analysis of renal biopsy tissues	CPH	C-index = 0.685	Predict renal prognosis in DKD patients	[141]
	Clinical indicators + PTBMIL	Pathology slides of renal biopsy specimens	CPH	C-index = 0.780	Predict prognosis of DKD	[142]
	Analytical renal pathology system	Digital pathology slides	CNN	Glomerulosclerosis: F1 = 0.928; K-W lesions: F1 = 0.953	Assist identification and pathological classification of glomerular lesions in DKD patients	[143]
Multimodal data	Metabolites and peptides	Metabolomics + Peptidomics + Transcriptomics	SVM,DT,NB,LR,LDA,KNN,LASSO	Correctly classified 69.6% early DKD patients and 75.7% clinical DKD patients	Investigate DKD pathogenesis	[144]
	Clinical indicators, fMRI markers, serum and urine biomarkers	Clinical, imaging, serum and urine multi-omics data	LR	AUC = 0.923	Classification and prognosis in DKD	[96]
	339 multidimensional indicators	Clinical features, retinal imaging, genetics, and blood metabolites	LR,LASSO,EN,CART,RF,GBDT,XGB,SVM,NB	EN performed best (AUC = 0.851)	Improve DKD prediction	[145]
	AI-assisted model	Clinical information, untargeted metabolomics, targeted lipidomics, and genome-wide SNP datasets	RF,SVM,LASSO	ACC = 0.70, AUC = 0.76	Identify biomarker signatures for high DKD risk in diabetic patients	[11]

Abbreviation: ADC, apparent diffusion coefficient; ASL, arterial spin labeling; DKI, diffusion kurtosis imaging; DSI, diffusion spectrum imaging; DTI, diffusion tensor imaging; fMRI: functional magnetic resonance imaging; Adaboost, adaptive boosting; CART, classification and regression trees; CNN, convolutional neural network; CPH: cox proportional hazard; DT, decision tree; DNN, deep neural network; EN, elastic net; GBDT, gradient boosting decision tree; GBM, gradient boosting machine; KNN, k-nearest neighbors; LDA, linear discriminant analysis; LightGBM, light gradient boosting machine; LASSO, least absolute shrinkage and selection operator; LR, logistic regression; MARS, multivariate adaptive regression splines; NB, naive Bayes; NN, neural network; PTBMIL, pathological tissue basement membrane thickness-based multi-instance learning; RDA, regularized discriminant analysis; RF, random forest; RPCT, recursively partitioned classification tree; SBGT, stochastic gradient boosting tree; SVM, support vector machine; SVM-RFE, support vector machine-recursive feature elimination; XGB, extreme gradient boosting

Fundus parameters

Given the shared pathophysiology of microvascular complications between the retina and kidneys, the inclusion of fundus features can provide highly relevant biological information for DKD prediction models, potentially enhancing their risk stratification capabilities [146]. In a Chinese study, incorporating optical coherence tomography angiography (OCTA) parameters into conventional clinical variables improved the model’s AUC to 0.681, with deep capillary plexus (DCP) density and the FD-300 area identified as key predictors [134]. Concurrently, researchers in Korea developed a hybrid model integrating Visual Geometry Group 16 (VGG16) with a deep neural network using large-scale cohorts from Korea and the United Kingdom. This model achieved an AUC of 0.722 for predicting DKD [135]. Furthermore, a large-scale real-world data study utilizing over 730,000 retinal fundus images developed an AI deep learning system named DeepDKD. This model demonstrated consistent performance across multiple ethnic datasets, with AUC values ranging from 0.791 to 0.826 [136] (Table 5). Collectively, these studies confirm the feasibility and generalizability of fundus features for DKD prediction.

Renal pathology

Renal pathology slides are considered the “gold standard” for diagnosing DKD. The integration of digital pathology with AI technology enables high-throughput mining of microscopic features using pathology slide-based ML models, offering the potential to further enhance diagnostic accuracy and prognostic assessment for DKD [147]. Some studies have utilized renal pathology images to quantitatively analyze glomerular pathological changes or predict tubulointerstitial lesions, achieving diagnostic accuracy and AUC values exceeding 0.90 [137, 138]. Other research has focused on evaluating kidney prognosis risk in DKD patients. By employing pathological slides of renal biopsy tissues and utilizing CPH models, prognostic models based on pathological parameters have been developed, with C-index values ranging from 0.685 to 0.932 [139–142]. Recently, analytical renal pathology systems based on a series of convolutional neural networks (CNNs) have been developed. These systems demonstrate classification performance comparable to that of pathologists in predicting baseline and long-term renal function in DKD patients. They achieved F1-scores of 0.928 and 0.953 for identifying global glomerulosclerosis and Kimmelstiel-Wilson nodules, respectively [143] (Table 5). In summary, given that ML-based renal pathology analysis has already demonstrated exceptional performance, promoting its standardized clinical application has become a core objective in translational research.

Multidimensional data integration

The data sources in the aforementioned studies are relatively homogeneous, sometimes combined with clinical variables. Given that the onset and progression of DKD involve complex interactions across multiple dimensions such as metabolism, hemodynamics, and inflammation, the scientific community is now attempting to build more holistic disease prediction models by integrating data from a wider range of sources and types, including images, text, genetic information, and clinical data [148]. For instance, one study utilized data from multi-omics sources (genomics, metabolomics, transcriptomics, lipidomics) combined with ML to identify high-risk DKD patients, achieving an AUC of approximately 0.76 [11, 144]. Another study integrated 339 features encompassing clinical indicators, retinal imaging, genetic information, and blood metabolites, further enhancing the predictive performance for DKD using an elastic net (EN) model, with an AUC reaching 0.851 [145]. A recent study also combined clinical indicators, imaging, and serum/urine multi-omics data to establish a multimodal model. The diagnostic model achieved a discrimination of 0.923, while the prognostic model maintained high performance with discriminations of 0.975 and 0.932 at 2-year and 3-year follow-ups, respectively [96] (Table 5). It is important to note that multimodal data fusion can enhance the biological interpretability and mechanism discovery potential of models. However, significant attention must be paid to key challenges. These include data acquisition difficulty, model complexity, and the risk of overfitting.

The clinical application, practical challenges and care pathways of DKD prediction models

The core value of DKD prediction models ultimately lies in optimizing and improving clinical diagnosis and treatment practices. For precision medicine to exert tangible impact, it is crucial that clinicians can stratify patients into corresponding complication risk categories based on model outputs and accordingly refine treatment strategies to prolong complication-free survival. It must be emphasized that prediction models themselves are not the ultimate goal; their true value resides in positively influencing medical decision-making. Possessing a validated and accurate risk prediction tool serves as a critical foundation for achieving this objective. Nevertheless, numerous challenges remain in the journey from development to integration into clinical practice.

Application tools for translation into clinical practice

The predictive tools for DKD are undergoing a significant evolution, transitioning from traditional methods towards AI. The eGFR calculated by the CKD-EPI or MDRD equations and the UACR remain fundamental indicators in clinical practice for estimating the progression of renal function and require long-term dynamic monitoring [10]. Furthermore, multifactorial tools such as the UKPDS risk engine [76] (incorporating UACR, eGFR, HbA1c, and blood pressure), the RECODe model [51], and various diagnostic and prognostic risk scoring systems presented via nomograms [66] or risk calculators [69] have been extensively developed, validated, and refined by researchers due to their relative simplicity and strong interpretability. Currently, the research frontier is increasingly focusing on more complex ML prediction models, such as the KidneyIntelX™ risk score [119], the Klinrisk model [149], the kidney age index [150], and various online web-based risk tools [69, 151]. These tools are capable of uncovering complex relationships between EHRs, vast amounts of medical data, and DKD, thereby enabling higher-precision early risk stratification. In summary, these prediction tools, developed based on multi-dimensional data, facilitate convenient and rapid dynamic assessment and quantification of an individual’s absolute risk of DKD, providing a basis for formulating risk alert strategies. To more intuitively present the clinical advantages of DKD prediction models compared to conventional diagnostics, this review constructs three hypothetical case analyses. The cases detail patients’ examination parameters and analyze the limitations of conventional diagnostic workflows versus the assessment advantages of prediction models. It is hoped that these cases will provide more concrete references for the clinical translation of prediction models (for details, see ESM Cases).

Barriers to the application of prediction models

Despite some tools having been deployed in clinical practice, the widespread adoption of DKD prediction models still faces multiple challenges spanning data source limitations, methodological, practical, and ethical aspects.

At the data quality level, real-world data often face systemic deficiencies. First, data missingness is a common problem. Due to the non-research nature of clinical practice, key variables such as dynamic renal function indicators, lifestyle information, or biomarkers often exhibit varying degrees of missingness, which may lead to representational bias in model training samples and instability in predictive performance [152]. Second, coding variability severely affects data standardization. Differences among medical institutions in disease coding (e.g., ICD systems), definitions and units of laboratory indicators, and even the recording of names for the same indicators exist. This requires complex data processing when deploying models across institutions, increasing implementation difficulty [153]. Third, follow-up inconsistency is a major constraint for prognostic models. In real-world settings, patient follow-up often involves irregular time intervals, interruptions, or loss to follow-up. This makes it difficult for dynamic risk prediction models based on longitudinal data to obtain complete time-series information, thereby affecting their ability to accurately capture disease progression trajectories [152]. Finally, heterogeneity in treatment interventions poses a significant challenge to the models. Treatment regimens for DKD patients are diverse and evolve over time. The widespread use of novel glucose-lowering drugs (e.g., SGLT2 inhibitors, GLP-1 receptor agonists) and RAS inhibitors has significantly altered the natural course of DKD. This treatment confounding effect means that models trained on historical data may fail to accurately reflect disease risks within the current treatment context, leading to predictions that deviate from actual clinical trajectories [154]. Additionally, efficacy-based dynamic treatment adjustments introduce time-dependent confounding effects that further complicate modeling [155]. These data limitations collectively lead to a performance-implementation gap between model development environments and real-world applications.

At the methodological level, the “real-world reliability” of DKD prediction models remains a significant challenge. On one hand, the issue of insufficient external validation is particularly pronounced. Most prediction models are confined to internal validation within their development cohorts. The lack of external validation across multiple centers or diverse ethnicities often leads to a marked decline in their performance in real-world clinical settings [28]. Furthermore, the unclear extent of clinical benefit impedes their clinical application. A key aspect of this uncertainty is the current lack of robust, real-world interventional studies demonstrating that using these models, even those with favorable DCA profiles, actually improves clinician decision-making or patient outcomes. For models to be genuinely useful in clinical practice, the introduction of clinical decision analysis metrics is necessary. For instance, DCA determines the optimal probability threshold for risk stratification. This is achieved through calculating the net benefit across different threshold probabilities, thereby comprehensively weighing the clinical consequences of false positives and false negatives. [26]. However, DCA provides a theoretical estimate of clinical utility; translating this into tangible clinical impact requires focusing on integrating DCA-validated models into electronic health record systems and conducting multicenter, prospective clinical trials to systematically evaluate their practical effectiveness in guiding personalized treatment, optimizing resource allocation, and improving patient outcomes [26].

In terms of practical challenges, acquiring additional data beyond routinely available information (such as conducting omics testing and screening for novel unconventional biomarkers) requires financial resources and institutional support, which may exacerbate disparities in healthcare resource distribution, reduce service accessibility, and even widen health inequalities. Therefore, cost-effectiveness analysis is particularly important, as it can assist policymakers in improving prediction accuracy while avoiding further deprivation of medical resources for vulnerable populations [156]. Moreover, even if prediction accuracy improves, the absence of matching intervention options or misalignment between interventions and patient needs and preferences may impose additional psychological burdens on patients due to more precise decision-making. Furthermore, issues such as the technical limitations of prediction models, particularly the “black-box” nature of many high-performance machine learning models in contrast to the interpretability of traditional statistical models like CPH, the complexity of clinical decision-making, and uncertainties in medical liability attribution contribute to a general lack of trust among clinicians in AI models, which also restricts their practical application [157]. To overcome the challenges arising from insufficient model trustworthiness and promote clinical adoption, the introduction of interpretability methods becomes particularly important. Techniques such as SHapley Additive exPlanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME) can interpret model predictions by quantifying the contribution of individual features to the model output (globally or for specific cases) [158]. Such explanations not only enhance model transparency but also align with the expectations of regulatory agencies (e.g., the FDA’s requirement for traceability of predictions) [158]. Future DKD prediction models should prioritize the integration of interpretability frameworks to generate clinically understandable reports, thereby supporting informed decision-making and enhancing the clarity of accountability attribution. It should be noted that, from the perspective of the practical feasibility of clinical translation of models, although novel data sources such as multi-omics, imaging, and fundus parameters demonstrate considerable potential, their application in routine clinical practice still faces significant obstacles. First, cost and scalability are major limitations. Multi-omics assays, high-resolution imaging examinations (e.g., MRI, OCTA), and digital pathology analysis are typically expensive and require specialized equipment and trained personnel. This makes them difficult to implement in resource-limited primary care settings or in low- and middle-income regions, potentially exacerbating disparities in healthcare resource distribution [159]. Second, data standardization and integration pose a substantial challenge. Data generated from different research centers, equipment brands, and imaging protocols exhibit high heterogeneity, and the fusion of multimodal data lacks unified technical standards and formats. This severely hinders the external validation and large-scale deployment of models [160]. Finally, integration into clinical workflows is equally complex. Seamlessly embedding prediction models into existing electronic health record systems and clinical decision-making pathways requires addressing a series of issues, including technical interfaces, user-friendliness, clinician acceptance, and the alignment of model outputs with established clinical guidelines [148]. Therefore, when assessing a model’s “clinical readiness,” it is essential to look beyond its discriminative performance and comprehensively consider its economic cost, technical accessibility, data generalizability, and workflow compatibility.

From a regulatory and ethical perspective, issues such as model transparency, data privacy, and health equity also present significant challenges. From a regulatory perspective, the U.S. FDA explicitly requires that medical AI models provide traceable justifications for their predictions. However, most current DKD prediction models are unable to clearly explain the rationale behind risk assessments. For instance, they cannot delineate the specific contributions of different features (such as HbA₁c or urinary protein) to risk judgments, thereby complicating regulatory approval. Ethically, there is a lack of unified standards for data privacy protection. The development of these models relies on extensive multidimensional data, which carries the risk of information leakage during collection, storage, and usage [161]. Moreover, structural social inequalities (e.g., in healthcare and education) exacerbate disparities in kidney health outcomes. This may lead to models trained on data from high-income regions failing to accurately capture risk characteristics of low-income populations [8].

Application scenarios and impact of prediction models in the clinical pathway of DKD patients

The key to the successful translation and application of prediction models lies in integrating them with established clinical care scenarios. In the future, the concepts and strategies of precision management can be implemented throughout all stages of DKD prevention, diagnosis, treatment, and management. In primary care settings, using simple diagnostic models based on routine indicators can facilitate efficient initial screening of diabetic patients, enabling early identification of high-risk individuals for DKD and optimizing referral efficiency [64]. In specialized diagnostic and treatment settings, models incorporating multi-omics or imaging features can provide more refined risk stratification for diagnosed DKD patients (e.g., distinguishing between rapid and slow progressors) and reveal potentially dominant pathological mechanisms, thereby assisting clinicians in developing individualized initial treatment plans [126]. In long-term follow-up management, dynamic prognostic models can continuously update risk predictions based on the patient’s latest clinical data, providing real-time evidence for adjusting treatment plans, and enabling dynamic alerts (such as the occurrence of ESRD and all-cause mortality) through remote management platforms [162]. In multidisciplinary collaboration settings, models capable of simultaneously predicting composite renal and cardiovascular endpoints help assess the patient’s overall risk and optimize the prioritization of treatment decisions [84]. Overall, the value of prediction models is not to replace clinical decision-making but to provide quantitative risk assessment in the aforementioned various clinical scenarios, thereby supporting medical decision-making, ultimately reducing healthcare resource consumption and improving patient outcomes (ESM Table 3).

Future directions: building a precise and accessible DKD prediction system

Based on the aforementioned research progress and existing limitations, this article proposes the following prioritized improvement directions and specific implementation pathways to promote the deepening of DKD research and its clinical translation.

Directions for action

To systematically advance the clinical translation of DKD prediction models, future efforts should adhere to the following prioritized sequence: First, promote the international standardization of DKD definitions and predictor variables by unifying diagnostic thresholds (e.g., UACR) and prognostic composite endpoints, and establish a consensus-based core variable set to lay the foundation for model comparability, reproducibility, and clinical implementation [52]. Second, there is an urgent need to establish large-scale, multi-ethnic, prospective cohorts that encompass multiple centers, diverse geographic regions, and underrepresented populations, providing long-term follow-up data to support the robust development, external validation, and generalizability of models [8]. It is important to note during this process that a prospective cohort design should be prioritized, and diverse samples should be proactively included to reduce selection bias at the source. For potential confounding bias, its sources must be identified and corrected using methods such as resampling, weighting, and propensity score matching. For missing values, direct deletion should be avoided, and model-based methods like multiple imputation should be prioritized. Additionally, outliers need to be detected, evaluated, and handled with caution. The entire data processing workflow should maintain methodological transparency and reproducibility, along with ongoing data quality monitoring [152]. Furthermore, develop interpretable multimodal integration frameworks, prioritizing modeling approaches that balance high performance with interpretability (e.g., explainable AI (XAI)) and establishing standardized data integration processes when integrating multi-source data such as genomics, imaging, and pathology. Concurrently, focus on embedding validated models into electronic health record systems and conducting prospective clinical trials to empirically evaluate their practical utility in improving clinical decision-making and patient outcomes [163]. Fourth, establish a tiered evaluation framework for model clinical readiness, clearly distinguishing between the “research potential” and “clinical usability” of models, and formulate corresponding translational pathways accordingly [160]. Finally, before promoting tools that integrate novel biomarkers or complex AI models, priority must be given to conducting health economic and health equity assessments to ensure technological accessibility and avoid exacerbating healthcare resource disparities [164].

Specific implementation pathways

Based on the aforementioned priorities, future research can advance through the following specific pathways. First, it is essential to promote the establishment of internationally unified standards for DKD definitions and predictor variables. To translate this goal into practice and reconcile existing discrepancies (e.g., in UACR thresholds or composite endpoints), a coordinated strategy is needed. This includes: (i) validating models across common alternative definitions during development; (ii) designing clinical tools with adjustable thresholds to accommodate different guidelines; (iii) employing meta-analysis and diverse external validation to calibrate performance under varied standards; (iv) mandating transparent reporting of all definitions in publications and shared data; and (v) ultimately establishing a consensus-based core variable set. Authoritative international organizations could lead efforts to define thresholds (e.g., for eGFR decline) and develop such core sets [52]. These should then undergo large-scale external validation across globally diverse regions and ethnic populations, aiming to identify robust models suitable for different healthcare settings and avoid a “one-size-fits-all” application [8]. Secondly, the development of predictive models must balance predictive accuracy with interpretability, while systematically integrating treatment variables and their temporal evolution. Given that the widespread adoption of novel drugs has significantly altered the progression trajectory of DKD, models should actively incorporate key information such as treatment type, initiation time, duration, and intensity, and employ time-dependent covariate models or dynamic updating mechanisms to more accurately reflect the dynamic impact of treatment strategies on disease risk. It is recommended to include treatment-stratified analysis during the model validation phase to assess model performance across different therapeutic subgroups, ensuring its generalizability in the new treatment era [162, 165]. Meanwhile, the active integration of XAI techniques is essential to clarify prediction rationales and mitigate the “black box” effect [163]. Additionally, an implementation learning mechanism should be established within EHRs systems, enabling models to dynamically adapt to evolving population characteristics and ensuring sustained predictive performance [163]. Future research should emphasize the comparison between ML and traditional statistical models (such as CPH models), using statistical tests and clinical utility assessments (such as calibration curves, DCA, and cost-effectiveness analysis) to determine whether their differences are statistically significant and clinically meaningful, rather than relying solely on performance metrics [166]. Moreover, to clearly assess and advance the clinical translation of different models, it is imperative to establish a corresponding phased implementation pathway. For models still in the proof-of-concept stage and reliant on high-cost or non-conventional data (such as most multi-omics or high-resolution imaging), their current primary value should be recognized as mechanistic exploration and biomarker discovery, with a focus on optimizing their discriminative performance and biological interpretability in well-controlled research settings. For models that have already been validated using routine clinical data, are cost-effective, and easily integrated, efforts should be accelerated to advance their external validation and simplification studies in real-world, multi-center environments. This involves exploring whether core performance can be maintained using fewer, more accessible variables, and ultimately evaluating their practical utility in improving patient outcomes through the development of user-friendly tools and integration into clinical workflows [48, 148, 160]. Most critically, efforts should actively integrate prediction models with clinical treatment decisions and outcomes, establishing a seamless management pathway that connects risk prediction, treatment evaluation, and targeted intervention. For example, the Lifetime Prognosis Model for Diabetes can use clinical data from patients with T2D to evaluate the effects of different treatment strategies (such as smoking cessation, blood pressure control, and glucose-lowering) on extending survival years free of myocardial infarction or stroke. Based on this, the model can provide personalized recommendations regarding whether to initiate treatment, the intensity of therapy, and combination drug use [167]. Additionally, while current models predominantly focus on “patients already diagnosed with DKD,” the predictive focus must shift earlier. Future work should develop risk assessment tools targeting individuals with prediabetes or early-stage diabetes, incorporating early indicators like genetic risk, to reduce the incidence of DKD from its source.

Conclusion and future perspectives

This review systematically synthesizes the current research landscape, methodological approaches, and clinical translation progress in DKD prediction models. Presently, these models have evolved into two core directions: diagnostic and prognostic. Leveraging cross-sectional studies and cohort studies respectively, and empowered by AI algorithms including machine learning, diagnostic and prognostic models can effectively achieve non-invasive early screening for DKD and predict the risk of complications such as ESRD, CVD, and death. Furthermore, the integration of multimodal data has become pivotal for breakthrough performance. Models that further incorporate omics, imaging, fundus parameters, and renal pathological features not only compensate for the informational limitations of traditional clinical indicators but also enhance discriminative capability. However, prevailing challenges include insufficient external validation, limited clinical translation, and suboptimal interpretability. Future research should focus on advancing multi-center, large-sample external validation, enhancing model transparency and interpretability, and actively exploring deep integration with EHR systems. Simultaneously, these models must undergo continuous updates by incorporating the latest therapeutic advances, thereby constructing a predictive system capable of dynamically adapting to the evolving treatment landscape. These efforts aim to build individualized risk-stratified management systems capable of dynamic monitoring, thereby laying a groundwork for future precision medicine and ultimately improving long-term patient outcomes.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Author contributions

All authors contributed to the interpretation and discussion of the evidence. YZW performed the literature search and drafted the manuscript. JZW contributed to the conceptualization and overall design of the review. HC revised the manuscript. All authors read and approved the final manuscript.

Funding

The author(s) declare that financial support was received for the research and/or publication of this article. This study was supported by the National Natural Science Foundation project of China (C031002), Gansu Provincial Science Foundation of China (21JR7RA416, 25JRRA597) and Lanzhou University Second Hospital’s “Cuiying Technology Innovation” Program (CY2018-ZD02).

Data availability

Not applicable.

Declarations

Ethics approval and consent to participate

Not applicable.

Conflict of interest

The authors declare that they have no competing interests.