Ultrasound radiomics-based machine learning models for differentiating diabetic kidney disease from non-diabetic kidney disease in type 2 diabetes Authors: Wenye Chen, Dan Zhao, Lehao Wu, Wenkun Bai, Xintong Wang, Ying Ying Zhang, Jian-Ying Niu, Chen Yu Published: 2026 ABSTRACT Background and objectiveNoninvasive differentiation of diabetic kidney disease (DKD) from non-diabetic kidney disease (NDKD) in patients with type 2 diabetes mellitus (T2DM) remains challenging, as conventional clinical markers and ultrasound imaging lack sufficient accuracy and reliability. This limitation often leads to unnecessary renal biopsies or delayed diagnosis. To address this gap, this study aimed to develop and validate a machine learning model integrating renal ultrasound radiomic features and clinical predictors to distinguish DKD from NDKD in T2DM patients.MethodsPatients with T2DM who underwent renal biopsy were retrospectively enrolled from three centers. Based on biopsy findings, patients were classified as DKD or NDKD, with overlapping lesions assigned to DKD. Renal ultrasound images were collected for ROI delineation and radiomic feature extraction. Clinical predictors were selected using logistic regression, correlation analysis, and LASSO. SVM, KNN, RF, and XGBoost models were developed using radiomic features alone or combined with clinical variables. Performance was evaluated by AUC, accuracy, precision, recall, and F1-score, and SHAP was used for model interpretation.ResultsUnivariate analysis identified serum creatinine, eGFR, albumin and LDL‑C as significant variables, and eGFR was selected as the final clinical predictor after LASSO and multivariable logistic regression. The radiomics‑only model performed well in the training cohort but showed lower performance in the validation cohorts. The integrated model combining radiomic features with eGFR achieved AUCs of 0.991, 0.895 and 0.721 in the training, internal validation and external validation cohorts, respectively, with corresponding F1‑scores of 0.939, 0.815 and 0.714. DeLong’s test further showed that XGBoost performed better than several comparator algorithms in the training and internal validation cohorts, whereas no significant differences were observed among models in the external validation cohort.ConclusionThe XGBoost model integrating renal ultrasound radiomic features with eGFR showed favourable performance for differentiating DKD from NDKD in patients with T2DM. By using routinely available ultrasound images and a readily accessible clinical indicator, this model may serve as a non-invasive auxiliary tool for preliminary screening in primary-level hospitals or nephrology departments, providing supportive evidence for individualized assessment and renal biopsy decision-making.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12882-026-05353-7. Introduction Diabetic kidney involvement primarily manifests as diabetic kidney disease (DKD), which is one of the most common and serious chronic complications of diabetes mellitus and is the leading cause of chronic kidney disease (CKD) [1, 2]. Renal involvement in Type 2 Diabetes Mellitus (T2DM) is often presumed to be attributable to the (DKD; also known as diabetic nephropathy) [3]. However, a substantial proportion of diabetic patients may suffer from non-diabetic kidney disease (NDKD) or NDKD superimposed on DKD [3–9]. Reports from centres that routinely perform kidney biopsy (KB) in patients with T2DM have reported that the prevalence of NDKD ranges from 33% to 72.5% in T2DM with kidney disease [3]. Accurate classification of renal pathology in these patients is essential for predicting disease progression, guiding therapeutic decisions, and ultimately improving clinical outcomes. Nevertheless, a definitive diagnosis of renal involvement in these patients can be reliably distinguished only by kidney biopsy [3, 10, 11], which is a time-consuming, invasive, and costly procedure [12–14], making it unsuitable for routine use in all patients. Therefore, a non-invasive diagnostic method should be developed and evaluated to differentiate DKD from NDKD in patients with T2DM. Extensive efforts have been devoted to the identification and validation of alternative biomarkers, and the discovery of novel biomarkers remains promising [15]. However, to date, only a limited number of these biomarkers have been successfully translated into clinical practice, so far, because the specificity and sensitivity of most of these biomarkers lack rigorous external clinical validation [16, 17]. With advances in the rapid development of imaging technology, current analyses demonstrated that ΔADC, an imaging biomarker derived from DWI (based on MRI), may serve as a non-invasive method to distinguish DKD from NDKD in a biopsy-proven diabetic CKD cohort [18]. Nevertheless, the high cost limits its feasibility as a population-wide screening tool. Ultrasound is widely recognized as the primary imaging modality for evaluating kidneys [17]. Previous studies have successfully assessed the progression of DN using Doppler ultrasonography (US) [19], contrast-enhanced ultrasound [20], and shear wave elastography (SWE) [20, 21]. Despite these technical advances, the subtle sonographic differences among distinct diabetic renal pathologies often exceed the limits of visual interpretation. To overcome this limitation, recent studies have applied artificial intelligence (AI) to evaluate renal involvement in patients with T2DM [22]. However, many existing models either fail to achieve adequate diagnostic sensitivity in primary care settings or do not effectively integrate multi-modal data [23–26]. In this study, we focus on developing a robust diagnostic screening framework for patients with T2DM by implementing a binary classification model that integrates two-dimensional ultrasound imaging with key clinical variables. This approach is intended to provide a reliable, non-invasive tool to support early intervention and timely referral decisions in primary care settings. Methods Study design and patients The study cohort was assembled by screening patients with type 2 diabetes mellitus (T2DM) who were hospitalized for renal biopsy at the Department of Nephrology, Tongji Hospital, between January 2021 and November 2025, at the Department of Nephrology, Shanghai Fifth People’s Hospital, between January and December 2024; and at Shanghai Ninth People’s Hospital over the same period (January–December 2024). Only patients with clinical diagnoses of T2DM in their Electronic Health Record (EHR)were included. The flowchart of patient selection is shown in Fig. 1, and the modelling pipeline can be seen in Fig. 2. This study was conceived when the analysis of pragmatic data collected during routine clinical care was conducted following the KB approach. We did not influence the decision or indications to perform KB in these patients. The exclusion criteria were: (a) incomplete clinical data, including outpatients; (b) poor image quality; and(c) massive renal space-occupying lesions (e.g. renal cysts, renal tumours). In the study, which involved human participants, was reviewed and approved by Ethics Committee of Shanghai Tongji Hospital (approval notice number K-2022-022). The requirement for informed consent was waived by the Ethics Committee of Shanghai Tongji Hospital because this retrospective study only used fully anonymized clinical data without any intervention or interaction with patients. Fig. 1A flowchart of the participants A flowchart of the participants Fig. 2The modeling pipeline The modeling pipeline Fig. 3The receiver operating characteristic curves of support vector machine, K-nearest neighbours, random forest and XGBoost models are shown for the training cohort, internal validation cohort and external independent validation cohort. (A–C) ROC curves of standalone ultrasound radiomics models constructed using radiomic features alone. (D–F) ROC curves of integrated radiomics–clinical models constructed by combining ultrasound radiomic features with eGFR. The XGBoost-based integrated model showed favourable discriminatory performance across cohorts and was selected as the final diagnostic model The receiver operating characteristic curves of support vector machine, K-nearest neighbours, random forest and XGBoost models are shown for the training cohort, internal validation cohort and external independent validation cohort. (A–C) ROC curves of standalone ultrasound radiomics models constructed using radiomic features alone. (D–F) ROC curves of integrated radiomics–clinical models constructed by combining ultrasound radiomic features with eGFR. The XGBoost-based integrated model showed favourable discriminatory performance across cohorts and was selected as the final diagnostic model Clinical data collection Clinical data of all patients, including demographic data of these patients, including (a) Demographic characteristics: age, sex, height, weight, and body mass index (BMI); (b) Clinical and laboratory parameters: systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), glycated hemoglobin (HbA1c), cystatin C (CysC), blood urea nitrogen (BUN), serum creatinine (Scr), uric acid (UA), albumin (Alb), hemoglobin (Hb), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), urinary albumin/creatinine ratio (UACR), urinary albumin/creatinine ratio (UACR), 24-hour urinary total protein (24 h Upro), and 24 h urine volume; (c) Pathological data. The eGFR was calculated using the CKD-EPI (Chronic Kidney Disease Epidemiology Collaboration) equation [27]. All participants underwent percutaneous renal biopsy to obtain renal tissue specimens, which were subsequently examined by light microscopy, immunofluorescence and electron microscopy to differentiate DKD from NDKD. The distribution and proportions of NDKD pathological subtypes are presented in Supplementary Tables 1, and representative light-microscopic glomerular features of the NDKD, DKD + NDKD and DKD subgroups are shown in Supplementary Fig. 1. Data imputation and normalization Supplementary Fig. 2A visualizes missing data patterns. The blue cells represent observed data, while the red cells indicate missing values. The left margin displays the frequency of each missing data pattern, and the right margin shows the number of missing variables for each pattern. The bottom margin indicates the total count of missing entries for each variable. Missing data were assumed to be missing at random (MAR). It should be noted that the external validation cohort had complete data for all variables and therefore did not undergo any imputation; all imputation procedures were performed solely on the combined training and internal validation cohorts prior to dataset splitting. To minimize bias and maintain statistical power, we performed multiple imputation using the Multivariate Imputation by Chained Equations (MICE) algorithm with the mice package in R software version 4.4.2. Predictive Mean Matching (PMM) was applied to continuous variables to preserve the underlying distributional structure of the data. We generated five imputed datasets (m = 5) with 10 iterations per imputation (maxit = 10). To ensure the validity of the imputation, diagnostic checks were performed. We visually inspected the trace plots (Supplementary Fig. 2B) for variables with high missingness rates or substantial clinical relevance, including UPCR, UACR, 24-hour urinary total protein, 24-hour urine volume, FBG, and CysC. These plots demonstrated adequate mixing of the imputation chains without discernible trends, indicating satisfactory model convergence. In addition, density plots (Supplementary Fig. 2C) were examined to confirm that the distributions of the imputed values aligned with those of the observed data. The imputed dataset exhibiting the closest distributional match to the original observations was retained for subsequent analyses. Clinical feature analysis and predictor selection All statistical analyses were performed using Python software, version 3.9. Baseline demographic and clinical characteristics were summarized according to data distribution. Continuous variables with an approximately normal distribution are presented as the mean ± standard deviation (SD), whereas skewed variables are expressed as the median with interquartile range (IQR). Categorical variables are reported as counts and percentages [n (%)]. The Shapiro–Wilk test was used to assess normality. For comparisons between groups, normally distributed continuous variables were analysed using the independent-samples Student’s t-test or one-way analysis of variance, as appropriate. Non-normally distributed continuous variables were compared using the Mann–Whitney U test or Kruskal–Wallis H test. Categorical variables were compared using the chi-square test or Fisher’s exact test, where appropriate. To identify candidate predictors associated with renal involvement, univariate logistic regression analysis was first performed for baseline clinical variables. Variables with a P value < 0.05 in the univariate logistic regression analysis were retained for further feature screening. To reduce redundancy and potential multicollinearity, correlation analysis was subsequently conducted among the retained variables, highly correlated features were removed according to clinical relevance and statistical considerations. After correlation-based feature filtering, two complementary approaches were applied for predictor selection. First, the remaining variables were entered into a multivariable logistic regression model to identify factors independently associated with renal involvement. Second, least absolute shrinkage and selection operator regression with 10-fold cross-validation was performed to further select robust predictive features. Variables consistently identified by both multivariable logistic regression and LASSO regression were considered the final candidate predictors for subsequent model construction. All statistical tests were two-sided, and a P value < 0.05 was considered statistically significant. Kidney ultrasound scan method A 3–5 MHz or 3-10 MHz convex probe was used for the adult kidney scans. For the kidney, patients with left/right lateral or dorsal positions were scanned in the corneal plane, and B-mode images were included. The maximum longitudinal kidney images, including the renal sinus and parenchyma, were saved. The doctors were unaware of the patient’s clinical and biochemical indicators or pathological results. Image processing and analysis Image segmentation There is still controversy regarding whether to delineate the renal cortex, the renal medulla, or the entire kidney when defining the region of interest (ROI) in renal imaging. While some studies have investigated the whole kidney [28–30]. Growing evidence suggests that focusing on the renal cortex or representative cortical areas may provide more sensitive diagnostic information [31–35]. Given that the renal cortex is the primary site for key pathological processes, we opted to delineate the renal cortex as the ROI. Image segmentation was performed using the open-source 3D Slicer software version 5.6.2 (available at www.slicer). Radiomic feature extraction The open-source “PyRadiomics” package (Python version 3.9) was used to extract features from the original images. These features included first-order statistics, shape descriptors, Grey Level Co-occurrence Matrix (GLCM), Grey Level Run Length Matrix (GLRLM), Neighbouring Grey Tone Difference Matrix (NGTDM), Laplacian of Gaussian filter, and wavelet features. After loading the images and masks, they were converted into NumPy arrays, normalized, and standardized. All features were normalized using the z-score method to fall within a standardized value range. Feature selection and model development Feature stability The inter-observer agreement was evaluated using the interclass correlation coefficient (ICC) analysis. ICC values greater than 0.75 indicated good consistency of the extracted. Renal ultrasound images from 50 patients were randomly selected, and the ROI was separately delineated by Reader 1 and Reader 2. An inter-group consistency test was then performed. The radiomics variables with ICC > 0.75 were selected in this study (Supplementary Fig. 3A). Selection of radiomics features To mitigate overfitting arising from the high dimensionality of radiomic datasets, we implemented a stepwise feature dimensionality-reduction workflow. After excluding unstable and poorly reproducible radiomic features from the initially extracted high-throughput descriptors, the minimum redundancy maximum relevance algorithm was applied to rank the remaining candidate features according to their relevance and redundancy. The top 20 predictors with the greatest predictive potential were retained for subsequent screening. Subsequently, Least Absolute Shrinkage and Selection Operator regression with 10-fold cross-validation was performed to further refine the feature space and identify robust predictive variables, as illustrated by the LASSO coefficient path curve in Supplementary Fig. 3. As log λ increased along the x-axis of the LASSO path plot, most feature coefficients progressively contracted towards zero. Features that maintained non-zero coefficients at the optimal lambda threshold were ultimately retained. Together with the pre-specified clinical variable eGFR, seven predictive variables were incorporated into the final Model2_Rad_Clinic_XGBoost model. These comprised six radiomic features—wavelet-HLH_gldm_LargeDependenceHighGrayLevelEmphasis, wavelet-HHH_glcm_Imc1, wavelet-HLH_firstorder_MeanAbsoluteDeviation, original_glszm_SmallAreaEmphasis, original_shape_SurfaceVolumeRatio, and wavelet-HHL_firstorder_MeanAbsoluteDeviation—as well as the clinical indicator eGFR. Model construction To predict renal involvement, we developed and evaluated a series of predictive models based on four machine-learning algorithms: support vector machine, k-nearest neighbours, random forest, and extreme gradient boosting. For each algorithm, two distinct predictive signatures were constructed: Model 1 (Rad), derived exclusively from the selected radiomic features; and Model 2 (Clinic + Rad), a fusion model integrating clinical parameters with radiomic features. All models were implemented in Python version 3.9.18 using scikit-learn version 1.6.1, NumPy version 2.2.3, and associated computational libraries. Supplementary Table 2 summarizes the hyperparameter settings for each machine learning algorithm. Model evaluation and statistical performance The predictive performance of the machine learning models was evaluated using the Receiver Operating Characteristic (ROC) curve, with the Area Under the Curve (AUC) serving as the primary metric. To provide a comprehensive assessment, we further calculated the Accuracy, Precision, Recall (Sensitivity), and F1-score. Comparison of AUCs among different models was conducted using the DeLong test. A confusion matrix was constructed to visualize the diagnostic correspondence between the predicted classifications and the gold-standard pathological labels. To assess model reliability, calibration curves were generated to evaluate the agreement between the predicted probabilities and observed outcomes. Furthermore, Decision Curve Analysis (DCA) and Clinical Impact Curves (CIC) were employed to quantify the net benefit and determine the clinical utility of the models across a range of threshold probabilities, ensuring their applicability in primary care settings. Model interpretation To evaluate the influence of individual variables on the final model, we applied the Shapley Additive Explanations (SHAP) method. This technique facilitates a deeper understanding of feature contributions within the best-performing model when predicting outcomes on the test set. In this binary classification task, SHAP values were calculated to quantify the impact of each feature on the predicted probability of the target outcome. For each sample, a positive SHAP value indicates a feature that drives the model toward the positive class, while a negative value suggests a contribution toward the negative class. Results Clinical feature analysis revealed that eGFR was independently associated with renal pathology A total of 289 consecutive patients with biopsy-confirmed renal lesions and concomitant T2DM were recruited from two clinical centres: Shanghai Tongji Hospital and Shanghai Fifth People’s Hospital. The study population was divided into three independent cohorts: a primary training cohort (n = 155), an internal validation cohort (n = 64), and an external validation cohort (n = 70). Based on renal biopsy findings, which served as the reference standard, all enrolled patients were classified into either the NDKD or DKD group. Specifically, the training cohort comprised 94 patients with NDKD and 61 with DKD; the internal validation cohort included 39 patients with NDKD and 25 with DKD; and the external validation cohort consisted of 38 patients with T2DM-related NDKD and 32 with DKD. In the overall training cohort, 64 patients (41.3%) were male and 91 (58.7%) were female, with a median age of 60.00 years (IQR, 42.00–69.00 years). Baseline clinical characteristics of the NDKD and DKD groups in each cohort are summarized in Tables 1, 2 and 3. In the training cohort, significant between-group differences were observed in Cys-C, BUN, Scr, eGFR, Alb, and LDL-C (P < 0.05). In contrast, no significant differences were found in sex, age, blood pressure, FBG, HbA1c, uric acid, TC, TG, HDL-C, hemoglobin, UPCR, UACR, 24-h Upro, or 24 h urine volume between the two groups (P > 0.05). In the internal validation cohort, significant differences between the NDKD and DKD groups were identified in Cys-C, BUN, Scr, eGFR, UPCR, and UACR (P < 0.05), whereas the remaining baseline parameters were comparable between groups (all P > 0.05). In the external validation cohort, none of the collected baseline clinical or laboratory parameters differed significantly between the NDKD and DKD groups P > 0.05). Table 1The baseline clinical features in the training cohortCharacteristicTotal (n = 155)NDKD (n = 94)DKD (n = 61)p-valueGender(%) Male64 (41.3%)40 (42.6%)24 (39.3%)0.819 Female91 (58.7%)54 (57.4%)37 (60.7%)Age (year)60.00 (42.00, 69.00)54.00 (39.00, 71.00)63.00 (54.00, 66.00)0.376Height (cm)170.00(164.00, 173.00)170.00 (163.50, 173.00)170.00 (165.00, 173.00)0.451Weight (kg)69.00 (60.00, 80.00)68.00 (57.25, 81.50)70.00 (60.00, 80.00)0.643BMI (kg/m2)24.14 (22.22, 27.78)24.08 (22.03, 28.40)24.20 (22.84, 27.78)0.097SBP (mmHg)135.00 (130.00, 147.00)134.50 (130.00, 140.00)139.00 (130.00, 157.00)0.346DBP (mmHg)80.00 (77.00, 87.00)80.00 (76.25, 87.75)80.00 (80.00, 85.00)0.451FBG (mmol/L)7.87 (5.26, 10.00)8.22 (5.49, 9.86)7.04 (5.18, 10.00)0.856HbA1c (%)8.10 (6.90, 8.70)8.00 (6.90, 8.60)8.20 (7.00, 9.50)0.197Cys-C (mg/L)1.64 (1.07, 2.51)1.21 (0.95, 2.03)2.12 (1.62, 2.63) < 0.001*** BUN (mmol/L)8.62 (5.80, 12.45)7.12 (5.11, 11.45)9.50 (7.86, 13.04) < 0.001*** Scr (µmol/L121.60 (75.55, 179.05)98.20 (70.22, 150.45)154.10 (121.60, 263.60) < 0.001*** eGFRmL/min/1.73m250.60 (29.65, 89.92)64.20 (38.33, 97.40)40.80 (22.43, 55.90) < 0.001*** UA (µmol/L)405.55 ± 116.99397.65 ± 117.06417.73 ± 116.790.298Alb (g/L)28.50 (22.95, 35.15)25.55 (20.75, 34.48)31.40 (25.70, 36.90) 0.002** Hb (g/L)118.00 (96.00, 136.50)117.50 (99.00, 134.00)122.00 (92.00, 141.00)0.955TC (mmol/L)6.56 (5.10, 8.54)6.76 (5.04, 9.25)6.35 (5.58, 7.29)0.253TG (mmol/L)1.77 (1.32, 2.64)1.69 (1.31, 2.40)2.04 (1.43, 2.69)0.409HDL-C (mmol/L)1.49 (1.25, 1.72)1.53 (1.26, 1.71)1.41 (1.23, 1.72)0.258LDL-C (mmol/L)3.94 (2.90, 5.50)4.39 (3.17, 6.32)3.82 (2.75, 4.17) 0.003** UPCR (mg/g)3497.15 (1103.24, 7543.98)2685.35 (1222.33, 8186.65)4755.57 (1103.24, 7543.98)0.682UACR (mg/g)2108.91 (784.52, 4454.25)2047.98 (759.68, 4487.74)2184.10 (1028.12, 4285.53)0.88924 h Upro (g/d)4.03 (1.62, 8.25)4.06 (1.82, 8.37)3.55 (1.36, 8.25)0.56924 h Urine Volume (mL/d)1600.00 (1300.00, 2225.00)1600.00 (1300.00, 2250.00)1600.00 (1300.00, 2000.00)0.824Body mass index (BMI); systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), glycated hemoglobin (HbA1c), cystatin C (Cys-C), blood urea nitrogen (BUN), serum creatinine (Scr), estimated glomerular filtration rate (eGFR), uric acid (UA), albumin (Alb), hemoglobin (Hb), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), urinary albumin/creatinine ratio (UACR), 24-hour urinary total protein (24h Upro) Table 1The baseline clinical features in the training cohortCharacteristicTotal (n = 155)NDKD (n = 94)DKD (n = 61)p-valueGender(%) Male64 (41.3%)40 (42.6%)24 (39.3%)0.819 Female91 (58.7%)54 (57.4%)37 (60.7%)Age (year)60.00 (42.00, 69.00)54.00 (39.00, 71.00)63.00 (54.00, 66.00)0.376Height (cm)170.00(164.00, 173.00)170.00 (163.50, 173.00)170.00 (165.00, 173.00)0.451Weight (kg)69.00 (60.00, 80.00)68.00 (57.25, 81.50)70.00 (60.00, 80.00)0.643BMI (kg/m2)24.14 (22.22, 27.78)24.08 (22.03, 28.40)24.20 (22.84, 27.78)0.097SBP (mmHg)135.00 (130.00, 147.00)134.50 (130.00, 140.00)139.00 (130.00, 157.00)0.346DBP (mmHg)80.00 (77.00, 87.00)80.00 (76.25, 87.75)80.00 (80.00, 85.00)0.451FBG (mmol/L)7.87 (5.26, 10.00)8.22 (5.49, 9.86)7.04 (5.18, 10.00)0.856HbA1c (%)8.10 (6.90, 8.70)8.00 (6.90, 8.60)8.20 (7.00, 9.50)0.197Cys-C (mg/L)1.64 (1.07, 2.51)1.21 (0.95, 2.03)2.12 (1.62, 2.63) < 0.001*** BUN (mmol/L)8.62 (5.80, 12.45)7.12 (5.11, 11.45)9.50 (7.86, 13.04) < 0.001*** Scr (µmol/L121.60 (75.55, 179.05)98.20 (70.22, 150.45)154.10 (121.60, 263.60) < 0.001*** eGFRmL/min/1.73m250.60 (29.65, 89.92)64.20 (38.33, 97.40)40.80 (22.43, 55.90) < 0.001*** UA (µmol/L)405.55 ± 116.99397.65 ± 117.06417.73 ± 116.790.298Alb (g/L)28.50 (22.95, 35.15)25.55 (20.75, 34.48)31.40 (25.70, 36.90) 0.002** Hb (g/L)118.00 (96.00, 136.50)117.50 (99.00, 134.00)122.00 (92.00, 141.00)0.955TC (mmol/L)6.56 (5.10, 8.54)6.76 (5.04, 9.25)6.35 (5.58, 7.29)0.253TG (mmol/L)1.77 (1.32, 2.64)1.69 (1.31, 2.40)2.04 (1.43, 2.69)0.409HDL-C (mmol/L)1.49 (1.25, 1.72)1.53 (1.26, 1.71)1.41 (1.23, 1.72)0.258LDL-C (mmol/L)3.94 (2.90, 5.50)4.39 (3.17, 6.32)3.82 (2.75, 4.17) 0.003** UPCR (mg/g)3497.15 (1103.24, 7543.98)2685.35 (1222.33, 8186.65)4755.57 (1103.24, 7543.98)0.682UACR (mg/g)2108.91 (784.52, 4454.25)2047.98 (759.68, 4487.74)2184.10 (1028.12, 4285.53)0.88924 h Upro (g/d)4.03 (1.62, 8.25)4.06 (1.82, 8.37)3.55 (1.36, 8.25)0.56924 h Urine Volume (mL/d)1600.00 (1300.00, 2225.00)1600.00 (1300.00, 2250.00)1600.00 (1300.00, 2000.00)0.824Body mass index (BMI); systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), glycated hemoglobin (HbA1c), cystatin C (Cys-C), blood urea nitrogen (BUN), serum creatinine (Scr), estimated glomerular filtration rate (eGFR), uric acid (UA), albumin (Alb), hemoglobin (Hb), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), urinary albumin/creatinine ratio (UACR), 24-hour urinary total protein (24h Upro) The baseline clinical features in the training cohort eGFR mL/min/1.73m2 Body mass index (BMI); systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), glycated hemoglobin (HbA1c), cystatin C (Cys-C), blood urea nitrogen (BUN), serum creatinine (Scr), estimated glomerular filtration rate (eGFR), uric acid (UA), albumin (Alb), hemoglobin (Hb), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), urinary albumin/creatinine ratio (UACR), 24-hour urinary total protein (24h Upro) Table 2The baseline clinical features in the training cohorts Internal validation cohortCharacteristicTotal (n = 64)NDKD (n = 39)DKD (n = 25)p-valueGender(%) Male29 (45.3%)20 (51.3%)9 (36.0%)0.347 Female35 (54.7%)19 (48.7%)16 (64.0%)Age (year)56.50 (42.00, 68.00)51.00 (41.50, 67.50)63.00 (54.00, 70.00)0.182Height (cm)168.00 (162.75, 172.00)170.00 (161.00, 172.50)168.00 (165.00, 170.00)0.907Weight (kg)69.50 (60.00, 80.00)70.00 (57.50, 80.00)65.00 (60.00, 70.00)0.615BMI (kg/m2)24.18 (22.02, 27.45)24.91 (21.96, 27.80)23.39 (22.04, 27.34)0.587SBP (mmHg)138.00 (130.00, 149.50)135.00 (129.00, 141.00)140.00 (130.00, 169.00)0.092DBP (mmHg)80.00 (78.00, 91.25)80.00 (77.50, 84.50)81.00 (80.00, 101.00)0.238FBG (mmol/L)8.59 ± 3.878.77 ± 4.308.32 ± 3.140.648HbA1c (%)7.85 (6.90, 8.70)7.70 (6.90, 8.55)8.40 (7.30, 9.00)0.079Cys-C (mg/L)1.61 (1.04, 2.30)1.17 (0.95, 1.70)2.29 (1.72, 2.55) < 0.001*** BUN (mmol/L)9.07 (5.08, 14.21)6.18 (4.82, 11.11)13.80 (9.13, 15.44) 0.002** Scr (µmol/L117.00 (85.58, 160.50)101.60 (74.85, 125.30)155.90 (137.00, 344.70) < 0.001*** eGFRmL/min/1.73m255.85 (35.67, 90.22)64.70 (50.50, 93.90)38.40 (17.19, 46.96) < 0.001*** UA (µmol/L)375.50 (327.40, 443.45)364.00 (310.30, 419.30)404.70 (340.80, 472.10)0.160Alb (g/L)30.40 ± 9.9231.33 ± 10.5228.95 ± 8.920.354Hb (g/L)116.69 ± 24.45119.90 ± 22.51111.68 ± 26.900.192TC (mmol/L)5.97 (4.64, 7.97)5.33 (4.53, 8.53)6.46 (4.95, 7.58)0.762TG (mmol/L)1.91 (1.45, 2.61)1.77 (1.38, 2.43)2.18 (1.54, 2.69)0.189HDL-C (mmol/L)1.40 (1.14, 1.58)1.40 (1.17, 1.58)1.39 (1.12, 1.58)0.891LDL-C (mmol/L)3.88 (2.95, 5.38)4.15 (2.90, 5.72)3.49 (2.98, 5.21)0.726UPCR (mg/g)3606.30 (832.44, 8819.11)1975.28 (527.40, 6680.44)7437.12 (1821.06, 9760.35) 0.017* UACR (mg/g)2330.38 (598.92, 4630.05)1438.05 (386.68, 3919.23)3992.50 (1273.67, 5676.34) 0.020* 24 h Upro (g/d)4.21 (0.94, 10.43)3.41 (0.77, 8.71)7.02 (1.14, 13.45)0.09524 h Urine Volume (mL/d)1790.55 ± 810.801940.64 ± 874.961556.40 ± 647.930.064Body mass index (BMI); systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), glycated hemoglobin (HbA1c), cystatin C (Cys-C), blood urea nitrogen (BUN), serum creatinine (Scr), estimated glomerular filtration rate (eGFR), uric acid (UA), albumin (Alb), hemoglobin (Hb), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), urinary albumin/creatinine ratio (UACR), 24-hour urinary total protein (24h Upro) Table 2The baseline clinical features in the training cohorts Internal validation cohortCharacteristicTotal (n = 64)NDKD (n = 39)DKD (n = 25)p-valueGender(%) Male29 (45.3%)20 (51.3%)9 (36.0%)0.347 Female35 (54.7%)19 (48.7%)16 (64.0%)Age (year)56.50 (42.00, 68.00)51.00 (41.50, 67.50)63.00 (54.00, 70.00)0.182Height (cm)168.00 (162.75, 172.00)170.00 (161.00, 172.50)168.00 (165.00, 170.00)0.907Weight (kg)69.50 (60.00, 80.00)70.00 (57.50, 80.00)65.00 (60.00, 70.00)0.615BMI (kg/m2)24.18 (22.02, 27.45)24.91 (21.96, 27.80)23.39 (22.04, 27.34)0.587SBP (mmHg)138.00 (130.00, 149.50)135.00 (129.00, 141.00)140.00 (130.00, 169.00)0.092DBP (mmHg)80.00 (78.00, 91.25)80.00 (77.50, 84.50)81.00 (80.00, 101.00)0.238FBG (mmol/L)8.59 ± 3.878.77 ± 4.308.32 ± 3.140.648HbA1c (%)7.85 (6.90, 8.70)7.70 (6.90, 8.55)8.40 (7.30, 9.00)0.079Cys-C (mg/L)1.61 (1.04, 2.30)1.17 (0.95, 1.70)2.29 (1.72, 2.55) < 0.001*** BUN (mmol/L)9.07 (5.08, 14.21)6.18 (4.82, 11.11)13.80 (9.13, 15.44) 0.002** Scr (µmol/L117.00 (85.58, 160.50)101.60 (74.85, 125.30)155.90 (137.00, 344.70) < 0.001*** eGFRmL/min/1.73m255.85 (35.67, 90.22)64.70 (50.50, 93.90)38.40 (17.19, 46.96) < 0.001*** UA (µmol/L)375.50 (327.40, 443.45)364.00 (310.30, 419.30)404.70 (340.80, 472.10)0.160Alb (g/L)30.40 ± 9.9231.33 ± 10.5228.95 ± 8.920.354Hb (g/L)116.69 ± 24.45119.90 ± 22.51111.68 ± 26.900.192TC (mmol/L)5.97 (4.64, 7.97)5.33 (4.53, 8.53)6.46 (4.95, 7.58)0.762TG (mmol/L)1.91 (1.45, 2.61)1.77 (1.38, 2.43)2.18 (1.54, 2.69)0.189HDL-C (mmol/L)1.40 (1.14, 1.58)1.40 (1.17, 1.58)1.39 (1.12, 1.58)0.891LDL-C (mmol/L)3.88 (2.95, 5.38)4.15 (2.90, 5.72)3.49 (2.98, 5.21)0.726UPCR (mg/g)3606.30 (832.44, 8819.11)1975.28 (527.40, 6680.44)7437.12 (1821.06, 9760.35) 0.017* UACR (mg/g)2330.38 (598.92, 4630.05)1438.05 (386.68, 3919.23)3992.50 (1273.67, 5676.34) 0.020* 24 h Upro (g/d)4.21 (0.94, 10.43)3.41 (0.77, 8.71)7.02 (1.14, 13.45)0.09524 h Urine Volume (mL/d)1790.55 ± 810.801940.64 ± 874.961556.40 ± 647.930.064Body mass index (BMI); systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), glycated hemoglobin (HbA1c), cystatin C (Cys-C), blood urea nitrogen (BUN), serum creatinine (Scr), estimated glomerular filtration rate (eGFR), uric acid (UA), albumin (Alb), hemoglobin (Hb), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), urinary albumin/creatinine ratio (UACR), 24-hour urinary total protein (24h Upro) The baseline clinical features in the training cohorts Internal validation cohort eGFR mL/min/1.73m2 Body mass index (BMI); systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), glycated hemoglobin (HbA1c), cystatin C (Cys-C), blood urea nitrogen (BUN), serum creatinine (Scr), estimated glomerular filtration rate (eGFR), uric acid (UA), albumin (Alb), hemoglobin (Hb), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), urinary albumin/creatinine ratio (UACR), 24-hour urinary total protein (24h Upro) Table 3The baseline clinical features in the external validation cohortCharacteristicTotal (n = 70)T2DM (n = 38)DKD (n = 32)p-valueGender(%) Male25 (35.7%)12 (31.6%)13 (40.6%)0.431 Female45 (64.3%)26 (68.4%)19 (59.4%)Age (year)58.50 (42.25, 70.25)62.00 (45.50, 71.75)54.00 (41.75, 65.75)0.089Height (cm)169.30 (162.93, 173.00)169.50 (163.05, 173.57)169.30 (162.88, 172.10)0.575Weight (kg)72.15 (64.58, 79.17)71.55 (64.62, 76.45)73.55 (64.10, 85.25)0.209BMI (kg/m2)25.09 (22.81, 28.67)24.13 (22.62, 26.76)26.59 (23.25, 30.19)0.099SBP (mmHg)138.00 (129.25, 151.50)137.00 (129.00, 152.00)138.00 (130.75, 147.00)0.809DBP (mmHg)82.00 (76.00, 87.00)80.00 (70.00, 86.75)83.50 (79.50, 88.50)0.194FBG (mmol/L)7.00 (5.47, 8.52)6.59 (5.36, 7.80)8.13 (5.90, 8.99)0.060HbA1c (%)7.42 (6.12, 8.81)7.30 (6.12, 8.93)7.46 (6.29, 8.73)0.962Cys-C (mg/L)1.71 (0.97, 2.45)1.53 (0.95, 2.52)1.81 (1.08, 2.36)0.746BUN (mmol/L)9.13 (5.57, 13.75)9.48 (5.75, 13.75)8.86 (5.37, 13.47)0.556Scr (µmol/L133.00 (88.95, 209.30)141.85 (76.78, 265.65)123.80 (94.20, 184.40)0.596eGFRmL/min/1.73m246.72 (24.46, 89.28)45.77 (22.59, 92.60)49.71 (30.50, 80.82)0.451UA (µmol/L)393.40 (342.57, 511.93)376.17 (345.98, 452.82)417.10 (337.60, 523.25)0.175Alb (g/L)30.74 ± 8.0531.73 ± 8.0829.57 ± 7.990.266Hb (g/L)121.66 ± 28.07121.29 ± 27.24122.09 ± 29.460.906TC (mmol/L)6.25 (4.73, 7.77)6.40 (4.67, 7.52)6.18 (5.02, 8.10)0.878TG (mmol/L)2.12 (1.47, 2.85)2.05 (1.41, 2.59)2.18 (1.53, 3.44)0.253HDL-C (mmol/L)1.38 (1.10, 1.58)1.37 (1.13, 1.57)1.39 (0.91, 1.59)0.517LDL-C (mmol/L)3.87 (3.08, 5.17)3.75 (2.80, 4.80)3.88 (3.35, 5.49)0.189UPCR (mg/g)2277.96 (813.56, 6767.23)2089.43 (708.19, 6767.23)2375.03 (1128.59, 6118.86)0.724UACR (mg/g)1706.85 (669.99, 3910.74)1543.26 (526.96, 4166.99)1968.26 (834.21, 3740.76)0.57124 h Upro (g/d)3.48 (0.90, 8.59)3.48 (0.79, 7.81)3.69 (1.32, 8.73)0.61624 h Urine Volume (mL/d)1631.86 ± 581.601654.04 ± 560.271605.53 ± 613.930.731Body mass index (BMI); systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), glycated hemoglobin (HbA1c), cystatin C (Cys-C), blood urea nitrogen (BUN), serum creatinine (Scr), estimated glomerular filtration rate (eGFR), uric acid (UA), albumin (Alb), hemoglobin (Hb), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), urinary albumin/creatinine ratio (UACR), 24-hour urinary total protein (24h Upro) Table 3The baseline clinical features in the external validation cohortCharacteristicTotal (n = 70)T2DM (n = 38)DKD (n = 32)p-valueGender(%) Male25 (35.7%)12 (31.6%)13 (40.6%)0.431 Female45 (64.3%)26 (68.4%)19 (59.4%)Age (year)58.50 (42.25, 70.25)62.00 (45.50, 71.75)54.00 (41.75, 65.75)0.089Height (cm)169.30 (162.93, 173.00)169.50 (163.05, 173.57)169.30 (162.88, 172.10)0.575Weight (kg)72.15 (64.58, 79.17)71.55 (64.62, 76.45)73.55 (64.10, 85.25)0.209BMI (kg/m2)25.09 (22.81, 28.67)24.13 (22.62, 26.76)26.59 (23.25, 30.19)0.099SBP (mmHg)138.00 (129.25, 151.50)137.00 (129.00, 152.00)138.00 (130.75, 147.00)0.809DBP (mmHg)82.00 (76.00, 87.00)80.00 (70.00, 86.75)83.50 (79.50, 88.50)0.194FBG (mmol/L)7.00 (5.47, 8.52)6.59 (5.36, 7.80)8.13 (5.90, 8.99)0.060HbA1c (%)7.42 (6.12, 8.81)7.30 (6.12, 8.93)7.46 (6.29, 8.73)0.962Cys-C (mg/L)1.71 (0.97, 2.45)1.53 (0.95, 2.52)1.81 (1.08, 2.36)0.746BUN (mmol/L)9.13 (5.57, 13.75)9.48 (5.75, 13.75)8.86 (5.37, 13.47)0.556Scr (µmol/L133.00 (88.95, 209.30)141.85 (76.78, 265.65)123.80 (94.20, 184.40)0.596eGFRmL/min/1.73m246.72 (24.46, 89.28)45.77 (22.59, 92.60)49.71 (30.50, 80.82)0.451UA (µmol/L)393.40 (342.57, 511.93)376.17 (345.98, 452.82)417.10 (337.60, 523.25)0.175Alb (g/L)30.74 ± 8.0531.73 ± 8.0829.57 ± 7.990.266Hb (g/L)121.66 ± 28.07121.29 ± 27.24122.09 ± 29.460.906TC (mmol/L)6.25 (4.73, 7.77)6.40 (4.67, 7.52)6.18 (5.02, 8.10)0.878TG (mmol/L)2.12 (1.47, 2.85)2.05 (1.41, 2.59)2.18 (1.53, 3.44)0.253HDL-C (mmol/L)1.38 (1.10, 1.58)1.37 (1.13, 1.57)1.39 (0.91, 1.59)0.517LDL-C (mmol/L)3.87 (3.08, 5.17)3.75 (2.80, 4.80)3.88 (3.35, 5.49)0.189UPCR (mg/g)2277.96 (813.56, 6767.23)2089.43 (708.19, 6767.23)2375.03 (1128.59, 6118.86)0.724UACR (mg/g)1706.85 (669.99, 3910.74)1543.26 (526.96, 4166.99)1968.26 (834.21, 3740.76)0.57124 h Upro (g/d)3.48 (0.90, 8.59)3.48 (0.79, 7.81)3.69 (1.32, 8.73)0.61624 h Urine Volume (mL/d)1631.86 ± 581.601654.04 ± 560.271605.53 ± 613.930.731Body mass index (BMI); systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), glycated hemoglobin (HbA1c), cystatin C (Cys-C), blood urea nitrogen (BUN), serum creatinine (Scr), estimated glomerular filtration rate (eGFR), uric acid (UA), albumin (Alb), hemoglobin (Hb), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), urinary albumin/creatinine ratio (UACR), 24-hour urinary total protein (24h Upro) The baseline clinical features in the external validation cohort eGFR mL/min/1.73m2 Body mass index (BMI); systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), glycated hemoglobin (HbA1c), cystatin C (Cys-C), blood urea nitrogen (BUN), serum creatinine (Scr), estimated glomerular filtration rate (eGFR), uric acid (UA), albumin (Alb), hemoglobin (Hb), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), urinary albumin/creatinine ratio (UACR), 24-hour urinary total protein (24h Upro) As shown in Table 4. Univariate logistic regression identified four variables significantly associated with DKD diagnosis: Scr (OR = 1.00, 95% CI: 1.00–1.01, P = 0.004), eGFR (OR = 0.98, 95% CI: 0.97–0.99, P < 0.001), Alb (OR = 1.04, 95% CI: 1.00–1.08, P = 0.031), and LDL-C (OR = 0.79, 95% CI: 0.67–0.95, P = 0.010). Table 4Univariate and multivariate logistic regression analysis of clinical features in the training cohortCharacteristicUnivariate AnalysisMultivariate AnalysisOR(95%CI)p-valueOR(95%CI)p-valueGender(%) ( Male)1.14(0.59–2.20)0.69Age (year)1.02(0.99–1.04)0.11Height (cm)1.01(0.97–1.06)0.51Weight (kg)1.01(0.98–1.03)0.65BMI (kg/m2)1.01(0.93–1.10)0.76SBP (mmHg)1.02(1.00-1.04)0.06DBP (mmHg)1.01(0.98–1.04)0.43FBG (mmol/L)0.99(0.91–1.09)0.90HbA1c (%)1.22(0.96–1.54)0.10BUN (mmol/L)1.04(1.00-1.08)0.07Scr (µmol/L1.00(1.00-1.01)0.004**eGFRmL/min/1.73m20.98(0.97–0.99)< 0.001***0.98 (0.97, 0.99)< 0.001***UA (µmol/L)1.00(1.00–1.00)0.30Alb (g/L)1.04(1.00-1.08)0.031*1.03 (0.99, 1.07)0.16Hb (g/L)1.00(0.99–1.01)0.99TC (mmol/L)0.90(0.80–1.02)0.09TG (mmol/L)0.96(0.74–1.25)0.77HDL-C (mmol/L)0.56(0.26–1.20)0.14LDL-C (mmol/L)0.79(0.67–0.95)0.010**OR = 0.90 (0.74, 1.09)0.27UPCR (mg/g)1.00(1.00–1.00)0.93UACR (mg/g)1.00(1.00–1.00)0.5524 h Upro (g/d)1.00(0.93–1.07)0.9624 h Urine Volume (mL/d)1.00(1.00–1.00)0.96Body mass index (BMI); systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), glycated hemoglobin (HbA1c), cystatin C (Cys-C), blood urea nitrogen (BUN), serum creatinine (Scr), estimated glomerular filtration rate (eGFR), uric acid (UA), albumin (Alb), hemoglobin (Hb), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), urinary albumin/creatinine ratio (UACR), 24-hour urinary total protein (24h Upro) Table 4Univariate and multivariate logistic regression analysis of clinical features in the training cohortCharacteristicUnivariate AnalysisMultivariate AnalysisOR(95%CI)p-valueOR(95%CI)p-valueGender(%) ( Male)1.14(0.59–2.20)0.69Age (year)1.02(0.99–1.04)0.11Height (cm)1.01(0.97–1.06)0.51Weight (kg)1.01(0.98–1.03)0.65BMI (kg/m2)1.01(0.93–1.10)0.76SBP (mmHg)1.02(1.00-1.04)0.06DBP (mmHg)1.01(0.98–1.04)0.43FBG (mmol/L)0.99(0.91–1.09)0.90HbA1c (%)1.22(0.96–1.54)0.10BUN (mmol/L)1.04(1.00-1.08)0.07Scr (µmol/L1.00(1.00-1.01)0.004**eGFRmL/min/1.73m20.98(0.97–0.99)< 0.001***0.98 (0.97, 0.99)< 0.001***UA (µmol/L)1.00(1.00–1.00)0.30Alb (g/L)1.04(1.00-1.08)0.031*1.03 (0.99, 1.07)0.16Hb (g/L)1.00(0.99–1.01)0.99TC (mmol/L)0.90(0.80–1.02)0.09TG (mmol/L)0.96(0.74–1.25)0.77HDL-C (mmol/L)0.56(0.26–1.20)0.14LDL-C (mmol/L)0.79(0.67–0.95)0.010**OR = 0.90 (0.74, 1.09)0.27UPCR (mg/g)1.00(1.00–1.00)0.93UACR (mg/g)1.00(1.00–1.00)0.5524 h Upro (g/d)1.00(0.93–1.07)0.9624 h Urine Volume (mL/d)1.00(1.00–1.00)0.96Body mass index (BMI); systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), glycated hemoglobin (HbA1c), cystatin C (Cys-C), blood urea nitrogen (BUN), serum creatinine (Scr), estimated glomerular filtration rate (eGFR), uric acid (UA), albumin (Alb), hemoglobin (Hb), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), urinary albumin/creatinine ratio (UACR), 24-hour urinary total protein (24h Upro) Univariate and multivariate logistic regression analysis of clinical features in the training cohort eGFR mL/min/1.73m2 Body mass index (BMI); systolic blood pressure (SBP), diastolic blood pressure (DBP), fasting blood glucose (FBG), glycated hemoglobin (HbA1c), cystatin C (Cys-C), blood urea nitrogen (BUN), serum creatinine (Scr), estimated glomerular filtration rate (eGFR), uric acid (UA), albumin (Alb), hemoglobin (Hb), total cholesterol (TC), triglyceride (TG), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), urinary albumin/creatinine ratio (UACR), 24-hour urinary total protein (24h Upro) Given the strong collinearity between eGFR and Scr (ρ = −0.96), Scr was excluded to avoid redundant renal function information (Supplementary Fig. 4A). CysC was excluded from the primary analysis because of a missing rate greater than 20%. In a sensitivity analysis including patients with available CysC data, CysC was added as a candidate variable but was not retained after repeated feature selection, while eGFR remained selected (Supplementary Fig. 4C and Supplementary Table 3). The remaining variables, eGFR, LDL-C, and Alb, were subsequently entered into LASSO logistic regression. Using 10-fold cross-validation, lambda.min was selected as the optimal penalty parameter (λ = 0.029), under which all three variables retained non-zero coefficients: eGFR, LDL-C, and Alb. To identify the most robust predictor, LASSO-selected variables were further intersected with those showing independent significance in multivariable logistic regression (Supplementary Fig. 4B). Among the three candidate variables, only eGFR remained statistically significant in the final model. Therefore, eGFR was ultimately retained as the final robust predictor for DKD diagnosis. Radiomic features were extracted and selected, after which machine learning models were developed and validated In total, 289 patients were retrospectively enrolled, and 289 renal US images were collected from these patients. We extracted 851 radiomic features from the B-mode images for each participant. The radiomics features showed good consistency (ICC > 0.75), and after filtering, we retained 786 features. Table 5 and Fig. 3 summarize the diagnostic performance of the standalone radiomics models (Model1_Rad) and the integrated radiomics–clinical models (Model2_Rad_Clinic, in which eGFR was incorporated into the radiomics feature set) across the training, internal validation and external independent validation cohorts. Among the standalone radiomics models, Model1_Rad_XGBoost achieved near-perfect performance in the training cohort, with an AUC of 0.999, an accuracy of 0.981 and an F1-score of 0.980. However, its performance decreased in the internal validation cohort and declined further in the external validation cohort, with AUCs of 0.835 and 0.669, respectively, suggesting substantial overfitting and limited generalizability. In contrast, Model2_Rad_Clinic_XGBoost, which integrated radiomics features with the single clinical predictor eGFR, demonstrated more stable and favorable discriminatory performance across cohorts. The model achieved AUCs of 0.991, 0.895 and 0.721 in the training, internal validation and external validation cohorts, respectively, with corresponding F1-scores of 0.939, 0.815 and 0.714. Table 5Performance of the different models in the training and validation cohortsDatasetModelAUCAccuracyPrecisionRecallF1-scoreTrainModel1_Rad_SVM0.9010.8320.8320.8130.820Model1_Rad_KNN0.9250.8390.8350.8240.828Model1_Rad_RF0.9980.9610.9660.9540.959 Model1_Rad_ XGBoost 0.999 0.981 0.981 0.978 0.980 Model2_Rad_Clinic_SVM0.9170.8580.8530.8480.850Model2_Rad_Clinic_KNN1.0001.0001.0001.0001.000Model2_Rad_Clinic_Rad_RF0.9960.9550.9540.9510.953Model2_Rad_Clinic_XGBoost0.9910.9420.9430.9350.939Interna validationModel1_Rad_SVM0.7880.7340.7310.6960.702Model1_Rad_KNN0.8410.7660.7610.7360.743Model1_Rad_RF0.8430.7660.7560.7430.748 Model1_Rad_ XGBoost 0.835 0.766 0.754 0.750 0.752 Model2_Rad_Clinic_SVM0.8320.8130.8120.7890.796Model2_Rad_Clinic_KNN0.8800.7970.7910.7760.781Model2_Rad_Clinic_Rad_RF0.8850.8280.8340.8020.812 Model2_Rad_Clinic_XGBoost 0.895 0.828 0.826 0.809 0.815 External validationModel1_Rad_SVM0.7440.6710.6910.6830.670Model1_Rad_KNN0.6900.6570.6800.6690.655Model1_Rad_RF0.6830.5570.6390.5850.522 Model1_Rad_ XGBoost 0.669 0.714 0.725 0.722 0.714 Model2_Rad_Clinic_SVM0.7420.6570.6800.6690.655Model2_Rad_Clinic_KNN0.6640.6860.6960.6930.685Model2_Rad_Clinic_Rad_RF 0.656 0.5140.5920.5450.461 Model2_Rad_Clinic_XGBoost 0.721 0.714 0.720 0.720 0.714 Table 5Performance of the different models in the training and validation cohortsDatasetModelAUCAccuracyPrecisionRecallF1-scoreTrainModel1_Rad_SVM0.9010.8320.8320.8130.820Model1_Rad_KNN0.9250.8390.8350.8240.828Model1_Rad_RF0.9980.9610.9660.9540.959 Model1_Rad_ XGBoost 0.999 0.981 0.981 0.978 0.980 Model2_Rad_Clinic_SVM0.9170.8580.8530.8480.850Model2_Rad_Clinic_KNN1.0001.0001.0001.0001.000Model2_Rad_Clinic_Rad_RF0.9960.9550.9540.9510.953Model2_Rad_Clinic_XGBoost0.9910.9420.9430.9350.939Interna validationModel1_Rad_SVM0.7880.7340.7310.6960.702Model1_Rad_KNN0.8410.7660.7610.7360.743Model1_Rad_RF0.8430.7660.7560.7430.748 Model1_Rad_ XGBoost 0.835 0.766 0.754 0.750 0.752 Model2_Rad_Clinic_SVM0.8320.8130.8120.7890.796Model2_Rad_Clinic_KNN0.8800.7970.7910.7760.781Model2_Rad_Clinic_Rad_RF0.8850.8280.8340.8020.812 Model2_Rad_Clinic_XGBoost 0.895 0.828 0.826 0.809 0.815 External validationModel1_Rad_SVM0.7440.6710.6910.6830.670Model1_Rad_KNN0.6900.6570.6800.6690.655Model1_Rad_RF0.6830.5570.6390.5850.522 Model1_Rad_ XGBoost 0.669 0.714 0.725 0.722 0.714 Model2_Rad_Clinic_SVM0.7420.6570.6800.6690.655Model2_Rad_Clinic_KNN0.6640.6860.6960.6930.685Model2_Rad_Clinic_Rad_RF 0.656 0.5140.5920.5450.461 Model2_Rad_Clinic_XGBoost 0.721 0.714 0.720 0.720 0.714 Performance of the different models in the training and validation cohorts Pairwise AUC comparisons among the four Model2_Rad_Clinic algorithms, including SVM, KNN, RF and XGBoost, were performed using DeLong’s test, as shown in Supplementary Fig. 5. In the training cohort, XGBoost achieved significantly higher AUCs than SVM, KNN and RF (all P < 0.05). In the internal validation cohort, XGBoost also outperformed SVM (P = 0.02) and KNN (P = 0.01). In the external validation cohort, however, no statistically significant differences in AUC were observed between any pair of models (all P > 0.05). The confusion matrices of Model2_Rad_Clinic_XGBoost across the three datasets are presented in Fig. 4. The model maintained relatively low numbers of false-positive and false-negative classifications in the training, internal validation and external validation cohorts, further supporting its ability to distinguish patients with DKD from those with NDKD. Fig. 4Confusion matrix heatmaps of the optimal XGBoost classifiers corresponding to the datasets shown in Fig. 4. (A–C) Confusion matrices of the standalone ultrasound radiomics model in the training cohort, internal validation cohort and external validation cohort, respectively. (D–F) Confusion matrices of the integrated radiomics–clinical model in the corresponding cohorts. The true labels indicate pathological classification, with 0 representing NDKD and 1 representing DKD, whereas the predicted labels indicate the model-assigned category. The number in each cell represents the case count, and the colour gradient reflects the sample size within each classification category Confusion matrix heatmaps of the optimal XGBoost classifiers corresponding to the datasets shown in Fig. 4. (A–C) Confusion matrices of the standalone ultrasound radiomics model in the training cohort, internal validation cohort and external validation cohort, respectively. (D–F) Confusion matrices of the integrated radiomics–clinical model in the corresponding cohorts. The true labels indicate pathological classification, with 0 representing NDKD and 1 representing DKD, whereas the predicted labels indicate the model-assigned category. The number in each cell represents the case count, and the colour gradient reflects the sample size within each classification category On the basis of its cross-cohort stability, quantitative diagnostic performance and statistical advantage over comparator algorithms, Model2_Rad_Clinic_XGBoost, integrating radiomics features with eGFR, was selected as the final diagnostic model for differentiating DKD from NDKD. Calibration curves, decision curve analysis, and clinical impact curves demonstrated that Model2_Rad_Clinic achieved the best calibration and clinical utility Clinical impact curves further demonstrated the risk-stratification performance of the models (Supplementary Fig. 6A–F). For Model2_Rad_Clinic_XGBoost, the estimated number of patients classified as high risk closely paralleled the number of DKD cases across all three cohorts, indicating that the integrated model could reliably identify patients with DKD when assigning high-risk labels. In comparison, Model1_Rad_XGBoost showed larger discrepancies between the predicted high-risk population and the actual number of DKD-positive cases, especially in the external validation cohort. Clinical impact curves (Supplementary Fig. 6A–F), decision curve analysis (DCA) (Supplementary Fig. 7A–F), Calibration curves (Supplementary Fig. 8A–F), and were further used to evaluate the calibration consistency and potential clinical utility of the final optimal model, Model2_Rad_Clinic_XGBoost, which integrated radiomics features with the single clinical predictor eGFR, across the training, internal validation and external independent validation cohorts. For calibration assessment, Model1_Rad_XGBoost is shown in Supplementary Fig. 8A–C, whereas Model2_Rad_Clinic_XGBoost is shown in Supplementary Fig. 8D–F. In the training and internal validation cohorts, the observed-probability curve of Model2_Rad_Clinic_XGBoost closely aligned with the ideal diagonal reference line, with narrow 95% confidence intervals, indicating excellent agreement between the predicted probability of DKD and the observed event rate. Although a modest deviation was observed in the external validation cohort, Model2_Rad_Clinic_XGBoost still exhibited substantially better calibration than the standalone radiomics model, Model1_Rad_XGBoost, which showed more pronounced departures from the ideal diagonal, particularly in the external validation cohort. Decision curve analysis quantified the clinical net benefit of the two candidate models across a broad range of threshold probabilities and corresponding cost–benefit ratios (Supplementary Fig. 7A–F). Compared with Model1_Rad_XGBoost and the two reference strategies of treating all or treating none, Model2_Rad_Clinic_XGBoost consistently yielded higher net benefit across most clinically relevant threshold ranges in the training, internal validation and external validation cohorts. By contrast, the net benefit of Model1_Rad_XGBoost declined rapidly and fell below the treat-all reference line at relatively low threshold probabilities, suggesting limited clinical applicability for the radiomics-only signature. Collectively, calibration analysis, DCA and clinical impact curves consistently demonstrated that Model2_Rad_Clinic_XGBoost, integrating radiomics features with eGFR, achieved favorable calibration, superior clinical net benefit and reliable risk-stratification performance for distinguishing DKD from NDKD. SHAP analysis revealed the global feature importance and directional impact of key predictors in Model2_Rad_Clinic_XGBoost, while force plots demonstrated its potential for individualized, transparent risk assessment To elucidate feature contributions and enhance the clinical interpretability of the final optimal model, Model2_Rad_Clinic_XGBoost, which integrated radiomics features with the single clinical predictor eGFR, SHapley Additive exPlanations (SHAP) analysis was performed (Fig. 5). This approach decomposed the predictions of the XGBoost algorithm and quantified the marginal contribution of each feature to DKD classification. Global feature importance, ranked by the mean absolute SHAP value (Fig. 5A), identified wavelet-HLH_gldm_LargeDependenceHighGrayLevelEmphasis as the most influential predictor, followed by wavelet-HHH_glcm_Imc1, wavelet-HLH_firstorder_MeanAbsoluteDeviation, eGFR, original_glszm_SmallAreaEmphasis, original_shape_SurfaceVolumeRatio and wavelet-HHL_firstorder_MeanAbsoluteDeviation. This hierarchy indicates that the final integrated model captured complementary discriminatory information from higher-order wavelet-derived radiomics signatures and the clinically important renal function marker eGFR. The SHAP summary beeswarm plot further illustrated the distribution and directional effects of individual predictors on the predicted probability of DKD (Fig. 5B). For the key clinical variable eGFR, lower values were associated with positive SHAP values, thereby increasing the predicted risk of DKD. Conversely, higher values of the top-ranked radiomics feature, wavelet-HLH_gldm_LargeDependenceHighGrayLevelEmphasis, also contributed positively to the model output and increased the predicted probability of DKD. The remaining radiomics features showed coherent directional associations between feature magnitude and SHAP output across the study population. To demonstrate patient-level interpretability, SHAP waterfall and force plots were generated for a representative individual case (Fig. 5C, D). The expected baseline prediction value for the overall cohort was 0.403, whereas the final predicted probability of DKD for this patient was 0.302. Among the contributing predictors, original_glszm_SmallAreaEmphasis (+ 0.071) and eGFR (+ 0.0498) shifted the prediction towards a higher DKD probability. In contrast, wavelet-HLH_gldm_LargeDependenceHighGrayLevelEmphasis (− 0.0858), wavelet-HHH_glcm_Imc1 (− 0.0685), wavelet-HLH_firstorder_MeanAbsoluteDeviation (− 0.0601) and two additional radiomics features exerted negative contributions, collectively reducing the predicted probability below the cohort baseline. Collectively, the SHAP visualizations provided both global and individualized explanations of the decision-making process of the integrated XGBoost model. These analyses mitigated the opacity of the machine-learning framework and supported transparent, patient-specific interpretation for the differential diagnosis of DKD and NDKD in clinical practice. Fig. 5SHAP-based interpretation of the final multimodal radiomics–clinical XGBoost model for differentiating DKD from NDKD. (A) SHAP summary bar plot showing the mean absolute SHAP value of each incorporated predictive feature, reflecting its overall contribution to model prediction. Wavelet-transformed ultrasound radiomic features accounted for the three highest predictive weights, followed by the clinical indicator eGFR and original shape- and texture-based radiomic features. (B) SHAP dependence scatter plot showing the relationship between individual feature values and their corresponding SHAP values. The colour gradient denotes the raw feature magnitude, ranging from low values in purple to high values in orange, illustrating how changes in feature values shift the model output towards or away from a DKD prediction. (C) SHAP waterfall plot for a representative patient. Starting from the baseline expected model output, E[f(x)] = 0.403, features with positive SHAP values increased the predicted probability of DKD, whereas features with negative SHAP values decreased the predicted probability. The final predicted probability for this case was 0.302. (D) SHAP decision plot summarizing the cumulative marginal contribution of ordered features across the cohort, further demonstrating the dominant predictive roles of high-level texture radiomic features and eGFR in the final risk estimation SHAP-based interpretation of the final multimodal radiomics–clinical XGBoost model for differentiating DKD from NDKD. (A) SHAP summary bar plot showing the mean absolute SHAP value of each incorporated predictive feature, reflecting its overall contribution to model prediction. Wavelet-transformed ultrasound radiomic features accounted for the three highest predictive weights, followed by the clinical indicator eGFR and original shape- and texture-based radiomic features. (B) SHAP dependence scatter plot showing the relationship between individual feature values and their corresponding SHAP values. The colour gradient denotes the raw feature magnitude, ranging from low values in purple to high values in orange, illustrating how changes in feature values shift the model output towards or away from a DKD prediction. (C) SHAP waterfall plot for a representative patient. Starting from the baseline expected model output, E[f(x)] = 0.403, features with positive SHAP values increased the predicted probability of DKD, whereas features with negative SHAP values decreased the predicted probability. The final predicted probability for this case was 0.302. (D) SHAP decision plot summarizing the cumulative marginal contribution of ordered features across the cohort, further demonstrating the dominant predictive roles of high-level texture radiomic features and eGFR in the final risk estimation Discussion In this study, we developed and validated a machine learning model that integrates ultrasound radiomics features with clinical data for the non-invasive identification of DKD in patients with T2DM. The model was deliberately designed as a binary classification framework to distinguish NDKD from DKD. This design reflects real-world clinical practice, in which NDKD and mixed pathological forms are common among patients with CKD and diabetes, with reported prevalences ranging from 27% to 82.9% and 4% to 45.5%, respectively [11]. In patients with T2DM and renal involvement, the most clinically relevant question is the presence or absence of NDKD, as this directly affects the diagnostic evaluation, treatment strategies, and decisions regarding renal biopsy. Although mixed pathological forms with both DKD and NDKD are common, they are usually managed according to DN dominant pathological features in daily clinical practice, particularly in the presence of advanced diabetic lesions [36–40]. By integrating ultrasound radiomics with clinical indexes, our model offers a non-invasive and straightforward approach to identifying patients with T2DM who are more likely to develop NDKD. This approach may facilitate earlier diagnostic assessment, reduce delays in initiating appropriate management, and ultimately support more individualized patient care. The SHAP analysis highlights the primary contribution of wavelet-transformed features, which provide a multi-scale decomposition of renal parenchymal textures. Among these, wavelet-HLH_gldm_LargeDependenceHighGrayLevelEmphasis (LDHGLE) is the most significant feature in the global importance ranking. In radiomics, LDHGLE quantifies the joint occurrence of large connected components with high grey-level intensities. For example, a machine‑learning radiomics study in breast cancer reported that LDHGLE was significantly associated with locoregional recurrence and poor treatment response [41]. The second most contributive feature, wavelet‑HHH_glcm_Imc1, showed a negative correlation between higher values and DKD, indicating reduced texture complexity and less correlated gray‑level distributions [42]. Wavelet-HLH_firstorder_MeanAbsoluteDeviation (MAD) and wavelet-HHL_firstorder_MeanAbsoluteDeviation characterize intensity dispersion across different wavelet subbands. In the present model, lower HLH-MAD, lower original_glszm_SmallAreaEmphasis, and lower original_shape_SurfaceVolumeRatio were associated with an increased likelihood of DKD, whereas higher HHL-MAD showed a modest positive contribution to DKD prediction. These findings suggest that features reflecting intensity variability, micro-textural patterns, and global morphological characteristics jointly contribute to the discrimination between DKD and NDKD. In previous studies, Kong et al. reported significant historadiological correlations between CT‑based radiomic features and chronic histologic lesions in kidney biopsies, with top‑ranked individual radiomic features achieving an AUC of 0.89 for differentiating moderate‑to‑severe chronicity [43].Hassan et al. also found that Busyness was negatively correlated with glomerulosclerosis, Coarseness was positively associated with interstitial inflammation, and the radiomics feature Energy was significantly correlated with eGFR [44]. In a study of 219 patients, machine learning combining ultrasound and biochemical data effectively diagnosed early DKD, attributed to disease‑specific texture features that capture pathological alterations, supporting a mechanistic link between radiomics and renal histopathology [28]. To enhance the clinical applicability and translational relevance of the model, key clinical variables were incorporated into the analysis. Our clinical biomarker findings are consistent with previous studies identifying 24h_Upro, age, systolic blood pressure, and serum creatinine as pivotal indicators in the assessment of kidney disease [9, 37, 45–58]. In the univariate analysis, renal function-related variables, including Scr and eGFR, were significantly associated with the differentiation of DKD from NDKD. However, Spearman correlation analysis revealed substantial multicollinearity among these markers, particularly between eGFR and Scr. To ensure the model remains practical for primary healthcare settings, eGFR calculated via the CKD-EPI equation was prioritized for the multivariate analysis. Ultimately, eGFR was identified as independent clinical predictors in our final models. SHAP analysis further confirmed that eGFR significantly contributed to the diagnostic performance of Model2_Rad_Clinic_XGBOOST, particularly when combined with wavelet-based radiomics features. This integration enabled the model to achieve high diagnostic accuracy while retaining practical utility for population screening. In comparison, diabetic retinopathy (DR) has been recognised as a crucial predictor of DKD. However, DR screening is not well performed in rural areas of China due to the unaffordability of medical costs, a lack of medical facilities, and limited access to conventional screening programs [59]. Moreover, diabetic retinopathy is observed in only 60 to 65% of cases with DKD, making its reliability as a standalone diagnostic indicator [60]. Beyond predictive accuracy, the clinical utility of Model2_Rad_Clinic_XGBOOST was validated through Decision Curve Analysis (DCA), which showed a superior net benefit across a wide range of threshold probabilities compared to clinical-only or radiomics-only models. Calibration curves demonstrated excellent agreement between predicted and actual outcomes, indicating high model reliability. Furthermore, the clinical impact curve (CIC) confirmed that the model’s high-risk predictions closely aligned with actual positive cases. By integrating eGFR with advanced radiomics, we have developed a robust tool that provides accurate, individualized risk assessment while remaining feasible for broad clinical use. Comparison with existing models. Several non-invasive models for differentiating DKD from NDKD have been reported. A summary of these published predictive models is provided in Supplementary Table 4. For example, Zhang et al. developed a clinical risk score incorporating diabetes duration, eGFR, proteinuria, hematuria, and diabetic retinopathy (DR), achieving an AUC of 0.86–0.92 [61]. In the ultrasound radiomics domain, Liu et al. constructed a clinical-radiomics nomogram with an AUC of 0.881 [62]. Our Model2_Rad_Clinic_XGBoost integrated wavelet‑based radiomic features with eGFR, achieving AUCs of 0.991, 0.895, and 0.721 in training, internal validation, and external validation cohorts, respectively. In the external cohort, eGFR did not differ significantly between groups. These results suggest that adding ultrasound radiomics to clinical information offers additional diagnostic value beyond routine assessment. Nevertheless, this study has several limitations. First, although a range of clinical and laboratory variables were evaluated, the final clinical component of the model was limited to eGFR after rigorous feature selection. Therefore, the absence of other clinically relevant variables should not be interpreted as evidence of limited clinical importance, but rather as a result of data-driven model selection within the current biopsy-based cohort. Second, although both internal and external validation were performed, the overall sample size remained modest, particularly in the external validation cohort. The reduced external performance may be partly explained by limited clinical separability between DKD and NDKD in that cohort, as no baseline clinical or laboratory parameters differed significantly between the two groups. Thus, the proposed model should be regarded as an auxiliary tool based on biopsy-defined classification, rather than a substitute for comprehensive clinical assessment or renal biopsy. Third, NDKD represents a heterogeneous pathological category. Although the distribution of NDKD subtypes was summarized, the present model was designed to distinguish DKD from overall biopsy-confirmed NDKD, rather than to identify specific NDKD subtypes or overlapping pathological patterns. Larger pathologically stratified cohorts are needed to develop multiclass or multilabel diagnostic models. Fourth, ultrasound images were retrospectively acquired and were not fully standardised across centres. Inter-centre differences in acquisition protocols, scanner settings, image quality and ROI segmentation may have affected radiomic feature reproducibility. Future prospective, multi-device and multicentre studies with standardised imaging protocols and radiomics harmonisation strategies, such as ComBat-based batch-effect correction, are warranted. Fifth, feature selection and model training were not performed within a nested cross-validation framework. This may have led to optimistic estimates of model performance and increased the risk of overfitting, despite the use of internal and external validation cohorts. Therefore, the reported model performance should be interpreted with appropriate caution. Finally, although SHAP analysis improved model interpretability, the biological meaning of radiomic features remains exploratory. Integrating radiomics with histopathological findings and molecular profiles may help clarify the biological basis of ultrasound-derived radiomic signatures in DKD and NDKD. Conclusion In this multicenter study, we developed and validated machine learning models to distinguish DKD from NDKD in patients with type 2 diabetes mellitus. The integrated radiomics–clinical model, which combined wavelet-based radiomic features with eGFR, achieved the most stable diagnostic performance across the training, internal validation and external validation cohorts. Decision curve analysis and SHAP interpretation further supported its clinical utility and interpretability. By leveraging routinely available ultrasound images and clinical data, this model may provide a non-invasive and practical tool for preliminary screening in primary-level hospitals or nephrology departments, thereby offering auxiliary support for individualized diagnosis, renal biopsy decision-making and kidney-preserving treatment strategies in real-world clinical settings. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3 Supplementary Material 4 Supplementary Material 4 Supplementary Material 5 Supplementary Material 5 Supplementary Material 6 Supplementary Material 6 Supplementary Material 7 Supplementary Material 7 Supplementary Material 8 Supplementary Material 8 Supplementary Material 9 Supplementary Material 9 Supplementary Material 10 Supplementary Material 10