Predicting future kidney function in type 2 diabetes mellitus using machine learning and baseline health information

Abstract

Diabetic kidney disease (DKD) is a common type 2 diabetes mellitus (T2DM) complication, often leading to end-stage kidney disease. Predicting kidney function decline may enable timely interventions to slow disease progression. We developed machine learning (ML) models to predict future estimated glomerular filtration rate (eGFR) values at specific annual time points among patients with T2DM, utilizing only baseline characteristics obtained from routine medical check-ups. We followed 974 patients for a median of 5.3 years across multiple centers. Fifty-four baseline clinical features and a time-point variable were included in three ML algorithms (Light Gradient Boosting Machine, Random Forest, and Support Vector Machine [SVM]). Their predictive performance was compared with each other and with a benchmark multiple linear regression (MLR) model, evaluated using the coefficient of determination (R²) and root mean squared error, with 95% confidence intervals (95% CIs). The SVM model showed the best prediction performance across a wide eGFR range (R² 0.67 [95% CI 0.62–0.72]), maintaining moderate or better performance for up to 6 years, whereas the MLR model struggled to predict beyond 4 years. These findings suggest that SVM-based prediction using baseline data and time-point information could support long-term management of DKD progression in patients with T2DM.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-45500-6.

Introduction

Diabetic kidney disease (DKD) affects approximately 40% of patients with type 2 diabetes mellitus (T2DM) and is the leading cause of chronic kidney disease (CKD) globally^1–3. DKD is clinically significant because of its progression to end-stage kidney disease (ESKD) requiring dialysis and its link to mortality³. Current research aims to understand the mechanisms and develop diagnostic techniques to monitor the decline in kidney function in patients with T2DM^4,5.

The rate of kidney function decline in DKD varies considerably among patients with T2DM; a subset, known as “rapid decliners,” progresses rapidly to ESKD⁶. Our previous study showed that 14% of Japanese patients with T2DM and preserved renal function and normoalbuminuria experienced an early decline in estimated glomerular filtration rate (eGFR) during a median follow-up of 6.7 years⁷. In that study, a lower baseline hemoglobin level was identified as a risk factor for early eGFR decline; however, it was insufficient to distinguish early eGFR decliners. Therefore, a more precise and long-term predictive model is needed to identify rapid decliners for targeted early intervention.

Machine learning (ML) techniques have revolutionized CKD prediction by enabling the analysis of large datasets and improving predictive accuracy. Supervised learning algorithms, including decision trees, Random Forest (RF), and Support Vector Machines (SVM), are known to effectively classify CKD stages and predict risk⁸. However, only a few studies have focused on developing a long-term prediction model for future eGFR values at upcoming check-ups in patients with CKD^9,10. To date, none have leveraged ML to predict future long-term changes in eGFR utilizing only baseline characteristics routinely collected during medical check-ups among Japanese patients with T2DM.

In this study, we aimed to develop prediction models for eGFR values at specific future time points using three ML algorithms (Light Gradient Boosting Machine [LightGBM]¹¹, RF, and SVM) based solely on baseline clinical characteristics of patients with T2DM. The primary objective was to provide an early warning system for clinicians to identify patients at a higher risk of future kidney function decline using only information available at the initial visit, such as during an initial outpatient consultation, where longitudinal data may be limited. By predicting eGFR values at future follow-up visits, the early warning system may support timely, individualized clinical decision-making and help reduce the risk of progression to ESKD.

Methods

Study design and population

This study was a part of a multi-center, longitudinal, observational study on diabetic complications⁷. Between March 2010 and June 2020, 1,305 patients with T2DM aged 20–74 years were enrolled from three hospitals. Exclusion criteria included secondary diabetes; kidney diseases such as glomerulonephritis or nephrotic syndrome; dialysis; prior cardiovascular or cerebrovascular treatment; occlusive arteriosclerosis or lower limb amputation; a cancer diagnosis within 5 years before enrollment; or infection with hepatitis B, hepatitis C, human immunodeficiency virus, or syphilis. After exclusions, 1,112 patients were followed up longitudinally, with clinical and biochemical data collection at least annually (every 12 months). Follow-up was discontinued if patients developed any listed conditions. Patients with < 1 year of follow-up or fewer than three eGFR measurements were further excluded. The last observation date was June 10, 2022. Additional details regarding the study design and population are provided in the Supplementary Methods.

Ethics approval and informed consent

The study protocol was approved by the institutional review boards of all participating centers (chief approval number: NCGM-S-004622-03) and was conducted following the tenets of the Declaration of Helsinki. All data used were anonymized. Informed consent was obtained using an opt-out procedure.

Study visits and measurements

Demographics, comorbidities, and medication use information at baseline and during follow-up were obtained from medical records and laboratory test results (Supplementary Methods). eGFR was calculated using the standard Japanese formula based on inulin clearance¹²:

Candidate baseline predictors and outcome

Fifty-four baseline clinical and biochemical characteristics were selected as predictors, based on their clinical relevance to eGFR and availability in routine medical check-ups. These included demographic data (sex, age), diabetes metrics (duration, family history), physical measurements (body mass index [BMI], blood pressure), laboratory values (liver enzymes, lipids, kidney function markers, glycemic parameters, hematological indices), urine tests (albumin-to-creatinine ratio [ACR], dipstick results), comorbidities, and medication usage (Supplementary Methods). To preserve interpretability and clinical completeness, features were not explicitly excluded for collinearity.

The primary outcome was the eGFR value at specific future follow-up visits. Each model was trained to predict distinct eGFR values at discrete annual time points (including 1-year, 2-year, among others) rather than an overall slope or trajectory. To account for the temporal distance from baseline, a time-point variable (coded as 1–9 years) was included in all ML models, enabling time-aware prediction. No post-baseline variables were used to prevent data leakage.

Statistical analysis and model development/evaluation

Continuous variables are presented as medians with interquartile ranges (IQRs), and categorical variables as counts and proportions for each category. We did not explicitly model censoring processes (e.g., death or loss to follow-up). At each time point, observations were included only if the target eGFR measurement was available; otherwise, the observation was excluded for that time horizon. To predict future eGFR values at each annual time point, we developed three ML algorithms: LightGBM, RF, and SVM. For the SVM, we used support vector regression with a radial basis function kernel. The dataset was randomly divided into training (75%) and test (25%) sets on a per-patient basis, stratified by the number of eGFR measurements per patient, to ensure dataset independence and prevent information leakage. No temporal or hospital-wise splits were applied. The training dataset was further split into training (80%) and validation (20%) datasets.

Missing eGFR values were not imputed. For aspartate aminotransferase, alanine aminotransferase, and γ-glutamyl transpeptidase (γ-GTP), which had relatively high missing rates owing to optional testing in routine check-ups, binary indicators of missingness (1 = missing; 0 = observed) were included as additional model features, as the missing tests may, for example, reflect less clinical suspicion of liver dysfunction. For other features, missing values were not imputed for LightGBM, but were imputed using median values for RF, SVM, and multiple linear regression (MLR). All continuous variables were standardized for all models, and categorical variables were one-hot encoded for SVM and MLR. These preprocessing steps were applied consistently before model training and evaluation to ensure compatibility across models.

Hyperparameter optimization was performed using Optuna, a Bayesian optimization framework, to enhance predictive performance for LightGBM, RF, and SVM models¹³. The model settings, including parameter ranges and tuning tests, are summarized in Supplementary Table S1. Each model was trained using five-fold cross-validation, with hyperparameter tuning performed within the cross-validation loop. Predictions were generated independently for each time point using baseline features and the time-point variable, rather than extrapolating a continuous trajectory. Model performance was evaluated using the coefficient of determination (R²) and root mean squared error (RMSE), with 95% confidence intervals (CIs). Calibration was assessed by regressing observed eGFR values against predicted eGFR values to estimate the calibration slope and intercept. A joint hypothesis test was performed to test the null hypothesis of perfect calibration (intercept = 0, slope = 1), and the corresponding p-value was reported. All performance estimates were based on this internal test dataset, as no external validation was conducted. The best models were selected based on their R² scores on the test datasets. For comparative benchmarking, an MLR model (Gauss–Markov type) was also developed.

Selected feature importance was evaluated using SHapley Additive exPlanations (SHAP) values to clarify their contribution to each ML model’s predictive performance¹⁴. After model training, important features were ranked in descending order of SHAP values, and those with cumulative contributions exceeding 90% were retained to balance model interpretability and explanatory power. This feature selection was conducted outside the cross-validation loop to avoid influencing model fitting or validation. SHAP analysis served as a post-hoc interpretability tool to assess global feature contributions and ensure clinical plausibility, rather than to infer causality or estimate directional effects. Local explanations and interaction effects were not emphasized, as our main objective in this study was predictive performance.

For robustness assessment, reduced models were developed using only the top-ranked SHAP features, and their performance was evaluated using the same procedure. Additionally, to assess time-point-specific performance (1–9 years), both training and test datasets were divided into nine categories (each from 1 to 9 years), and the prediction performance was evaluated for each period.

All analyses were performed using Python version 3.12.3 and relevant libraries, including scikit-learn, NumPy, Pandas, and Optuna.

Subgroup analyses

Subgroup analyses were conducted to assess the consistency of model performance across clinically relevant patient subsets. Stratification variables included baseline eGFR stage (CKD stages 1–2 vs. 3–4), age (< 65 vs. ≥65 years), and sex (male vs. female). Reduced models were also developed using only the top-ranked SHAP features for each subgroup. Model performance was assessed using R² and RMSE, with 95% CIs. Calibration was assessed by regressing observed eGFR values against predicted eGFR values to estimate the calibration slope and intercept.

All analyses were performed using Python version 3.12.3 and relevant libraries, including scikit-learn, NumPy, Pandas, and Optuna.

Results

Baseline participant characteristics

Of the 1,305 enrolled patients, 974 met the inclusion criteria and were randomly assigned to the training (n = 730) and test (n = 244) datasets (Fig. 1). Among these patients, 719 (73.8%) were men. The median age was 62 (IQR: 55–66) years, eGFR was 72.4 (IQR: 62.2–82.9) mL/min/1.73 m², and urine ACR was 15.5 (IQR: 6.6–63.5) mg/gCr at enrollment (Table 1). Detailed baseline characteristics of the overall population, as well as training and test datasets, are shown in Supplementary Table S2. Supplementary Table S3 summarizes the baseline characteristics by hospital. Some differences were observed in lipid levels and medication use (such as glucagon-like peptide 1 receptor agonist and sodium-glucose cotransporter 2 inhibitor), particularly at the smallest hospital; however, the overall profiles were similar across centers.

Fig. 1.

Flow chart of participant enrollment and the process of building prediction models.

Table 1.

Clinical characteristics of patients with type 2 diabetes mellitus in the training and test datasets.

Baseline characteristics		Overall (n = 974)	Training dataset (n = 730)	Test dataset (n = 244)
Demographics, median (IQR)
Male, n (%)		719 (73.8)	538 (73.7)	181 (74.2)
Age at registration (years)		62 (55–66)	62 (55–66)	62 (54–66)
BMI (kg/m2)		25.1 (22.5–28.0)	25.0 (22.5–28.0)	25.1 (22.6–28.3)
Current smoker, n (%)		268 (27.5)	193 (26.4)	75 (30.7)
Laboratory characteristics, median (IQR)
sBP (mmHg)		129 (120–137)	129 (120–137)	130 (120–136)
dBP (mmHg)		75 (68–82)	76 (69–83)	74.0 (67–81)
TC (mg/dL)		184 (164–206)	183 (165–205)	188 (164–208)
HDL-C (mg/dL)		54 (44–64)	54 (44–64)	54 (45–67)
TG (mg/dL)		117 (84–177)	117 (84–177)	120 (85–177)
BUN (mg/dL)		14 (12–17)	14 (12–17)	15 (12–18)
eGFR (mL/min/1.73 m2)		72.4 (62.2–82.9)	72.5 (62.1–82.5)	72.3 (62.9–84.6)
Uric acid (mg/dL)		5.7 (4.9–6.6)	5.7 (4.9–6.6)	5.7 (4.8–6.6)
HbA1c (%)		6.9 (6.5–7.5)	6.9 (6.5–7.5)	7.0 (6.5–7.6)
WBC (103/mL)		6.2 (5.1–7.6)	6.2 (5.1–7.5)	6.4 (5.3–7.7)
RBC (106/mL)		4.6 (4.3–5.0)	4.6 (4.3–5.0)	4.7 (4.4–4.9)
Hemoglobin (g/dL)		14.4 (13.5–15.3)	14.3 (13.4–15.3)	14.6 (13.6–15.4)
Platelet count (103/mL)		225 (190–262)	226 (192–262)	222 (187–263)
ACR (mg/gCr)		15.5 (6.6–63.5)	15.5 (6.6–63.5)	15.8 (6.5–63.9)
Comorbidities, n (%)
Diabetic Retinopathy		267 (27.4)	192 (26.3)	75 (30.7)
Hypertension		360 (37.0)	271 (37.1)	89 (36.5)
Medications, n (%)
Insulin		175 (18.0)	130 (17.8)	45 (18.4)
Biguanide		558 (57.3)	409 (56.0)	149 (61.1)
ACE inhibitor or ARB		529 (54.3)	398 (54.5)	131 (53.7)
Statin		468 (48.0)	349 (47.8)	119 (48.8)

ACE inhibitor, angiotensin-converting enzyme inhibitor; ACR, urine albumin-to-creatinine ratio; ARB, angiotensin II receptor blocker; BMI, body mass index; BUN, blood urea nitrogen; dBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; HbA1c, hemoglobin A1c; HDL-C, HDL-cholesterol; RBC, red blood cell count; sBP, systolic blood pressure; TC, serum total cholesterol; TG, triglyceride; WBC, white blood cell count.

Results of developed ML algorithms

Hyperparameter tuning was conducted to optimize the performance of each ML model (Supplementary Table S4). The best-performing configurations revealed distinct characteristics across the models. The LightGBM model favored shallow trees with strong regularization (maximum tree depth = 1; L2 regularization parameter = 9.96), suggesting a preference for simplicity and generalization. Conversely, the RF model selected deeper trees with minimal constraints (maximum tree depth = 19; minimum number of samples required to split a node = 3; minimum number of samples required to form a leaf node = 1), indicating a tendency to capture more complex patterns. The SVM model was optimized with a high regularization parameter (C = 116) and a relatively wide ε-insensitive margin (ε = 0.83), reflecting a tolerance for prediction errors and a focus on smooth decision boundaries. Despite potential multicollinearity among clinical variables, tree-based models are inherently robust to this issue, and SVM performance was empirically stable during hyperparameter tuning; therefore, we retained all features to preserve clinical interpretability.

Regarding predictive performance, all three ML models showed comparable performance based on using R², RMSE, and 95% CIs; however, the SVM model achieved the highest R² score (0.67 [95% CI: 0.62–0.72]) and the lowest RMSE (11.29 [95% CI: 10.16–12.42]), indicating the best predictive accuracy (Table 2). Additionally, the MLR benchmark model showed an R² score and RMSE of 0.66 (95% CI: 0.61–0.71) and 11.50 (95% CI: 10.35–12.65) in the test dataset, respectively, similar to RF (R² = 0.66 [95% CI: 0.61–0.71], RMSE 11.50 [95% CI: 10.35–12.65]) and SVM models (R² = 0.67 [95% CI: 0.62–0.72], RMSE 11.29 [95% CI: 10.16–12.42]).

Table 2.

Performance and validity of predictive models generated using the three machine learning algorithms for predicting estimated glomerular filtration rate.

Methods		Training dataset		Test dataset
		R 2 (95% CI)	RMSE (95% CI)	R 2 (95% CI)	RMSE (95% CI)
Using all variables
LightGBM		0.80 (0.78–0.82)	8.26 (7.82–8.70)	0.65 (0.61–0.71)	11.72 (10.55–12.89)
Random Forest		0.89 (0.88–0.90)	6.21 (5.88–6.54)	0.66 (0.61–0.71)	11.50 (10.35–12.65)
Support Vector Machine		0.74 (0.71–0.77)	9.42 (8.92–9.92)	0.67 (0.62–0.72)	11.29 (10.16–12.42)
Using only the top-ranked variables
LightGBM		0.79 (0.76–0.82)	8.49 (8.05–8.93)	0.65 (0.59–0.71)	11.68 (10.59–12.77)
Random Forest		0.87 (0.85–0.89)	6.65 (6.31–6.99)	0.65 (0.58–0.72)	11.63 (10.57–12.69)
Support Vector Machine		0.75 (0.72–0.78)	9.24 (8.76–9.72)	0.68 (0.62–0.74)	11.25 (10.22–12.28)
Multiple Linear Regression		0.73 (0.70–0.76)	9.45 (8.95–9.95)	0.66 (0.61–0.71)	11.50 (10.35–12.65)

LightGBM, Light Gradient Boosting Machine; R², coefficient of determination; RMSE, root mean squared error; CI, confidence interval.

Figures 2a–c show the scatter plots of observed versus predicted eGFR values in the test dataset, which also serve as calibration plots. Each plot displays the calibration regression line, calibration slope, intercept (with 95% CI), and p-value for the joint hypothesis test of perfect calibration. For models using all features, calibration slopes were close to the ideal value of 1 (LightGBM: 0.98, RF: 1.04, SVM: 0.99), and the intercepts were close to the ideal value of 0 (LightGBM: 1.04, RF: -3.70, and SVM: 1.06). The joint hypothesis test for perfect calibration (intercept = 0 and slope = 1) was not rejected for any of the models (p > 0.05 for all), indicating good calibration across the eGFR range. Although all models showed good overall agreement, LightGBM and RF exhibited larger deviations at high eGFR values (≥ 100 mL/min/1.73 m²) (Figs. 2a–b), whereas the SVM model maintained closer agreement even at higher eGFR values (Fig. 2c).

Fig. 2.

(a–c) Model prediction results for Light Gradient Boosting Machine (LightGBM) (a), Random Forest (RF) (b), and Support Vector Machine (SVM) (c) using all features. (d–f) Model prediction results for LightGBM (d), RF (e), and SVM (f) using the top-ranked features. Predicted and observed estimated glomerular filtration rate (eGFR) values are plotted on the X and Y axes, respectively. These plots also serve as calibration plots. The blue line in each panel represents the linear regression of observed on predicted eGFR values. The calibration slope (β) and intercept (α), with 95% confidence intervals, as well as the p-value for the joint hypothesis test of perfect calibration (intercept = 0 and slope = 1), are also displayed. The black diagonal line represents perfect agreement between the predicted and observed eGFR values. Lighter coloring denotes later time-points.

Analysis of the features influencing the predictive performance of the ML models

Figure 3 shows the selected features ranked by descending SHAP values from the ML models using all variables. The number of selected features whose cumulative contribution exceeded 90% was 23, 14, and 29 for the LightGBM, RF, and SVM models, respectively. The top-ranked features common to all three ML models were eGFR, serum creatinine (sCr), ACR, age, high-density lipoprotein-cholesterol (HDL-C), hematocrit (Hct), white blood cell count (WBC), γ-GTP, and triglycerides (TGs) at baseline and the time-point variable. In the SHAP summary plots (Fig. 3, right panels), higher sCr, ACR, age, TG, and time-point values were associated with lower predicted future eGFR, whereas higher baseline eGFR, HDL-C, Hct, and γ-GTP were associated with higher predicted future eGFR.

Fig. 3.

Importance matrix plot of the Light Gradient Boosting Model (LightGBM) (a, left), Random Forest (RF) (b, left), and Support Vector Machine (SVM) (c, left) models. The features were selected in order of descending SHapley Additive exPlanations (SHAP) values from the machine learning models using all variables until the cumulative contribution exceeded 90%. SHAP summary plot of the clinical features of the LightGBM (a, right), RF (b, right), and SVM models (c, right). One dot represents a patient’s colored-feature value, where red indicates a higher value and blue indicates a lower value. A SHAP value greater than zero indicates an effect to increase the predicted eGFR. ACEi, angiotensin-converting enzyme inhibitor; ALT, alanine aminotransferase; ARB, angiotensin II receptor blocker; AST, aspartate aminotransferase; BG, biguanides; BMI, body mass index; BUN, blood urea nitrogen; CCB, calcium channel blockers; DPP4i, dipeptidyl peptidase 4 inhibitors; DR, diabetic retinopathy; FBG, fasting blood glucose; FHD, family history of diabetes; γ-GTP, γ-glutamyl-transpeptidase; Hb, hemoglobin; HbA1c, hemoglobin A1c; Hct, hematocrit; HDL-C, HDL-cholesterol; Plt, platelet count; sBP, systolic blood pressure; sCr, serum creatinine; TC, serum total cholesterol; TG, triglyceride; TZDs, thiazolidinediones; UA, uric acid; ACR, urine albumin-to-creatinine ratio; WBC, white blood cell count. Time-point, time from baseline.

For robustness assessments, reduced models were constructed using only the top-ranked features. As shown in Table 2, the reduced LightGBM and SVM models performed equivalently to or slightly better than their full-feature counterparts did, whereas the reduced RF model performed slightly worse (LightGBM: reduced R² = 0.65 [95% CI: 0.59–0.71], RMSE = 11.68 [95% CI: 10.59–12.77]; full R² = 0.65 [95% CI: 0.61–0.71], RMSE = 11.72 [95% CI: 10.55–12.89]; RF: reduced R² = 0.65 [95% CI: 0.58–0.72], RMSE = 11.63 [95% CI: 10.57–12.69]; full R² = 0.66 [95% CI: 0.61–0.71], RMSE = 11.50 [95% CI: 10.35–12.65]; SVM: reduced R² = 0.68 [95% CI: 0.62–0.74], RMSE = 11.25 [95% CI: 10.22–12.28]; full R² = 0.67 [95% CI: 0.62–0.72], RMSE = 11.29 [95% CI: 10.16–12.42]). The reduced SVM model continued to outperform the benchmark MLR model. Calibration analysis confirmed that these reduced models remained well-calibrated, with good agreement in the scatter plots and non-significant p-values for the joint hypothesis test of perfect calibration (p > 0.05 for all) (Figs. 2d–f). Notably, as with the full model, the reduced SVM model maintained better agreement at high eGFR values (≥ 100 mL/min/1.73 m²).

Prediction performance of the three ML models at different time points

A summary of the available data points at different time points is shown in Supplementary Table S5. For all three ML models using all features, R2 scores decreased and RMSE increased with longer follow-up in the test datasets (Table 3). Consistent with these results, the scatter plots showed larger deviations between the predicted and observed eGFR values at later time points (Fig. 2). The LightGBM model achieved R² scores ≥ 0.6 for up to 3 years from baseline in the test dataset, whereas the RF and the SVM models achieved R² ≥0.6 for up to 4 years from baseline (Supplementary Table S6A).

Table 3.

Performance and validity of predictive models generated using the three machine learning algorithms for predicting time-point estimated glomerular filtration rate.

Methods			Training dataset		Test dataset
		Time-point (year)	R2 (95% CI)	RMSE (95% CI)	R2 (95% CI)	RMSE (95% CI)
Using all variables
LightGBM		2	0.83(0.81–0.85)	7.58 (7.17–7.99)	0.67 (0.62–0.72)	12.52 (11.25–13.79)
		4	0.85 (0.83–0.87)	7.22 (6.71–7.73)	0.53 (0.46–0.60)	12.40 (10.72–14.08)
		6	0.81 (0.78–0.84)	7.86 (7.23–8.49)	0.41 (0.33–0.49)	12.66 (10.52–14.80)
RF		2	0.83 (0.81–0.85)	7.71 (7.30–8.12)	0.73 (0.68–0.78)	11.42 (10.27–12.57)
		4	0.81 (0.78–0.84)	8.06 (7.50–8.62)	0.60 (0.54–0.66)	11.53 (9.97–13.09)
		6	0.77 (0.73–0.81)	8.76 (8.06–9.46)	0.56 (0.49–0.63)	10.90 (9.05–12.75)
SVM		2	0.75 (0.72–0.78)	9.28 (8.78–9.78)	0.68 (0.63–0.73)	12.36 (11.11–13.61)
		4	0.70 (0.66–0.74)	10.11 (9.40–10.82)	0.61 (0.55–0.67)	11.27 (9.83–12.91)
		6	0.63 (0.58–0.68)	11.00 (10.12–11.88)	0.56 (0.49–0.63)	10.88 (9.04–12.72)
Using only the top-ranked variables
LightGBM		2	0.81 (0.79–0.83)	8.10 (7.67–8.53)	0.69 (0.63–0.75)	12.20 (11.05–13.35)
		4	0.82 (0.79–0.85)	7.85 (7.32–8.38)	0.52 (0.43–0.61)	12.53 (11.02–14.04)
		6	0.76 (0.72–0.80)	8.81 (8.13–9.49)	0.43 (0.32–0.54)	12.46 (10.69–14.23)
RF		2	0.81 (0.79–0.83)	7.98 (7.56–8.40)	0.73 (0.68–0.78)	11.44 (10.38–12.50)
		4	0.80 (0.77–0.83)	8.25 (7.70–8.80)	0.60 (0.51–0.69)	11.49 (10.15–12.83)
		6	0.76 (0.72–0.80)	8.96 (8.28–9.64)	0.55 (0.45–0.65)	11.00 (9.51–12.49)
SVM		2	0.77 (0.74–0.80)	8.85 (8.38–9.32)	0.73 (0.68–0.78)	11.42 (10.33–12.51)
		4	0.74 (0.70–0.78)	9.45 (8.81–10.09)	0.66 (0.59–0.73)	10.58 (9.28–11.88)
		6	0.71 (0.66–0.76)	9.82 (9.06–10.58)	0.62 (0.54–0.70)	10.13 (8.64–11.62)

LightGBM, Light Gradient Boosting Machine; RF, Random Forest; SVM, Support Vector Machine; R², coefficient of determination; RMSE, root mean squared error; CI, confidence interval.

In the reduced models, the R² scores and RMSE were almost comparable to those of the models using all features. For example, the R² scores for the reduced LightGBM and RF models remained ≥ 0.6 for up to 3 and 4 years from baseline in the test dataset, respectively, whereas the reduced SVM model slightly outperformed the full-feature model, maintaining moderate or better accuracy for up to approximately 6 years in the test dataset (R² = 0.60) (Table 3; Supplementary Table S6B).

The R² scores and RMSE of the MLR model showed comparable trends at each time point for the three ML models, with performance using all features in the test dataset declining over time: R² = 0.83 (95% CI: 0.81–0.85), RMSE = 8.6 (95% CI: 8.14–9.06) at 1 year; R² = 0.73 (95% CI: 0.70–0.76), RMSE = 11.35 (95% CI: 10.74–11.96) at 2 years; and R² = 0.69 (95% CI: 0.65–0.73), RMSE = 11.72 (95% CI: 11.01–12.43) at 3 years. Beyond 4 years, the R² values for MLR became negative, indicating that the MLR model at ≥ 4 years did not follow the test dataset trend (data not shown).

Subgroup performance and calibration

To further evaluate the robustness and clinical applicability of the developed models, we conducted additional subgroup performance and calibration assessments across clinically relevant strata. Subgroups were defined by baseline CKD stages (1–2 vs. 3–4), age (< 65 vs. ≥65 years), and sex (male vs. female). In the test dataset, the SVM model maintained relatively stable performance across all subgroups, with R² values ranging from 0.52 in older patients (≥ 65 years) to 0.72 in female patients, and calibration slopes close to 1.0 (Supplementary Table S7A; Supplementary Figs. S1–S6). Conversely, tree-based models (LightGBM and RF) showed greater variability, particularly in patients with advanced CKD (stages 3–4), where performance declined markedly (R² <0.50). Calibration analyses confirmed that all models remained well-calibrated within subgroups, as indicated by non-significant p-values for the joint hypothesis test of perfect calibration (p > 0.05 for all strata, Supplementary Figs. S1–S6).

SHAP-based interpretation revealed consistent top predictors (baseline eGFR, sCr, ACR, and time-point) across subgroups, with subgroup-specific patterns, such as greater contributions of HDL-C, and BMI in older patients (≥ 65 years) and stronger influence of blood urea nitrogen, and red blood cell count in females (Supplementary Figs. S7–S12). We also assessed the impact of feature reduction on subgroup performance. Reduced models using only the top-ranked SHAP features generally maintained comparable or slightly improved performance relative to full-feature models, particularly for SVM, except in patients with advanced CKD (stages 3–4) (Supplementary Table S7B).

Discussion

In this study, we developed long-term, ML-based models to predict future eGFR values at discrete annual time points in patients with T2DM for up to 9 years after baseline. We compared the performance of LightGBM, RF, and SVM models, with the SVM model demonstrating superior performance across a wide range of eGFR values; the difference in predictive performance was particularly pronounced at higher eGFR levels (≥ 100 mL/min/1.73 m²). The SVM model outperformed the benchmark MLR model, maintaining moderate or better predictive performance for up to approximately 6 years from baseline, whereas the MLR model had difficulty predicting future eGFR beyond 4 years from baseline. The superior performance of the SVM model likely reflects its ability to capture complex nonlinear relationships in the data that are not easily modeled by linear approaches, such as MLR. Additionally, the robustness of SVM in handling high-dimensional feature spaces likely contributes to its enhanced predictive accuracy over longer time horizons. These findings suggest that SVM-based models hold promise for improving long-term kidney function prediction in patients with T2DM.

This study has several strengths. First, the models were constructed using (1) readily available baseline information from routine medical check-ups for patients with T2DM and (2) long-term follow-up measurements collected across multiple centers. Recently, Gregorich et al. developed a multivariable mixed-effects model to predict future eGFR trajectories in patients with T2DM and CKD based on clinical characteristics¹⁵; however, to our knowledge, the present study is among the few to develop and evaluate ML-based models for forecasting future eGFR values at upcoming follow-up visits using baseline-only information. Although our study focused on predictive accuracy and calibration, we acknowledge that evaluating clinical utility using measures such as decision curve analysis and net benefit is essential for assessing real-world impact. These methods can help determine whether the model’s predictions lead to better clinical decisions and outcomes. We propose them as important directions for future research to assess the practical value of our model in clinical settings. In addition, unlike many DKD progression studies that included patients with prior cardiovascular or cerebrovascular disease^16,17, we analyzed a cohort without such histories at enrollment, thereby reducing potential confounding by major vascular events that may influence kidney function^18,19.

Although our primary aim was to develop accurate prediction models, we examined SHAP-derived feature importance to assess clinical plausibility. Across the three ML models, baseline eGFR, sCr, ACR, age, HDL-C, Hct, WBC, γ-GTP, TG, and the time-point variable were consistently identified as influential predictors of future eGFR. Among these features, age, sCr, and eGFR are well-established as eGFR-related features. A higher eGFR is a known protective factor, and a previous study suggested baseline eGFR should be included in the model, as it affects later eGFR measurements¹⁵. Other top-ranked features, including ACR^6,20 and WBC²¹, are established risk factors for DKD progression, whereas TG²², HDL-C²³, and Hct⁷ align with prior evidence linking dyslipidemia and anemia to renal decline. γ-GTP, although previously associated with cardiovascular and metabolic diseases, has not been directly linked to eGFR decline²⁴; however, it may influence CKD development through the progression of cardiovascular and metabolic diseases, a finding requiring further investigation.

Subgroup analyses revealed that the SVM model maintained a relatively stable performance across clinically relevant strata, supporting its potential for broad clinical applicability (Supplementary Table S7). Notably, SHAP-based interpretation confirmed the consistent importance of baseline eGFR, sCr, ACR, and time-point across all subgroups, while revealing subgroup-specific patterns: HDL-C, and BMI contributed more strongly in older patients (≥ 65 years), whereas blood urea nitrogen, and red blood cell count were more influential in females (Supplementary Figs. S7–S12). These findings highlight the need for tailored risk assessment strategies and suggest that incorporating subgroup-specific predictors could enhance model transparency and clinical utility. Local SHAP explanations and interaction analyses should also be incorporated in future studies to enhance model transparency. In contrast, prediction performance deteriorated in advanced CKD, which may reflect greater biological heterogeneity, more frequent treatment changes, and competing risks not captured by baseline-only predictors. These factors likely increase uncertainty in long-horizon forecasts and highlight the need for updated or longitudinal models for this subgroup. Future studies should investigate dynamic updating frameworks that incorporate follow-up measurements (e.g., at 1 and 2 years or at therapy initiation) to quantify individualized treatment-associated changes and improve predictions in advanced CKD.

Limitations

This study has some limitations. First, predicting long-term eGFR values is inherently more uncertain than near-term predictions, as patients’ clinical status, treatment regimens, and lifestyle factors may change significantly over time. Our models partially addressed this by including a time-point variable; however, they still relied solely on baseline data and assumed stability in unmeasured factors. Therefore, predictions at longer horizons should be interpreted cautiously, and external validation in diverse populations is required before clinical implementation. Although time-to-event models, such as mixed-effects or joint models, are well-suited for understanding the longitudinal trajectory of eGFR decline, our primary objective was to develop a practical early-warning tool based on data available at a single time point. This focus on prediction from the baseline, rather than on explaining the trajectory, justifies our use of high-performance ML algorithms, which are known for their predictive accuracy. Furthermore, we explicitly assessed the decay in model performance over time (Table 3), acknowledging that predictions for more distant time points are inherently more challenging. Mixed-effects or joint modeling approaches should be incorporated in future research to complement baseline-only predictions and explore trajectory-based risk assessments. Second, this study lacked temporal and external validation. Internal splits were used with a held-out test set, and regularization techniques were applied to reduce overfitting; however, these approaches cannot fully guarantee generalizability. Considering real-world developments, such as advances in diabetes care and variations in clinical practice across facilities, temporal validation strategies, such as training in the early years and validation in later years, should be incorporated. Furthermore, because our cohort consisted solely of Japanese participants with T2DM followed in an outpatient setting, external validation using an independent cohort that includes non-Japanese participants is warranted. Third, eGFR was calculated using the standard equation commonly used in Japan. Because eGFR equations differ across regions (e.g., CKD-EPI–based approaches), absolute eGFR values may not be directly interchangeable. Thus, international implementation would require external validation and, if necessary, recalibration or retraining using locally adopted eGFR equations. Fourth, the sample size was relatively small (n = 974). Although this dataset was sufficient for initial model development and internal validation, its limited size and potential lack of diversity may limit the generalizability of our findings.

Furthermore, this study began in 2010, with a median follow-up of 5.3 years. Certain medications approved for DKD treatment after study initiation (such as sodium-glucose cotransporter 2 inhibitors, glucagon-like peptide 1 receptor agonists, and non-steroidal mineralocorticoid receptor antagonists) were unavailable or not widely utilized during the study period. Although these medications may cause an acute decline in eGFR, they may improve long-term eGFR trajectories and renal outcomes in contemporary practice, potentially delaying progression to ESKD^25–27. Further research is needed to evaluate the predictive value of these medications for future eGFR changes. Finally, unknown factors may affect future eGFR in patients with T2DM. For example, Jiang et al. reported that including five genetic loci improved the prediction of eGFR trajectories in patients with T2DM⁶, whereas Mayer et al. showed that serum protein biomarkers enhanced prediction performance²⁸. We also found that urinary afamin levels were associated with DKD progression²⁹, although their utility in predicting future eGFR warrants further study. However, routine clinical practice does not include genetic variant testing, and the added value of protein biomarkers remains under investigation^30,31. Further studies are needed to validate novel risk factors affecting future eGFR in patients with T2DM and to reassess their utility in predictive models.

Clinical applicability and future directions

Regarding clinical applicability, our models achieved moderate predictive accuracy (including R² 0.68 [95% CI: 0.62–0.74] and RMSE 11.25 [95% CI: 10.22–12.28] for the reduced SVM model), which, while insufficient for precise point-predictions of eGFR values at 5 years, may be meaningful for long-term risk stratification. The model could be integrated into electronic health record systems to generate individualized forecasts of kidney function decline at initial visits, serving as a decision-support tool rather than replacing clinical judgment. For example, clinicians could (i) identify patients whose predicted eGFR at 3–5 years is substantially lower than expected for age and baseline kidney function and intensify monitoring (e.g., more frequent eGFR/ACR testing), and (ii) consider earlier nephrology referral when the predicted eGFR is projected to fall below clinically meaningful thresholds (e.g., < 60 or < 45 mL/min/1.73 m²), even if the current eGFR is stable. Additionally, because our approach provides predicted eGFR values at multiple annual follow-up time points, an annualized predicted change (i.e., a predicted “slope” over a specified horizon) can be derived from differences between predicted eGFR values across time points. Fixed “rapid decliner” thresholds (e.g., ≥ 5 mL/min/1.73 m²/year) remain clinically useful as a simple screening heuristic; however, patient-specific expected decline may complement this approach by contextualizing risk relative to an individual’s baseline profile and projected trajectory.

Before deployment, ethical and regulatory considerations should be addressed. Transparency is critical, and our SHAP analysis provides a foundation for model explainability, which is vital for clinical acceptance. Algorithmic bias is a primary concern; therefore, models must be validated across diverse populations to prevent the amplification of health disparities. From a regulatory perspective, the model would likely be classified as a clinical decision support tool, requiring robust external and temporal validation to demonstrate safety, fairness, and net clinical benefit—ideally through decision curve analysis—before integration into routine care.

In conclusion, our developed model adequately predicted future eGFR values for up to approximately 6 years from baseline in patients with T2DM. Integrating this tool into daily outpatient clinic workflows could support the management of CKD progression by enabling early interventions and individualized treatment decisions.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

H.U.: Conceptualization, Funding acquisition, Formal analysis, Project administration, Writing – original draft, Writing – review & editing. K.N.: Conceptualization, Formal analysis, Project administration. Y.S. and S.T.: Conceptualization, Formal analysis, Project administration, Writing – original draft. H.K., R.B., S.Y., Y.F., N.S., M.M., K.U., and M.N.: Investigation. Y.U.: Supervision. D.Y.: Writing – review & editing. Y.K.: Funding acquisition, Investigation, Supervision.

Funding

This work was supported by the Japan Health Research Promotion Bureau Research Fund for 6NC-EHRs [grant number 2022-E-03] (to H.U.); the NCGM Intramural Research Fund [grant number 24A1016] (to Y.K.); and the A&T corporation [grant number 21C027] (to H.U.). The funding organizations had no role in the conduct of the study, the analysis and interpretation of the data, or in manuscript preparation.

Data availability

The datasets that support the findings of this study are available from the corresponding author upon reasonable request.

Declarations

Conflict of interest

H.U. received a Japan Health Research Promotion Bureau Research Fund for 6NC-EHRs (2022-E-03), and a Grant from A&T corporation (21C027). Y.K. received the NCGM Intramural Research Fund (24A1016). K.N. is an employee of A&T Corporation and had a role in formal analysis and project administration. The rest of the authors have nothing to declare.