Diabetes is a complex metabolic disease, which can lead to a variety of complications, chronic damage to a variety of organs and tissues and a serious threat to the quality of life of patients [13]. In recent years, the incidence of diabetic nephropathy has increased year by year and has become the main source of end-stage nephropathy [14]. However, in diabetic patients with kidney injury, the diagnosis of diabetic nephropathy depends on the pathological results of kidney puncture, but it can cause damage to the blood vessels around the calyces, and even the risk of perforation of the abdominal organs or pleura [15, 16]. Therefore, for diabetic patients with chronic kidney injury, it is very important to find a better noninvasive method to identify diabetic nephropathy for targeted treatment.

The types of nephropathy that induce kidney damage in diabetic patients are diverse and intricate. Besides diabetic nephropathy, conditions such as chronic renal tubular injury, glomerulonephritis, IgA nephropathy and hypertensive nephropathy can also occur. Moreover, two or more types of nephropathy may coexist. The accuracy of diagnosing different types of kidney diseases significantly influences their clinical treatment strategies and goals, thereby altering the course and prognosis of kidney diseases. Identifying potential non-diabetic renal diseases (NDRDs) holds great clinical significance [17].

Whether biochemical indicators and ultrasonic medical indicators can be organically combined to evaluate the kidney status of patients more comprehensively, artificial intelligence came into being. Through the random forest algorithm of machine learning, the laboratory inspection and CEUS parameters are incorporated into the random forest model, and the data are automatically learned, the rules are found, and the rules are used to make predictions and decisions [18].

CEUS can quantitatively evaluate renal microcirculation perfusion through intravenous injection of microbubble contrast agent combined with real-time dynamic imaging technology, and its quantitative parameters can noninvasive and sensitive mirror the hemodynamic changes caused by microvascular injury in the glomeruli and renal tubulointerstitium [19]. This provides valuable insights into the microvascular status of the kidneys, enabling early detection and monitoring of potential renal pathologies related to microvascular injury. In this study, among the top eight important factors screened by single factors, there are two CEUS parameters, namely PEAK and AUC. Y. Wang [20] et al. also reached a similar conclusion, suggesting that quantification of CEUS parameters is a feasible parameter to identify DN, PE and AUC in patients with diabetes complicated with kidney injury.

In this model, in addition to eGFR/EPI, serum creatinine and serum protein concentrations [21,22,23], which are routinely used to assess renal function, biochemical indicators were also included in β2 microglobulin, serum cystatin and total cholesterol.

β2 microglobulin is a low molecular weight endogenous protein secreted mainly by lymphocytes and other nucleated cells, and is a sensitive indicator of glomerular filtration function and a marker of renal tubule injury [24]. Takayuki Uemura [25] et al. suggested that serum β2-MG could be predictive of kidney failure by studying patients with biopsy-confirmed diabetic nephropathy.

Serum cystatin C is produced stably by all nucleated cells and is completely filtered through the glomeruli. It is a sensitive marker of renal function, especially in the early stages of diabetic nephropathy, and its sensitivity is due to serum creatinine [26, 27]. Alexandra-Mihaela Visinescu [28] et al. concluded that CKD is an important complication in diabetic patients, and cystatin C is more sensitive than serum creatinine in the assessment of renal function, and combining cystatin C with other biomarkers can improve the diagnostic accuracy of CKD in diabetic patients. Zhao Ping [29] et al. investigated the relationship between serum cystatin C level and renal microvascular perfusion in patients with DN, suggesting that CEUS parameter reflects the changes of renal microvascular perfusion in patients with DN, and AUCs may be a useful indicator of GFR decline in DN patients with elevated serum cystatin C.

As a crucial indicator of lipid metabolism, total cholesterol is closely related to the occurrence and development of diabetic nephropathy. Hypercholesterolemia can induce inflammation and oxidative stress, leading to the proliferation of mesangial cells and the damage of renal tubule interstitial [30]. Concurrently, kidney function damage can disrupt lipid metabolism, resulting in increased cholesterol levels. In addition, the increase of total cholesterol can induce atherosclerosis, leading to vascular complications such as renal artery lesions and consequently reducing renal blood perfusion. The interplay between these factors exacerbates the clinical manifestations of diabetic nephropathy, driving its deterioration [31].

In this study, a random forest model was constructed by combining contrast-enhanced ultrasound parameters and biochemical indexes with the method of machine learning. Xuee Su [32] et al. used a machine learning model containing two-dimensional ultrasound imaging and biochemical data to diagnose patients with type 2 diabetes mellitus (T2DM). By incorporating 10 ultrasonic features and biochemical indicators such as total cholesterol, triglyceride, blood creatinine and urea nitrogen, she constructed a model for prediction and evaluation, and obtained good sensitivity, specificity and accuracy of evaluation. In our study, all the patients included in this model had obtained renal pathological results, which served as the gold standard to guide the model assessment, providing a good orientation for the evaluation. And it has achieved good consistency and accuracy.

Limitations of this study: (1) The sample size was limited to retrospective studies from a single center; the predictive model developed in this paper may reflect the distribution of biomarkers in a specific population (such as serum creatinine, estimated glomerular filtration rate based on the EPI formula), but these distributions do not apply to cohorts with different genetic or environmental backgrounds; these findings require further multi-institutional validation with larger sample sizes; (2) this was a retrospective study that could not completely avoid missing data and measurement bias, so further studies must include more candidate biomarkers to develop predictive models in the future.

The diagnostic model developed in this study, which is based on eight key biomarkers, provides a practical tool for the differential diagnosis of diabetic nephropathy (DN) and non-diabetic renal disease (NDRD). Its core advantage lies in enabling noninvasive diagnosis using routine laboratory and examination indicators, effectively compensating for the limitations of renal biopsy. It is particularly suitable for patients with biopsy contraindications or those with poor physical conditions, and holds significant value in optimizing clinical workflows and alleviating the burden on patients. From the perspective of clinical translation, the model can be seamlessly integrated into existing diagnostic and treatment pathways by incorporating routine tests and embedding into clinical decision support systems (CDSS) [33], thereby playing a role in scenarios such as disease stratification and treatment monitoring.

Of course, future efforts could focus on incorporating multi-dimensional indicators to enhance its efficacy and optimizing through interpretable algorithms to improve clinical acceptance. Additionally, clinicians will receive training on interpreting the model’s output to improve diagnostic consistency, with emphasis on the model serving as a complement to, rather than a replacement for, clinical judgment. The interface will include visual aids to contextualize results, ensuring usability across diverse clinical settings.

In conclusion, our approach demonstrates the significant role of integrating machine learning with contrast-enhanced ultrasound and clinical data in model construction and the enhancement of disease detection strategies, which is poised to find widespread application across various medical disciplines.