This study developed a model based on the machine learning ensemble algorithms, featuring simple, non-invasive, and easily accessible ultrasound imaging metrics with good interpretability, facilitating clinical application and dissemination. Screening 110 models to identify the optimal prediction model not only improved diagnostic accuracy but also provided strong support for clinical decision-making. By incorporating patients’ baseline data, serological indicators, and ultrasound imaging data, we identified potential new predictive factors, including gallbladder wall thickness, polyp size, polyp echo, and pedunculation. The SVM + RF model achieving a sensitivity and specificity of 85.52% and 97.73%, respectively, and a diagnostic accuracy of 90.08%, demonstrating excellent diagnostic performance.

Infiltrative gallbladder cancer accounts for approximately 5–10% of tumours that arise from malignant transformation of intra-gallbladder tumours, such as adenomas [21], fitting the ‘adenoma-carcinoma sequence’ concept. Approximately two-thirds of patients with intra-gallbladder tumours may eventually develop infiltrative cancer. Therefore, timely diagnosis and intervention for potentially malignant gallbladder adenomas are crucial. However, it is challenging to distinguish them from benign cholesterol polyps in clinical practice. This has led to some patients undergoing cholecystectomy as a precaution against gallbladder malignancy, even when the postoperative pathology reveals cholesterol polyps. In particular, the inability to effectively differentiate polyps > 10 mm from gallbladder adenomas increases the risk of unnecessary surgical removal [22,23,24,25]. Building on our previous research [16], this study utilised a multicentre retrospective data analysis combined with the latest ML algorithms to establish an effective preoperative prediction model, SVM + RF, aimed at enhancing the ability to differentiate gallbladder adenomas from cholesterol polyps, thereby providing better individualised treatment strategies for patients.

Risk factor analysis of gallbladder adenomas is essential for the formulation of individualised treatment strategies. Studies have shown that polyp size is the most common risk factor for gallbladder adenomas [26, 27]. Wang et al. [28] evaluated 89 patients with GBPs measuring 1–2 cm using conventional and contrast-enhanced ultrasonography and found that a maximum diameter of 1.45 cm was the optimal cutoff value for predicting adenomatous polyps. Similar studies have indicated that polyps > 1 cm significantly increase the risk of malignancy [29,30,31], consistent with our findings. Mellnick et al. [32] reported that adenomas are mostly sessile or pedunculated hypoechoic polyps. Although gallbladder wall thickness and polyp base thickness are considered risk factors for gallbladder cancer, no direct link between gallbladder wall thickness and adenoma formation has been established and requires further research. Through ML analysis of ultrasound imaging features, we identified polyp size, pedunculated hypoechoic polyps, and gallbladder wall thickness as independent risk factors for gallbladder adenomas. In addition to these conventional factors, previous studies have found a correlation between obesity, especially visceral fat accumulation, and an increased risk of GBPs [35, 36]. Yamin et al. highlighted the relationship between dyslipidaemia and GBP formation, particularly in the Chinese population [37], indicating that lifestyle and metabolic factors might play a role in the pathogenesis of GBPs. Our study also found that metabolic syndrome is a potential risk factor for gallbladder adenomas. To date, no study has reported a relationship between ADA and gallbladder adenomas. Only one study by Tounsi et al. [38] on the association between ADA and tumour angiogenesis in patients with gallbladder cancer under nitroxidative stress indicated that ADA influences microvascular density, possibly regulating adenosine levels and affecting immune and endothelial cells, indirectly participating in angiogenesis regulation. A high microvascular density/glutathione ratio is a potential biomarker for gallbladder cancer, suggesting a correlation between ADA levels and tumour progression. Nonetheless, controversy regarding the risk factors for gallbladder adenomas still remains, requiring further research to identify patients needing proactive intervention.

To improve preoperative diagnostic accuracy, various ultrasound imaging-based risk assessment models and prediction tools have emerged in recent years. These tools aim to combine multiple independent risk factors to predict the malignancy risk of GBPs and guide clinical decisions [16, 39, 40]. For example, Liu et al. [31] retrospectively analysed 423 patients with GBPs and developed a prediction model combining ultrasound imaging features, identifying solitary lesions, larger polyps, and irregular morphology as independent risk factors for gallbladder adenomas, thereby providing more accurate predictions for adenomas > 1 cm. Zhu et al. proposed a new risk-scoring system, the gallbladder reporting and data system (GB-RADS), to explore the risk factors for gallbladder adenomas [33]. This system evaluated the ultrasound imaging features of 136 patients, forming a scoring standard for gallbladder adenomas, including enhancement patterns, wash-out characteristics, and vascularity. Additionally, this system simplifies the assignment of different weights to each risk factor, ultimately calculating the total score and providing a straightforward risk assessment for physicians. Compared with traditional ultrasound imaging methods, the CEUS-based scoring system significantly improves diagnostic accuracy, as confirmed in previous studies [34]. Zhang et al. [41] developed a nomogram prediction model focusing on GBPs measuring 10–15 mm in diameter. By combining clinical and ultrasound imaging features, this model provides a quantitative risk score, allowing physicians to evaluate malignant tendencies and make appropriate treatment decisions. These tools enable a more precise malignancy risk assessment, avoiding overtreatment for low-risk patients while ensuring that high-risk patients receive timely medical intervention.

In recent years, the field of artificial intelligence-based radiomics has developed rapidly, utilising complex mathematical models to process large datasets and uncover patterns unrecognisable via traditional biostatistical methods [42]. For example, Yuan et al. [43] used ultrasound radiomics to analyse the spatial and morphological features of preoperative ultrasound imaging in 99 patients with GBPs, confirming that cholesterol polyps are smaller and more regular in morphology than are gallbladder adenomas, aiding in differentiating true from pseudo polyps. Similarly, Yin et al. [44] proposed a new risk assessment model for the preoperative differentiation of cholesterol and adenomatous GBPs and analysed the CT imaging parameters of 52 patients with polyps. This model served as a prediction model for gallbladder adenomas, showing that arterial phase, portal vein phase and delayed phase CT values and ∆CT values (including ∆CT1 and ∆CT2) can differentiate the nature of gallbladder polypoid lesions, thereby providing objective risk assessment for physicians. These radiomics models based on high-throughput data are less convenient and generalisable for clinical applications, with unclear indicator significance. Elmasry et al. [45] analysed bile viscosity, bile cholesterol, and age as independent risk factors and established a prediction model for GBPs and adenomas, demonstrating good specificity and sensitivity (80.2% and 90.9%, respectively), with an AUC of 0.845. They concluded that bile viscosity, bile cholesterol, and age are important predictors of tumorous polyps, although the invasive nature of the procedures limits their clinical implementation.

However, the current study has some limitations. Firstly, ultrasound is an examiner-dependent procedure, which may introduce certain biases in the results. Secondly, the sample size is small, and broader clinical validation is needed to verify the effectiveness of the prediction model. Future research should include larger sample sizes to supplement and improve the model. Additionally, future studies should utilise emerging ML algorithms combined with radiomics to explore the risk factors for gallbladder adenomas, further enhancing the model accuracy and reliability.