This study employed a prospective single-center cohort design to develop and validate a predictive diagnostic model for chronic kidney disease (CKD) based on ultrasound viscoelastic imaging. The study cohort comprised consecutive CKD patients admitted to the Department of Nephrology at the Fourth Affiliated Hospital of Zhejiang University School of Medicine between September 2024 and March 2025. The research protocol was approved by the Fourth Affiliated Hospital of Zhejiang University School of Medicine (K2025134), and written informed consent was obtained from all participants. The study strictly adhered to the KDIGO Clinical Practice Guidelines for CKD diagnosis (Stevens and Levin, 2013). The sample size estimation was based on a 9.1% prevalence rate of the disease in the population, with an anticipated 20% exclusion rate of cases. This study used “chronic kidney disease (CKD) status” as the primary endpoint (dichotomous outcome: 1 = CKD, 0 = non-CKD). To detect an association strength of an odds ratio = 2.0 between the target exposure factor and CKD (α = 0.05, power = 80%), the sample size was calculated based on a case-control study design using a logistic regression model framework for dichotomous outcomes. The calculation was performed using PASS 15.0 statistical software, which indicated a requirement of 140 CKD patients to be included. The inclusion criteria required participants to demonstrate either persistent renal dysfunction, defined as two consecutive measurements taken at least 12 weeks apart showing an estimated glomerular filtration rate <90 mL/min/1.73 m², or significant proteinuria indicated by an albumin-to-creatinine ratio (ACR) of at least 30 mg/g or a protein-to-creatinine ratio (PCR) of at least 150 mg/g. All CKD patients had ≥6 months of documented medical history to exclude acute kidney injury. Healthy controls met eGFR ≥90 mL/min/1.73 m² with exclusion of chronic conditions (diabetes, hypertension) and body mass index outside 18.5-24 kg/m² range. The study established standardized exclusion criteria to eliminate potential confounding factors, specifically excluding patients with complex renal cysts, renal artery stenosis characterized by a peak systolic velocity (PSV) exceeding 180 cm/s combined with a resistance index (RI) above 0.7, uncontrolled hypertension manifesting as systolic blood pressure greater than 160 mmHg or diastolic pressure surpassing 100 mmHg, and severe cardiac dysfunction classified as NYHA class Ⅲ-Ⅳ. To maintain optimal imaging quality standards, the study further excluded participants demonstrating subcutaneous fat thickness beyond 8 cm or exhibiting ultrasound image stability indices lower than 85%. All enrolled subjects completed thorough clinical evaluations incorporating demographic profiling, comprehensive laboratory testing, and detailed comorbidity analysis, with every case receiving precise staging following the established KDIGO classification system.

All renal viscoelastic imaging examinations were performed using the Mindray Resona A20 ultrasound system equipped with an SC7-1U high-frequency transducer (1.2-6.0 MHz) and dedicated viscoelastic analysis module, which simultaneously quantified tissue elasticity and viscosity parameters based on the Voigt biomechanical model, with E mean values reflecting tissue stiffness in kPa and Vi mean values in Pa·s derived from shear wave dispersion characteristics across the 50-400 Hz frequency band. Participants maintained a standardized prone position with arms crossed beneath the forehead for stability during examinations performed by two sonographers with over 5 years of specialized renal ultrasound experience following a double-blind protocol where operators remained unaware of clinical data, beginning with B-mode images of the maximal longitudinal kidney section followed by viscoelastic imaging in the identical plane during which a standardized 2 × 2 cm region of interest was selected within the renal cortex while avoiding vessels and collecting systems during a 6-s breath-hold, with five repeated measurements per kidney recorded for mean values (Figure 1). Stringent quality control criteria including measurement coefficient of variation below 30% within ROI, consistent 6 mm diameter ROI for all acquisitions, and sound wave attenuation correction at 0.5 dB/cm/MHz were implemented to ensure data reliability. Inter-rater reliability was assessed for all ultrasound parameters using intraclass correlation coefficient (ICC) for analysis according to a predefined standardized examination manual.

Statistical analyses were performed using R software (version 4.4.3, R Foundation for Statistical Computing, Vienna, Austria). The normality of all continuous variables was rigorously assessed using Shapiro-Wilk tests (n < 50) or Kolmogorov-Smirnov tests (n ≥ 50). Based on these evaluations, continuous variables were expressed as mean ± standard deviation (normally distributed) or median (interquartile range) (non-normally distributed), and compared using independent Student's t-tests or Mann-Whitney U tests. Categorical variables were presented as frequencies (percentages) and analyzed by chi-square or Fisher’s exact tests (for cell counts <5), with statistical significance set at P < 0.05.

Variables with >20% missingness were excluded from analysis. When the proportion of missing data exceeds 20%, the remaining data can hardly reflect the true distribution of variables. In particular, non-random missing data will exacerbate bias. In contrast, when the missing proportion is ≤ 20%, the remaining data retain core characteristics. Using 20% as the threshold helps retain valuable variables (detailed missing proportions in Supplementary Table S1) while avoiding interference from low-quality data, thus ensuring statistical power and model stability. For remaining variables, missing data were handled using multiple imputation by chained equations (MICE). Continuous variables were imputed using predictive mean matching (for normally distributed data) or quantile regression (for skewed data), while categorical variables used logistic regression imputation. The study strictly adhered to KDIGO guidelines for clinical data collection from both CKD patients and healthy controls. The analytical framework comprised dependent and independent variables, where the dependent variable represented the CKD diagnosis based on KDIGO criteria (Stevens and Levin, 2013). Independent variables included two categories of measures: conventional ultrasound-measured renal length and viscoelastic parameters, along with basic demographic characteristics (age, sex, and body mass index). All ultrasound examinations were conducted by experienced sonographers with over 5 years of practice following standardized image acquisition and processing protocols to ensure data consistency and comparability throughout the study.

Variable selection was performed using 10-fold cross-validated LASSO regression (repeated 100 times) (Van Beek et al., 2024) to optimize the penalty coefficient λ, where λ.min was selected as the value yielding minimum cross-validated mean squared error (MSE), and λ.1se represented the most parsimonious model within one standard error of the minimum MSE. Variables were retained only if selected in the majority of 1000 bootstrap iterations to ensure stability. Variance inflation factors (VIF) were calculated to confirm the absence of multicollinearity. Model performance was rigorously evaluated through discrimination (AUC with DeLong 95% CI, sensitivity/specificity at Youden-index optimal thresholds), calibration (calibration plots, Brier score and Hosmer-Lemeshow goodness-of-fit test), and clinical utility (decision curve analysis across 0%–100% risk thresholds). Internal validation was enhanced via 1000 bootstrap resamples to derive optimism-adjusted metrics. Results were implemented via a dynamic nomogram explicitly linked to KDIGO staging criteria and therapeutic decision points, with probability thresholds informed by decision curve analysis.

To evaluate the model’s predictive performance across different CKD progression stages, we implemented a multiclass ROC analysis framework with the following specifications: (1) Stage classification followed KDIGO guidelines (Class 1: normal controls; Class 2: G1-3 stages; Class 3: G4-5 stages); (2) Multiclass handling employed one-vs-rest (OvR) methodology to generate stage-specific ROC curves, reporting both sensitivity (True Positive Rate) and 1-specificity (False Positive Rate) metrics; (3) Comprehensive evaluation included micro-averaged AUC for overall discriminative ability alongside stage-specific AUC values. All performance metrics were derived with 1000 bootstrap iterations to generate 95% confidence intervals, ensuring robust variance estimation.

Following data preprocessing to handle missing values, the final analysis included 228 participants comprising 138 chronic kidney disease (CKD) patients and 90 healthy controls. Regarding the imbalance in sample size between the case group (138 cases) and the control group (90 cases), we implemented targeted statistical adjustments: robust standard errors were used in the logistic regression analysis to reduce the impact of inter-group sample size differences on the variance of parameter estimates. Furthermore, the model calibration curve showed that the mean absolute error between predicted probabilities and actual observed probabilities was 0.03 (<0.05), which further confirms that the calibration effect of the model under this sample structure is acceptable. All intraclass correlation coefficients (ICC) for measurements exceeded 0.8, indicating excellent intra-observer reliability (Supplementary Table S2). As shown in Table 1, comparative analysis of renal ultrasound parameters revealed significant intergroup differences across multiple indicators. CKD patients exhibited markedly reduced renal dimensions bilaterally compared to controls (left kidney: 9.76 ± 1.42 cm vs. 10.32 ± 0.88 cm; right kidney: 9.73 ± 1.42 cm vs. 10.23 ± 0.68 cm; both P < 0.001), consistent with the characteristic renal atrophy observed in CKD progression and confirming the diagnostic value of renal length measurements. Regarding tissue characterization parameters, the CKD group demonstrated significant histopathological alterations with elevated elastic moduli bilaterally (left kidney E mean: 6.02 ± 1.33 kPa vs. 5.34 ± 0.50 kPa; right kidney: 6.07 ± 1.17 kPa vs. 5.36 ± 0.34 kPa; both P < 0.001) and increased shear wave velocities (left kidney Cs mean: 1.39 ± 0.14 m/s vs. 1.32 ± 0.06 m/s; right kidney: 1.40 ± 0.12 m/s vs. 1.33 ± 0.04 m/s; both P < 0.001), suggesting increased renal tissue stiffness potentially associated with fibrotic processes. Most notably, viscosity parameters (Vi mean) reflecting inflammatory activity showed significant elevation in CKD patients (1.99 ± 0.31 Pa·s vs. 1.64 ± 0.11 Pa·s bilaterally, P < 0.001), providing quantitative biomarkers for renal inflammation assessment.

Demographic analysis revealed CKD patients were older (49.77 ± 16.92 years vs. 41.36 ± 10.76 years, P < 0.001) with comparable BMI between groups (24.05 ± 3.55 kg/m² vs. 24.34 ± 3.79 kg/m², P = 0.457), effectively excluding body mass index as a confounding factor. Gender distribution analysis showed female proportions of 42.0% (58/138) in CKD and 51.1% (46/90) in control groups (P = 0.178), with male predominance (58.0%) observed in CKD patients. Despite these non-significant demographic differences, both BMI and gender were incorporated as covariates in subsequent modeling to ensure the identified ultrasound parameter differences specifically reflected CKD-related pathological changes rather than demographic variations.

The LASSO regression analysis (Figures 2, 3) demonstrated that as the penalty parameter λ increased, the coefficients of each parameter gradually shrank toward zero. The model achieved optimal performance at log(λ) = −3.705, retaining four significant variables. As λ decreased, the number of variables systematically reduced from 10 to 0, confirming the stability of the variable selection process. Among the renal ultrasound parameters, the viscosity coefficients of both kidneys, the elasticity parameter of the right kidney (RE) and age were retained in the final model. Multicollinearity assessment confirmed all retained variables had Variance Inflation Factors (VIF) < 5 (Supplementary Table S3). Notably, the viscosity parameters (LVi/RVi) exhibited the highest absolute coefficient values (Table 2), indicating their dominant contribution to CKD diagnosis. In contrast, age showed the smallest absolute coefficient, suggesting its limited independent diagnostic value after adjusting for other factors.

ROC analysis of the training set revealed excellent diagnostic performance (Figure 5), with an AUC of 0.932 (95% CI: 0.908–0.956). At the optimal threshold determined by Youden’s index, the sensitivity was 87.8% (80.0%-93.3%) and specificity was 88.4% (80.2%-94.0%). The ROC curve (Figure 4) clearly illustrated the relationship between the true positive rate (TPR) and false positive rate (FPR) across different thresholds, demonstrating the model’s ability to maintain high TPR while effectively controlling FPR. In the validation set (Figure 5), at a cutoff of 0.383, sensitivity and specificity were both 0.884 and 0.878, respectively. The TPR-FPR curve further indicated that the model achieved TPR >0.8 at FPR <0.2, outperforming the random guessing line.

Calibration analysis (Figure 6; Table 3) demonstrated strong agreement between predicted probabilities and observed outcomes. The diagnostic model exhibited outstanding discriminative ability, with an AUC of 0.932, and explained approximately 63.7% of the outcome variance (R² = 0.637), confirming its high diagnostic accuracy. Calibration metrics indicated close alignment between logistic calibration curves, nonparametric calibration curves, and the ideal calibration line in the intermediate-risk range (actual probability = 0.4). (Supplementary analysis: Hosmer-Lemeshow goodness-of-fit test indicated potential departure from perfect calibration [P = 0.001], though this may be influenced by the test’s sensitivity to group partitioning and large sample size, consistent with other calibration evidence.) The Brier score of 0.104 suggested low overall prediction error, while the calibration intercept of 0.000 indicated slight systematic overestimation. A calibration slope of 1.000 suggested potential mild overfitting; however, the model maintained good accuracy across most risk ranges, with minor overestimation in the low-risk range (predicted probability: 0–0.3) and slight underestimation in the high-risk range (E90 = 0.086). Decision curve analysis (Figure 7) further confirmed clinical utility, demonstrating superior net benefit versus “treat-all” and “treat-none” strategies at risk thresholds of 20%-80%, with peak utility at 40% threshold probability. These results collectively confirm that the ultrasound-based diagnostic model provides well-calibrated risk predictions while maintaining excellent discriminative performance.

A visualized CKD risk scoring system was developed based on the selected ultrasound parameters (Figure 8). This integrated scoring system incorporates multidimensional diagnostic indicators, where the total score is calculated as: Total Points = RE (renal elasticity) + LVi/RVi (viscosity parameters) + age. Each parameter was assigned a specific weight according to its regression coefficient, and the total score was converted into a CKD probability ranging from 0.1 to 0.9 via a linear predictor using the logistic function specified in Supplementary Material Equation. The scoring system highlights the significant contribution of viscosity parameters, which reflect renal inflammatory activity, along with elasticity (RE) and age, forming a comprehensive multidimensional evaluation framework. Notably, the viscosity parameters alone could contribute over 30 points to the total score, reaffirming their pivotal diagnostic value in CKD assessment.

ROC analysis of the test set (Figure 9) revealed excellent stage-specific diagnostic accuracy. The model achieved highest discrimination for end-stage CKD (Class 3: AUC = 0.918, 95% CI: 0.890–0.946), followed by normal controls (Class 1: AUC = 0.884, 95% CI: 0.850–0.918) and early-stage CKD (Class 2: AUC = 0.774, 95% CI: 0.730–0.818). DeLong’s test confirmed significant AUC differences between Class 3 vs Class 2 (P < 0.001). The micro-average AUC (0.869) indicated consistent overall performance, supporting clinical utility in distinguishing high-risk patients.