Machine learning and deep learning in predicting the risk of diabetic kidney disease: A systematic review and meta-analysis

Abstract

BACKGROUND

Machine learning (ML) and deep learning (DL) algorithms have been utilised to predict the risk of diabetic kidney disease (DKD) in individuals with type 2 diabetes mellitus (T2DM). Despite promising results, concerns regarding the clinical applicability and performance of these artificial intelligence models remain.

AIM

To evaluates the performance of ML-based and DL-based models in predicting DKD risk among T2DM patients.

METHODS

A systematic search was conducted across five databases. The risk of bias was assessed using the prediction model risk of bias assessment tool. After data extraction, summary point estimates of the area under the receiver operating characteristic curve (AUC) were aggregated. Heterogeneity was evaluated with the I² statistic and Cochrane Q test, and subgroup analyses were performed to identify potential sources of heterogeneity.

RESULTS

Twelve eligible studies were included. The pooled AUC for the top-performing artificial intelligence models was 0.858 [95% confidence interval (CI): 0.779-0.912], with a prediction interval of 0.480-0.975. Significant heterogeneity was detected (I² = 99.7%). Studies employing cross-validation methods demonstrated significantly higher diagnostic accuracy (pooled AUC = 0.88; 95%CI: 0.79-0.94) compared to those using simple holdout validation (pooled AUC = 0.77; 95%CI: 0.59-0.89, P = 0.0231). Predictive factors most frequently used for DKD prediction included age, body mass index, estimated glomerular filtration rate, serum creatinine, urinary albumin, glycated hemoglobin, systolic blood pressure, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, and triglycerides.

CONCLUSION

ML and DL algorithms exhibit strong performance in predicting DKD in patients with T2DM. However, future research should focus on standardizing model development and validation processes.

Key Words: Diabetic kidney disease; Type 2 diabetes mellitus; Predicting; Deep learning; Machine learning; Artificial intelligence

Core Tip: Machine learning and deep learning algorithms show great performance in predicting diabetic kidney disease (DKD) among type 2 diabetes mellitus patients. Predictors, such as age, body mass index, estimated glomerular filtration rate, serum creatinine, urinary albumin, glycated hemoglobin, systolic blood pressure, low-density lipoprotein cholesterol, high-density lipoprotein cholesterol, and triglycerides, play significant roles in DKD prediction. Models employing cross-validation methods exhibit superior predictive capability for DKD compared to those using holdout validation approaches.

INTRODUCTION

Diabetic kidney disease (DKD), a major microvascular complication of type 2 diabetes mellitus (T2DM), is clinically characterized by proteinuria and a rapid decline in estimated glomerular filtration rate (eGFR)[1,2]. Currently, approximately 589 million adults worldwide live with diabetes, a figure projected to increase to 853 million by 2050[3]. Individuals with DKD are at risk of progressing to end-stage renal disease, posing a significant challenge to global public health systems[4,5]. Despite guidelines recommending screening for DKD at the time of T2DM diagnosis, which includes assessment of urinary albumin-to-creatinine ratio (UACR) and eGFR, many patients present with irreversible renal damage at the time of diagnosis due to the insidious progression of DKD[6]. Consequently, early identification of high-risk individuals is critical for preventing microvascular complications.

Recent advancements in machine learning (ML) and deep learning (DL) techniques have shown great promise in early disease detection and clinical management[7-9]. These computational methods can effectively integrate and analyze multi-modal data sources, including clinical text, medical imaging, behavioural patterns, and physiological parameters, to uncover complex pathophysiological relationships between DKD and its multifactorial determinants[10-12]. Although previous meta-analyses have evaluated ML applications in DKD research, they have not sufficiently distinguished between predictive and diagnostic modelling paradigms[13-17]. Such methodological confusion may obscure the practical value of predictive models in assessing the risk of DKD. To address this research gap, this systematic review focuses on predictive models for DKD based on longitudinal cohorts. The core objectives of this study are: First, to systematically evaluate the overall performance of ML and DL models in predicting the risk of DKD in T2DM; second, to explore the sources of heterogeneity across studies through subgroup analysis; and third, to provide a solid and specialized evidence base for the future development of DKD risk prediction tools with greater generalizability and clinical utility.

MATERIALS AND METHODS

Protocol registration

This study adhered to the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines and was conducted in accordance with the Cochrane Collaboration[18]. The protocol has been registered with PROSPERO (CRD420251080292). Review questions and data extraction were structured using the critical appraisal and data extraction for systematic reviews of prediction modelling studies checklist[19]. All aspects of the protocol, including literature searching, screening, data extraction, quality assessment, and risk of bias evaluation, were performed independently by two reviewers.

Study selection

A systematic search was conducted across five databases (PubMed, Web of Science, EMBASE, MEDLINE, and the Cochrane Library) up to June 25, 2025. Literature identification and screening followed the PICOTS (population, intervention, comparator, outcome, timing, setting) framework, with inclusion criteria as follows: (1) Population: Patients with confirmed T2DM and eGFR > 60 mL/minute/1.73 m²; (2) Index model: Any models utilising ML or DL algorithms; (3) Comparator model: Not applicable; (4) Outcome: Development of DKD among patients with T2DM; (5) Timing: From T2DM diagnosis onward, including all reported prediction timeframes; and (6) Setting: Both hospitalised and ambulatory care populations. The full search syntax is detailed in Supplementary Table 1.

Risk of bias assessment

The risk of bias was assessed using the prediction model risk of bias assessment tool[20], which evaluates four key bias domains: Participants, predictors, outcome, and analysis, with 20 signalling questions. For each domain, the risk was classified as “low”, “high”, or “unclear” based on responses to the signalling questions. The overall risk of bias assessment was derived from evaluations across all domains. Models lacking external validation were automatically considered high risk, regardless of domain-specific assessments.

Data extraction

Data extraction followed the critical appraisal and data extraction for systematic reviews of prediction modelling studies and Transparent Reporting of a multivariable prediction model of Individual Prognosis Or Diagnosis + artificial intelligence checklists, utilising standardised forms[21]. Comprehensive study data extracted included data sources, sample characteristics (size and demographics), outcome definitions, predictors, missing data handling, statistical methods, validation approaches, and performance metrics.

Statistical analysis

Statistical analyses were performed using R software (version 4.4.2). Continuous variables with normal distribution are reported as mean ± SD, while non-normally distributed data are presented as median (interquartile range). For the meta-analysis of model discrimination performance, area under the receiver operating characteristic (ROC) curve [area under the receiver operating characteristic curve (AUC)], values were pooled using a logit transformation approach. Each AUC value was logit-transformed, and the standard error (SE) was calculated using the delta method. Pooled logit-transformed AUC values and SEs were then used in the meta-analysis. Heterogeneity among studies was assessed using I² statistics and the Cochrane Q test. If significant heterogeneity was found (I² > 50% or Q test P < 0.10), a random-effects model was employed for pooling. Subgroup analyses were conducted to explore potential sources of heterogeneity. Sensitivity analyses using the leave-one-out method were performed to evaluate the robustness of the findings. A two-sided P value < 0.05 was considered statistically significant.

RESULTS

Study selection

After systematically searching five databases, 1125 studies were initially identified. Duplicate records were removed using EndNote X9 software, resulting in 206 exclusions. Following title and abstract screening, 810 records were excluded. A full-text review of the remaining studies led to the exclusion of an additional 97 records due to the following reasons: Reviews, non-human studies, incompatible demographic characteristics, absence of ML and DL, diagnostic model construction, and mismatched outcomes. Ultimately, 12 studies met all predefined inclusion criteria and were included in the final analysis[22-33]. The study selection process, following Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines, is illustrated in Figure 1.

Figure 1  PRISMA flow diagram.

Study characteristics

The characteristics of the 12 studies are summarised in Table 1[22-33]. Nine studies obtained data from electronic health records (EHRs)/electronic medical records (EMRs)[22,23,25-27,29-32], one study utilised data from the Action to Control Cardiovascular Risk in Diabetes study[24], another from the Singapore Epidemiology of Eye Diseases study[28], and one study used data from the DApagliflozin Real-World evIdeNce-Renal study[33]. Five studies were single-centre[22,26,29,31,32], while seven were multicentre[23-25,27,28,30,33]. Most studies enrolled T2DM participants who were free from DKD at baseline. For DKD diagnostic criteria, all 12 studies primarily relied on urinary albumin measurements, eGFR, and UACR[22-33].

Table 1 Characteristics of included studies.

Ref.	Country	Data source	Recruitment period	Single-center/multi-center	Sample size	Inclusion criteria	Definition of positive outcome
Cho et al[22], 2008	South Korea	EHRs	1996-2005	Single-center	292	Confirmed T2DM; age ≥ 18 years; patients underwent up to 20 clinical tests	20 μg/minute to 200 μg/minute in urinary albumin; no microalbumin or renal failure at diabetes diagnosis; prior evidence of diabetic retinopathy
Dagliati et al[23], 2018	Italy	EHRs	2012-2016	Multi-center	943	Confirmed T2DM; no pre-existing complications; sufficient follow-up records	eGFR < 60 mL/minute/1.73 m2; UACR=30 mg/gram to 299 mg/gram (≥ 2 morning samples)
Rodriguez-Romero et al[24], 2019	United States	ACCORD dataset	2001.01-2001.06; 2003.02-2005.10	Multi-center	10251	Confirmed T2DM by ADA criteria; age ≥ 40 with CVD history; age ≥ 55 with high CVD risk; HbA1c 7.5%-9% (multiple drugs); 7.5%-11% (fewer drugs)	Baseline SCr doubling; eGFR < 60 mL/minute/1.73 m2; UACR ≥ 30 mg/gram or 3.4 mg/mmol; UACR ≥ 300 mg/gram or 33.9 mg/mmol; renal failure
Allen et al[25], 2022	United States	EHRs	2007-2020	Multi-center	111046	Confirmed T2DM by ICD-9/10 codes; age ≥ 18 years; ≥5 years follow-up records; required baseline tests (1-year pre-T2DM); albuminuria/reduced eGFR allowed at baseline; no pre-existing CKD or renal transplant	DKD confirmed by ICD-9/10 codes
Dong et al[26], 2022	China	EMRs	2008.10-2019.12	Single-center	2809	Confirmed T2DM by ADA criteria; age ≥ 18 years; 3 years follow-up records	UACR > 30 mg/gram; protein excretion rate > 150 mg/24 hours; urine dipstick test ≥ 1+; eGFR < 60 mL/minute/1.73 m2
Nicolucci et al[27], 2022	Italy	EMRs	NA	Multi-center	147664	Confirmed T2DM by ICD-9 CM codes	Confirmed DKD by ICD-9 CM codes; no pre-existing complication at baseline
Sabanayagam et al[28], 2023	Singapore	SEED study	2004-2011	Multi-center	1365	Random glucose ≥ 11.1 mmol/L; HbA1c ≥ 6.5% (48 mmol/mol); self-reported antidiabetic medication use; diagnosed with diabetes by a physician	eGFR < 60 mL/minute/1.73 m2; ≥ 25% eGFR decline during follow-up
Hosseini Sarkhosh et al[29], 2023	Iran	EHRs	2012-2021	Single-center	1907	Confirmed T2DM by ADA criteria; 5 years follow-up records; no pre-existing DKD at baseline	UACR ≥ 30 mg/gram; eGFR ≤ 60 mL/minute/1.73 m2
Yun et al[30], 2024	China	EHRs	NA	Multi-center	6040	Confirmed T2DM; age ≥ 18 years; ≥ 7 years follow-up records; ≥ 2 hospital visits per year; no pre-existing DKD at baseline or within 2 years	UACR ≥ 30 mg/gram (≥ 3 months); eGFR < 60 mL/minute/1.73 m2; exclude other kidney diseases
Lin et al[31], 2025	China	EHRs	2011-2023	Single-center	3291	Confirmed T2DM by ICD-9/10 codes; age ≥ 18 years; at least one hospital visit following the end of the baseline time window	eGFR < 60 mL/minute/1.73 m2; UACR ≥ 30 mg/gram over 3 months; DKD confirmed by ICD-9/10 codes; protein in urine dipstick test ≥ 1+
Zou et al[32], 2025	China	EHRs	2014.01-2022.12	Single-center	9572	Confirmed T2DM; no pre-existing DKD; ≥ 3 years follow-up records; exclude glomerular/systemic diseases	UACR ≥ 30 mg/gram in 2/3 tests within 3 months to 6 months; eGFR < 60 mL/minute/1.73 m2 for more than 3 months; renal biopsy was consistent with DKD pathological changes
Dei Cas et al[33], 2025	Italy	DARWIN-Renal	2015.01-2021.09	Multi-center	22379	Confirmed T2DM; age ≥ 18 years; ≥ 1 year follow-up records	5 renal disease severity thresholds established by the KDOQI scale based on the eGFR

EHRs: Electronic health records; T2DM: Type 2 diabetes mellitus; eGFR: Estimated glomerular filtration rate; UACR: Urine albumin-creatinine ratio; ACCORD: Action to Control Cardiovascular Risk in Diabetes; ADA: American Diabetes Association; CVD: Cardiovascular disease; SCr: Serum creatinine; ICD-9/10 codes: International Classification of Diseases, 9th/10th Revision; CKD: Chronic kidney disease; DKD: Diabetic kidney disease; EMRs: Electronic medical records; NA: Not available; SEED: Singapore Epidemiology of Eye Diseases; HbA1c: Glycated hemoglobin; DARWIN-Renal: DApagliflozin Real-World evIdeNce-Renal; KDOQI: Kidney Disease Outcomes Quality Initiative.

The details of prediction model development and validation are presented in Table 2. Of the 12 included studies[22-33], 10 developed models using ML approaches[22-29,31,32], while 2 employed DL methods[30,33]. For missing data imputation, three studies used multiple imputation by chained equations[31-33], three used the random forest (RF) imputation algorithm (missForest)[23,25,26], two applied mean/mode imputation[28,29], and one study filled missing values with an extreme value (-999)[27]. To address class imbalance, four studies used oversampling[23,24,27,32], including three employing synthetic minority over-sampling technique[24,27,32]. Most studies randomly allocated datasets into development and validation sets with varying proportions. Regarding validation methods, most studies utilised holdout or K-fold validation[24,26-28,32,33], except for Cho et al[22] and Dagliati et al[23], who applied leave-one-out cross-validation (LOOCV), and Hosseini Sarkhosh et al[29], who used recursive feature elimination with cross-validation. Four studies performed external validation to assess model generalizability[25,27,29,33].

Table 2 Characteristics of the best-performing model from each included study.

Ref.	Number	Algorithm type	Missing data imputation	Oversample	Train-test split	Validation type	Internal validation method	Best model	Prediction horizon, year	AUC (95%CI)
Cho et al[22], 2008	292	ML	NA	NA	NA	Internal	LOOCV	SVM	1	0.969 (0.941-0.997)
Dagliati et al[23], 2018	943	ML	missForest	Oversampling	NA	Internal	LOOCV	LR	3	0.808 (0.772-0.845)
Rodriguez-Romero et al[24], 2019	10251	ML	NA	SMOTE	66:34	Internal	CV-10	RF	1-1.9	0.730 (0.715-0.745)
Allen et al[25], 2022	111046	ML	missForest	NA	7:1	External	Holdout	XGBoost	5	0.750 (0.734-0.766)
Dong et al[26], 2022	861	ML	missForest	NA	8:2	Internal	CV-5	LightGBM	3	0.815 (0.747-0.882)
Nicolucci et al[27], 2022	147664	ML	Extra-values imputation	SMOTE	Yes	External	CV-10	XGBoost	2	0.970 (0.968-0.972)
Sabanayagam et al[28], 2023	1365	ML	Mean/mode	NA	8:2	Internal	CV-5	Elastic Net	6	0.851 (0.847-0.856)
Hosseini Sarkhosh et al[29], 2023	3444	ML	Mean/mode	NA	8:2	External	RFECV	RF	5	0.790 (0.770-0.820)
Yun et al[30], 2024	6040	DL	NA	NA	7:3	Internal	Holdout	LSTM	5	0.830 (0.807-0.853)
Lin et al[31], 2025	3291	ML	MICE	NA	7:3	Internal	Holdout	SuperLearner	2.53	0.714 (0.673-0.755)
Zou et al[32], 2025	12190	ML	MICE	SMOTE	8:2	Internal	CV-5	LightGBM	3	0.918 (0.906-0.930)
Dei Cas et al[33], 2025	32379	DL	MICE	NA	8:2	External	CV-5	RNN	2	0.887 (0.869-0.904)

AUC: Area under the receiver operating characteristic curve; ML: Machine learning; NA: Not available; LOOCV: Leave-one-out cross-validation; SVM: Support vector machine; LR: Logistic regression; SMOTE: Synthetic minority oversampling technique; CV-10: 10-fold cross-validation; RF: Random forest; CV-5; 5-fold cross-validation; RFECV: Recursive feature elimination with cross-validation; DL: Deep learning; LSTM: Long short-term memory; RNN: Recurrent neural network.

Risk of bias assessment

Risk of bias assessment revealed that 12 studies[22-33] were rated as having a high risk of bias based on prediction model risk of bias assessment tool (Supplementary Figure 1). Bias in participant selection was identified in studies that extracted data from single-center EHRs or EMRs. In the analytical domain, one study had fewer than 10 events per variable, three failed to handle missing data adequately[22,24,30], and eight studies did not address potential overfitting[22,25,26,28-31,33]. Additionally, five studies did not perform feature selection[24,25,27,30,31].

Synthesised results

The AUC values across the 12 included studies ranged from 0.714 to 0.970[22-33] (Figure 2). Due to substantial heterogeneity, a random-effects model was employed for synthesising AUC values. After applying the Hartung-Knapp adjustment, the pooled AUC was estimated at 0.858 (95%CI: 0.779-0.912), with a prediction interval spanning (0.480-0.975).

Figure 2 Pooled area under the receiver operating characteristic curve values in internal validation sets. HK: Hausman-Kuipers; CI: Confidence interval.

Source of heterogeneity

This meta-analysis revealed extremely high heterogeneity in AUC values among the studies (I² = 99.7%, P = 0, Figure 2). The Baujat plot indicated that Nicolucci et al[27], contributed most to the overall heterogeneity and was identified as a primary source of variation (Supplementary Figure 2A). The Galbraith plot further highlighted that the effect sizes of Nicolucci et al[27] and Rodriguez-Romero et al[24], significantly deviated from the regression line (Supplementary Figure 2B). Although excluding Nicolucci et al[27], slightly reduced the pooled estimate, the change was minimal, and heterogeneity remained largely unchanged (I² = 98.4%, P < 0.0001). Consequently, Nicolucci et al[27], was retained in the analysis, and subgroup analyses were conducted to explore other potential sources of heterogeneity.

Subgroup analyses were performed based on factors such as region, study type, algorithm type, study center design, use of external validation, and prediction horizon. All subgroup analyses yielded non-significant P value (all P > 0.05; Table 3), suggesting that none of these factors significantly explained the observed heterogeneity. However, when the studies were categorised into holdout and non-holdout groups based on internal validation methods, statistically significant heterogeneity was observed (P = 0.0231, Figure 3). Studies employing cross-validation methods, including k-fold cross-validation, LOOCV, and repeated cross-validation, exhibited markedly superior diagnostic accuracy (pooled AUC = 0.88; 95%CI: 0.79-0.94; Figure 3) compared to those utilising simple holdout validation (pooled AUC = 0.77; 95%CI: 0.59-0.89; Figure 3).

Figure 3 Subgroup analysis comparing cross-validation and holdout validation methods. CI: Confidence interval.

Table 3 Subgroup analysis of pooled area under the receiver operating characteristic curve values in internal validation.

Subgroup variable	Level	Studies	AUC (95%CI)	I2	t11	P value	χ2	P value
Region	Overall				7.34	< 0.0001	0.01	0.9422
	Asian	7	0.86 (0.74-0.92)	96.7%
	Western	5	0.86 (0.78-0.91)	99.7%
Study type	Overall	12	0.86 (0.78-0.91)	85.8%	7.34	< 0.0001	0.47	0.4951
	Prospective study	3	0.90 (0.40-0.99)	99.7%
	Retrospective study	9	0.85 (0.74-0.91)	99.7%
Algorithm type	Overall				7.34	< 0.0001	0.00	0.9563
	ML	10	0.86 (0.76-0.92)	99.7%
	DL	2	0.86 (0.23-0.99)	99.7%
Center	Overall	12	0.86 (0.78-0.91)	99.7%	7.34	< 0.0001	0.02	0.8770
	Single-center	5	0.86 (0.66-0.96)	97.7%
	Multicenter	7	0.85 (0.78-0.91)	99.8%
Validation	Overall	12	0.86 (0.78-0.91)	99.7%	7.34	< 0.0001	0.26	0.6107
	Internal validation	8	0.84 (0.74-0.91)	98.5%
	External validation	4	0.86 (0.78-0.91)	99.9%
Prediction horizon	Overall	12	0.86 (0.78-0.91)	99.7%	7.34	< 0.0001	0.92	0.3362
	< 3 years	5	0.90 (0.64-0.98)	99.8%
	≥ 3 years	7	0.83 (0.77-0.88)	98.0%

AUC: Area under the receiver operating characteristic curve; CI: Confidence interval.

Sensitivity analysis and publication bias

A leave-one-out sensitivity analysis was conducted (Supplementary Figure 3). Point estimates consistently ranged between 0.82 and 0.88 after iteratively excluding individual studies, indicating minimal variation. The confidence intervals remained similar across all iterations. Egger’s and Begg’s tests were used to assess publication bias, and no significant bias was detected (Supplementary Figure 4). After imputing one study using the Trim-and-Fill method, the pooled AUC was 0.867 (95%CI: 0.796-0.917). The effect size showed only marginal change after imputation, and the width of the confidence interval remained largely unchanged.

Predictors

A systematic analysis was performed on the characteristic variables integrated into the optimal prediction models developed in 12 studies on DKD[22-33]. These variables were categorised into three primary groups based on their attributes: Demographic characteristics, medical history, and laboratory indicators. Demographic characteristics mainly included age, gender, body mass index (BMI), and diabetes duration (Table 4). Age was incorporated in eight studies, BMI in gender, gender in four, and diabetes duration in three. Medical history variables focused on key risk factor exposures, such as hypertension, smoking, and alcohol consumption. Hypertension history appeared in four studies, smoking in three, and alcohol consumption in two. Laboratory indicators formed the core of the prediction models, covering various dimensions including blood pressure parameters, renal function indicators, glycemic control levels, lipid profiles, complete blood count, and liver function. Glycated hemoglobin (HbA1c) was included in eight studies, low-density lipoprotein (LDL) in seven, eGFR in gender, and systolic blood pressure (SBP) in gender. Additionally, Sabanayagam et al[28], and Dei Cas et al[33], also included antidiabetic medication usage.

Table 4 Predictors in diabetic kidney disease prediction models.

Ref.	Number	Demographic	Past history	Laboratory
Cho et al[22], 2008	39	BMI (mean, minimum)	NA	SBP (mean, EST, and initial), DBP (slope and maximum), WBC (initial), WBC (latest), hemoglobin, platelet (slope), platelet (var), platelet (K), platelet (EST), cholesterol (K and EST), AST (K and initial), ALT (initial, minimum, and K), ALP (minimum and latest), creatinine (mean and var), uric acid (var), Na (EST), K (slope and K), triglyceride (EST), HDL-C (var and initial), LDL-C (EST and initial), microalbumin (mean, var, maximum, and minimum)
Dagliati et al[23], 2018	4	BMI	Smoking, hypertension	HbA1c
Rodriguez-Romero et al[24], 2019	18	Age	NA	UAlb at baseline, eGFR at baseline, UCr at baseline, eGFR from baseline to year 1, trig at baseline, FPG at baseline, CPK at baseline, age, eGFR at month 4, K at baseline, LDL at baseline, Chol at baseline, eGFR at month 12, FPG at month 4, FPG from baseline to year 1, trig from baseline to year 1, UACR at baseline, FPG at month 8
Allen et al[25], 2022	15	Age, gender, BMI	Acute kidney injury, chronic heart failure, smoking, drinking	SBP, DBP, blood urea nitrogen, creatinine, eGFR, cholesterol (HDL and LDL), white cell count
Dong et al[26], 2022	8	Age, BMI	NA	Hcy, HbA1c, ALB, eGFR, bicarbonate, LDL
Nicolucci et al[27], 2022	46	Gender, age, height, weight, BMI, waist circumference, diabetes duration	NA	SBP, DBP, ankle/brachial index DX, ankle/brachial index SX, fasting blood glucose, blood glucose after breakfast, blood glucose before dinner, blood glucose before lunch, blood glucose at 11:00 pm, blood glucose after lunch, blood glucose after dinner, pre-prandial blood glucose, post-prandial blood glucose, HbA1c, albuminuria, serum creatinine, creatinine clearance, total cholesterol, LDL cholesterol, HDL cholesterol, triglycerides, fibrinogen, GGT, ALT, AST, alkaline phosphatase, amylase, CPK, hemoglobin, platelets, BUN, uric acid, glycosuria, urinary amylase, urinary ketones, urinary potassium, urinary sodium, urine creatinine, urine culture
Sabanayagam et al[28], 2023	15	Ethnicity (Malay/Chinese)	Diabetic retinopathy, hypertension	Acetate, diabetic retinopathy, SBP, DHA, GFR-EPI, HbA1c, IDL-CE%, M-HDL-PL%, M-VLDL-PL%, S-HDL-FC%, XL-HDL-CE%, anti-DM meds
Hosseini Sarkhosh et al[29], 2023	6	Diabetes duration	Hypertension, CVD	HbA1c, eGFR, ACR
Yun et al[30], 2024	10	Age	NA	PP, SBP, variabilities of PP and SBP, Scr, HDL-C, HbA1c, TG, variabilities of HbA1c
Lin et al[31], 2025	46	Gender, age	Smoking, drinking	WBC count, neutrophil count, lymph count, mono count, eosinophil count, basophil count, red blood cell count, hemoglobin, hematocrit, mean corpuscular volume, mean corpuscular hemoglobin, mean corpuscular hemoglobin concentration, red cell distribution width-coefficient of variation, platelet, mean platelet volume, platelet distribution width, lymph, mono, eosinophil, basophil, ALT, AST, AST/ALT (AST ALT), total bilirubin, creatinine, Ca, total cholesterol, triglyceride, HDL, LDL, prealbumin, HbA1c, D-Dimer, pH, glucose, billing, ketone, obstetric, nitrogen, urology
Zou et al[32], 2025	5	Age	NA	UACR, Cystatin C, eGFR, Neutrophil
Dei Cas et al[33], 2025	34	Gender, age, weight, weight-past, BMI, diabetes duration	CKD	SBP, DBP, SBP in the past, DBP in the past, HbA1c, HbA1c in the past, eGFR, eGFR in the past, AER, ACR > 30, AER in the past, and anti-DM Meds

BMI: Body mass index; NA: Not available; SBP: Systolic blood pressure; EST: Estimated value on the date of prediction; DBP: Diastolic blood pressure; WBC: White blood cell; AST: Aspartate aminotransferase; ALT: Alanine aminotransferase; ALP: Alkaline phosphatase; var: Variance; HDL: High-density lipoprotein; LDL: Low-density lipoprotein; HbA1C: Glycosylated hemoglobin; eGFR, Estimated glomerular filtration rate; UAlb: Uninary albumin; UCr: Urinary creatinine; Trig: Triglycerides; FPG: Fasting plasma glucose; CPK: Creatinine phosphokinase; UACR: Urine albumin-creatinine ratio; Chol: Cholesterol; Hcy: Homocysteine; ALB: Albumin; GGT: Gamma-glutamyl transferase; CVD: Cardiovascular disease; BUN: Blood urea nitrogen; DHA: Docosahexaenoic acid; GFR-EPI: Glomerular filtration rate estimated using; IDL-CE%: Cholesterol esters to total lipids ratio in IDL; M-HDL-PL%: Phospholipids to total lipids ratio in medium HDL; M-VLDL-PL%: Phospholipids to total lipids ratio in medium VLDL; S-HDL-FC%: Free cholesterol to total lipids ratio in small HDL; XL-HDL-CE%: Cholesterol esters to total lipids ratio in very large HDL; AER: Albumin excretion rate; ACR: Albumin-creatinine ratio.

DISCUSSION

DKD accounts for approximately 30% to 50% of global end-stage renal disease cases[34]. Therefore, early prediction of DKD is critical for preserving renal function in patients with T2DM. Conventional screening primarily relies on cross-sectional measurements of single or limited biomarkers, such as a spot UACR > 30 mg/gram or an eGFR < 60 mL/minute/1.73 m²[35]. This approach essentially identifies established kidney injury[36]. In contrast, ML and DL models can integrate multi-dimensional longitudinal data. This enables the earlier identification of high-risk individuals before irreversible structural kidney damage occurs. Therefore, multiple research teams develop various DKD prediction models using diverse statistical methods[7]. To evaluate the clinical applicability and reliability of these AI models, 12 eligible studies were included, and a meta-analysis of the highest-performing prediction models was conducted[22-33]. The pooled AUC from internal validation sets was 0.858 (95%CI: 0.779-0.912), indicating that the top-performing models from the included studies demonstrate strong discriminatory ability in predicting DKD, consistent with findings from other meta-analyses[12-17].

The pooled AUC in our analysis exhibited considerable heterogeneity, which may be attributed to variations in internal validation methods used across the included studies. Four distinct validation approaches were employed: K-fold cross-validation, LOOCV, repeated cross-validation, and hold-out validation. Our analysis revealed that studies utilising hold-out validation demonstrated a lower pooled AUC compared to those employing cross-validation. Despite a statistically significant difference in pooled effect estimates between the hold-out and cross-validation groups, substantial heterogeneity remained within both subgroups, suggesting that the validation method accounts for only part of the variability. Leave-one-out sensitivity analyses confirmed that no single study significantly influenced the findings. Furthermore, other subgroup analyses, examining factors such as region, study type, algorithm type, center design, external validation, and prediction horizon revealed no significant differences. The persistent high heterogeneity implies that the outcomes are influenced by complex, multifactorial interactions or unmeasured covariates, such as subtle population differences, variations in model architecture or hyperparameters, or differing clinical outcome definitions. Further investigation with a larger sample of studies, where possible, and individual participant data meta-analysis, is needed to better identify and adjust for these factors.

DKD is a complex condition whose development and progression are driven by a combination of hemodynamic abnormalities and metabolic disturbances. The pathological mechanisms of DKD include hyperglycemia, accumulation of advanced glycation end products, proteinuria, and lipid overload[37,38]. Our analysis identified the core predictive factors for DKD across the 12 included studies as gender, BMI, eGFR, serum creatinine, urinary albumin, HbA1c, SBP, LDL-cholesterol (LDL-C), high-density lipoprotein cholesterol (HDL-C), and triglycerides[22-33]. These factors are critical for the early identification and risk stratification of DKD. Established predictors such as BMI, LDL-C, fasting blood glucose, SBP, and HbA1c have been confirmed as key risk factors for DKD in previous non-ML studies[39,40]. Unlike traditional statistical methods, which rely on clinically pre-defined variable relationships and linear assumptions to test associations with DKD, ML and DL are data-driven. They autonomously learn and capture complex patterns in the data, including nonlinear relationships and interactions among variables. For example, Wu et al[41], demonstrated that in patients with DKD, T2DM diagnosed before age 40 was associated with poorer glycemic control and more severe proteinuria. Other research shows that for every 5 years to 10 years increase in age, the risk of DKD rises by 38%[42]. In the SHapley Additive exPlanation plot presented by Dong et al[26], both age < 40 years and > 60 years were associated with positive SHapley Additive exPlanation values, suggesting that both younger and older age are independent extremes linked to heightened DKD risk. However, this does not imply that complex algorithms are superior to logistic regression. A meta-analysis found that in the field of chronic kidney disease research, the alleged superiority of non-logistic regression models over traditional logistic regression lacks substantial empirical support. This suggests that the value of ML and DL lie in their abilities to capture complex feature relationships, rather than in the complexities of the algorithms themselves. Beyond age, several modifiable factors significantly influence DKD progression. Intensive glycemic control has been shown to reduce DKD risk[43,44]. Hypertension exacerbates renal dysfunction and increases cardiovascular risk, emphasising the importance of blood pressure control in DKD management[45]. Dyslipidemia, particularly elevated LDL-C or reduced HDL-C levels, plays a critical role in DKD development and progression[46,47], Additionally, smoking has been established as an independent risk factor for both the onset and progression of DKD[48,49].

A key strength of this study is that all included studies were based on longitudinal cohorts for constructing prediction models for DKD. In contrast to diagnostic models derived from cross-sectional data, this design more accurately captures the dynamic progression of the disease and the cumulative effects of long-term risk factors, significantly enhancing the reliability and clinical applicability of the predictions.

However, several methodological limitations should be acknowledged: First, the majority of the included studies relied on EHRs or EMRs for retrospective analysis, which inherently introduces a high risk of selection bias and information bias; second, there was substantial heterogeneity in study designs, with most being single-center retrospective cohorts, further increasing the potential for bias; third, many studies had shortcomings in data preprocessing, including inconsistent methods for handling missing values and inadequate attention to class imbalance issues, which may impact the predictive performance and robustness of the models; fourth, the lack of a standardised model evaluation framework led to the use of diverse evaluation metrics across studies, and some studies did not report complete metrics. This limited the synthesis and comparability of results and increased the complexity of interpretation; fifth, our meta-analysis aims to assess the performances of models to predict the risk of early-stage DKD, its findings are primarily applicable to T2DM populations with normal or mildly impaired baseline renal function (eGFR ≥ 60 mL/minute/1.73 m²). Consequently, our findings are not generalizable to patients with a baseline eGFR < 60 mL/minute/1.73 m², as their disease drivers are more complex, warranting future development of dedicated prediction models; and finally, due to limited access to raw data, an individual patient data meta-analysis could not be conducted, limiting the ability to further adjust for potential confounding factors.

CONCLUSION

In conclusion, ML and DL algorithms demonstrate strong performance in predicting DKD in patients with T2DM. Key predictors such as age, BMI, eGFR, serum creatinine, urinary albumin, HbA1c, SBP, LDL-C, HDL-C, and triglycerides play significant roles in DKD prediction. Models utilising cross-validation methods show superior predictive capability compared to those employing holdout validation. There is, therefore, a need to standardise disease prediction research. In summary, ML and DL have significant clinical value in predicting DKD, offering an effective solution for early prevention and management of the disease.