Elderly living with T2DM were recruited from The Affiliated Hospital of Hangzhou Normal University between May 2021 and October 2022. The inclusion criteria were (1): age ≥ 60 years, (2) diagnosis of T2DM, and (3) urinary albumin-to-creatinine ratio (UACR) < 300 mg/g. Patients were excluded if they: (1) were under 60 years of age, (2) had non-T2DM diagnoses, (3) had UACR ≥ 300 mg/g, uremia, or renal failure, (4) had severe infections, systemic immune diseases, organ failure, or malignancies, or (5) lacked essential clinical information. After applying these criteria, 786 patients were included in the study (Figure 1).

T2DM and early DN were diagnosed according to the 1999 World Health Organization criteria and Kidney Disease: Improving Global Outcomes (KDIGO) guidelines (25, 26). Early DN was defined as a UACR between 30–300 mg/g, confirmed by at least two of three tests conducted over 3–6 months.

Patient data were obtained from the hospital’s electronic medical records and a questionnaire survey. The data were categorized into three domains:

Data were subjected to an extensive cleaning process to ensure quality and completeness. Records with missing values exceeding 15% were excluded (32, 33), while missing values under 15% were addressed using simple interpolation. We used SPSS 22.0 for simple interpolation (< 15% missing): imputation methods (1): continuous: series mean imputation; (2) categorical: mode imputation. Duplicate and extreme outliers were removed.

Statistical analyses were conducted using SPSS 22.0 and R software (v4.2.0). Continuous variables with normal distributions were compared using t-tests or ANOVA and reported as mean ± standard deviation (SD). Non-normally distributed data were analyzed with non-parametric tests (Wilcoxon rank-sum or Kruskal-Wallis H test) and reported as medians with interquartile ranges (P25, P75). Categorical variables were compared using chi-square or Fisher’s exact tests and expressed as counts and percentages. A two-tailed p-value < 0.05 was considered statistically significant.

Least absolute shrinkage and selection operator (LASSO) regression was performed to identify key predictors, followed by multivariate logistic regression (34). A stepwise backward elimination method was used to finalize the model variables. Model discrimination was evaluated using receiver operating characteristic (ROC) curves, with the area under the curve (AUC) calculated using the “pROC” package in R. AUC comparisons between models were conducted with the DeLong test. Calibration and clinical utility were assessed using calibration curves and decision curve analysis (DCA), with the “rms” and “nricens” packages, respectively. Nomograms were constructed for visualization.

The final model variables were used to develop predictive models using seven machine learning algorithms: logistic regression (LR), decision tree (DT), random forest (RF), k-nearest neighbors (KNN), light gradient boosting machine (Light GBM), support vector machine (SVM), and XGBoost (XGB). The performance of each algorithm was evaluated using accuracy, precision, recall, F1 score, and AUC.

A total of 786 participants were included in this study, with 550 allocated to the training set and 236 to the validation set in a 7:3 ratio. Baseline comparisons revealed significant differences in the variables “cold extremities”, diabetes duration, total bilirubin (TB), and indirect bilirubin (IB) between the training and validation sets (P < 0.05). No significant differences were observed for other variables (P > 0.05), confirming the comparability between the training and validation sets (Table 1).

LASSO regression was applied to identify relevant predictors, using early DN as the dependent variable and 48 candidate variables as independent variables. The optimal model was determined through ten-fold cross-validation (Figure 2). LASSO analysis identified 21 significant factors, including cold extremities, lumbar and knee weakness, blurred vision, nocturnal polyuria, limb numbness, hypertension, cerebral infarction, hyperlipidemia, diabetic retinopathy (DR), diabetes duration, systolic blood pressure (SBP), indirect bilirubin (IB), blood calcium (Ca), blood urea nitrogen (BUN), uric acid (UA), retinol-binding protein (RBP), triglycerides (TG), C-reactive protein (CRP), renal parenchymal calcification, carotid intima thickening, and carotid artery plaque.

Through backward stepwise logistic regression, 15 independent risk factors were identified: lumbar and knee weakness, blurred vision, nocturnal polyuria, hypertension, cerebral infarction, hyperlipidemia, DR, SBP, IB, Ca, UA, RBP, CRP, carotid intima thickening, and carotid artery plaque (Supplementary Figure S1).

Three prediction models were developed:

Model performance was evaluated using the area under the receiver operating characteristic curve (AUC). In the training set, AUCs were 0.852 (95% CI: 0.819–0.884) for Mod A, 0.881 (95% CI: 0.852–0.910) for Mod B, and 0.902 (95% CI: 0.876–0.928) for Mod C. In the validation set, AUCs were 0.801 (95% CI: 0.741–0.860), 0.832 (95% CI: 0.781–0.883), and 0.855 (95% CI: 0.808–0.902) for Mod A, Mod B, and Mod C, respectively. Calibration and decision curve analyses further confirm that Mod C exhibited the best predictive performance (Figure 3). Based on these findings, Mod C was selected to construct a nomogram for clinical application (Figure 4).

Using the multi-feature data from Mod C, seven machine learning algorithms were evaluated: logistic regression (LR), decision tree (DT), random forest (RF), k-nearest neighbors (KNN), light gradient boosting machine (Light GBM), support vector machine (SVM), and XGBoost (XGB). Among these, RF demonstrated the highest performance. In the training set, RF achieved an AUC, sensitivity, specificity, precision, recall, and F1 score of 1.000. In the validation set, RF yielded an AUC of 0.894, sensitivity of 0.667, specificity of 0.877, precision of 0.769, recall of 0.667, and F1 score of 0.714 (Supplementary Figure S2, Figure 5, Table 2). SHAP analysis identified DR, UA, and carotid artery plaque as the top three predictors influencing early DN in elderly T2DM patients (Supplementary Figure S3).

To account for potential age-related differences in DN, subgroup analyses were performed by stratifying participants into three age groups: 60–69 years, 70–79 years, and ≥ 80 years. In the training set, AUCs were 0.933 (95% CI: 0.902–0.964) for ages 60–69, 0.935 (95% CI: 0.900–0.970) for ages 70–79, and 0.837 (95% CI: 0.756–0.917) for ages ≥ 80. In the validation set, AUCs were 0.879 (95% CI: 0.814–0.944), 0.923 (95% CI: 0.861–0.985), and 0.718 (95% CI: 0.559–0.878), respectively (Supplementary Figure S4). Forest plots and age-specific nomograms were generated to visualize risk factors for each age group and to facilitate clinical decision-making (Supplementary Figures S5, S6).

The predictive model identified several key factors related to early DN in older patients. TCM symptoms such as lumbar and knee weakness, blurred vision, and nocturnal polyuria emerged as critical predictors. These symptoms reflect the underlying deficiencies in kidney Yin and Yang, which are central to TCM pathophysiology (37, 38). TCM classifies DN based on syndrome differentiation (e.g., Qi-Yin deficiency), which reflects systemic pathological changes even in early-stage DN when conventional biomarkers (e.g., microalbuminuria) may still be within normal ranges (39). Lumbar and knee weakness, blurred vision, and nocturnal polyuria may serve as early warning signs, aiding in the identification of high-risk patients before structural kidney damage becomes evident. Age-related declines in kidney function further exacerbate these deficiencies, as documented in prior studies (40, 41).

Ultrasound imaging indicators, particularly carotid intima thickening and carotid artery plaque, were also identified as significant predictors. These markers are commonly associated with diabetic macroangiopathy and atherosclerosis, which are closely linked to DN progression (42–48). Although not a gold standard, ultrasound provides valuable structural and hemodynamic insights, such as: carotid intima thickening and carotid artery plaque (suggesting diabetic macroangiopathy and atherosclerosis) (49). These findings, combined with laboratory tests, enhance early DN detection sensitivity. While the relationship between carotid intima-media thickness (IMT) and DN remains controversial, our study reinforces its role as an independent risk factor for early DN. TCM syndrome progression (e.g., from Qi-Yin deficiency to Yang deficiency) correlates with ultrasound-documented structural decline (e.g., carotid intima thickening and carotid artery plaque), offering a holistic view of disease progression. We acknowledge the need for further research on correlations between specific TCM syndromes (e.g., Spleen-Kidney Qi deficiency) and ultrasound parameters (e.g., carotid intima thickening and carotid artery plaque). Prospective studies validating a combined TCM-ultrasound predictive model would strengthen clinical utility.

Clinical factors such as hypertension, cerebral infarction, hyperlipidemia, diabetic retinopathy (DR), systolic blood pressure (SBP), indirect bilirubin (IB), calcium (Ca), uric acid (UA), retinol-binding protein (RBP), and C-reactive protein (CRP) were also significant predictors. Known risk factors, including DR, hypertension, and hyperlipidemia, align with established evidence (7, 11, 50). Notably, UA, an oxidative stress marker, is strongly associated with proteinuria, glomerular filtration rate decline, and DN progression (51–54). RBP, an early diagnostic marker of proximal tubular dysfunction, also emerged as a significant predictor (55). Interestingly, IB, with its antioxidant properties, was identified as a protective factor against DN, consistent with emerging research on its role in mitigating oxidative stress (56). Further studies are warranted to explore the mechanisms underlying these associations.

Given the known epidemiological and physiological differences in DN risk across age groups, subgroup analyses were conducted. Predictive models showed robust performance across all age groups, with AUC values exceeding 0.7 in both training and validation datasets. The performance was particularly strong for patients aged 60–79 years, with slightly lower predictive accuracy for those aged ≥ 80 years, potentially due to smaller sample sizes. These findings emphasize the adaptability and applicability of the model for various age cohorts, providing a practical tool for early DN risk stratification in older populations.

The integration of clinical, TCM, and ultrasound imaging data was further validated through ML approaches. Among the seven algorithms evaluated, the random forest (RF) model exhibited the best predictive performance, with an AUC of 0.894, sensitivity of 0.667, specificity of 0.877, precision of 0.769, recall of 0.667, and F1 score of 0.714 in the validation set. These results confirm RF’s robustness, aligning with prior research that highlighted its superiority in predicting progression to end-stage renal disease (ESRD) (35, 57, 58). Notably, few existing ML models incorporate TCM and imaging data, underscoring the novelty and clinical relevance of our approach. Risk Equations for Complications Of type 2 Diabetes (RECODe) were derived from the Action to Control Cardiovascular Risk in Diabetes (ACCORD) study (59, 60). Compared with established DN prediction tools like RECODe, our model incorporates TCM syndromes (e.g., lumbar and knee weakness, blurred vision, and nocturnal polyuria) and ultrasound parameters (e.g., carotid intima thickening and carotid artery plaque). This multimodal design targets earlier prediction windows than RECODe. At the same time, the RECODe predicts 3–5 year renal function decline, our model identifies pre-clinical risks (TCM manifestations precede lab abnormalities).

This study provides a comprehensive predictive framework for early DN in elderly living with T2DM, incorporating multidimensional data to improve accuracy. The integration of TCM symptoms and ultrasound imaging data represents a novel and clinically valuable approach, facilitating early detection and targeted management. However, studying has several limitations.