Adrenal tumors (ATs) are a rare type of tumor that usually occurs in the cortex or medulla of the adrenal gland (1). Depending on their type and size, these tumors can be benign or malignant (2). ATs can cause many symptoms, including high blood pressure, palpitations, headaches, insomnia, anxiety, and obesity (3). In some cases, these symptoms may be mistaken for symptoms of other diseases, so further testing is needed to determine the diagnosis (4–6).

Treatment for AT includes surgery, radiation therapy, and chemotherapy. Surgery is the most common treatment method and can altogether remove the tumor (7). The gold standard treatment for AT is laparoscopic surgery, which can be divided into two main approaches: transperitoneal laparoscopic adrenalectomy (TPLA) and retroperitoneal laparoscopic adrenalectomy (RPLA) (8). The RPLA involves entering the retroperitoneal cavity through laparoscopic surgery, avoiding interference with abdominal organs, and reducing surgical trauma and recovery time. Compared with traditional open surgery, this technique has fewer complications and faster recovery (6, 9, 10).

In the field of medicine, machine learning (ML) has wide-ranging applications (3). For example, ML can be used for medical image recognition to help doctors diagnose diseases. It can also be used to predict the health status of patients, assisting doctors to develop better treatment plans. In addition, ML can be used for drug development and clinical trials to speed up the development and launch of new drugs (11). For example, a study has used ML to differentiate between adrenal pheochromocytoma and adrenocortical adenoma (12).

Radiomics is an emerging field of medicine that combines computer science, mathematics, and medical imaging to understand better and diagnose diseases (13, 14). Radiomics analyzes large amounts of medical imaging data to extract useful information, helping doctors make more accurate diagnoses and treatment decisions (15).

This study aimed to collect data retrospectively from patients with AT who underwent RPLA at Shanxi Bethune Hospital from August 2014 to December 2020. The study utilized ML to analyze their clinical and radiomics features and develop a predictive model for the difficulty of RPLA. The goal was to improve preoperative preparation, reduce surgical risks, and enhance patient benefits.

AT has become a hot topic in the medical field, and surgery is the primary treatment method. TPLA and RPLA were proposed in 1992 (8, 23), respectively, and have continuously improved. The advantage of RPLA is that it results in less surgical trauma and bleeding, faster recovery, and fewer complications. Moreover, it is also suitable for some cases that traditional surgery finds challenging, such as obesity and complex AT. There are some relatively objective analysis systems for the surgical difficulty of TPLA, while there is less analysis on RPLA (10, 17).

This study retrospectively analyzed 396 patients who underwent RPLA for AT. The LASSO analysis of radiomics features was used to calculate the Rad-score. By combining the Rad-score with preoperative clinical characteristics, ML models such as CART, KNN, LASSO, NB, RF, SVM, and xGBoost were constructed and compared. It was found that RF had a more stable prediction accuracy, while xGBoost could bring more significant benefits to patients. The ML model suggested that in addition to the most influential Rad-score, the clinical characteristics such as Hb, age, BMI, gender, and diabetes mellitus also greatly influenced surgical difficulty. Through the validation of the validation set and prospective set, it was found that the ML models had high predictive ability. Through the comprehensive comparison of different models, it was found that the RF model exhibits the best prediction performance, thus making it our recommended model. Furthermore, in comparison to clinical models in previous study (16), our RF model exhibited superiority as evidenced by 2000 Bootstrap tests (D = 7.155, P < 0.001). The discrimination power of models can be effectively compared using two measures: the Net Classification Index (NRI) and the Integrated Discrimination Improvement (IDI). In comparison to previous studies, the RF model in this study demonstrated an NRI of 0.308 (95% CI: 0.194-0.422, p < 0.001) and an IDI of 0.165 (95% CI: 0.119-0.210, p < 0.001).

The Rad-score calculated based on LASSO significantly impacts the surgical difficulty of RPLA. When performing univariate logistic regression, 279 features were statistically significant. After LASSO, 18 variables were retained and used to construct the Rad-score. The final retained variables included “Shape Features” like “Maximum3DDiameter”. Moreover, many studies have generally confirmed that the maximum diameter of the tumor is an essential factor affecting the difficulty of removing AT (9, 10, 16–20). In addition, “First Order Features”, which are linearly correlated with the CT value of the tumor, such as “90Percentile” were also included. Malignant and benign AT have different degrees of enhancement during arterial enhancement, which increases the risk of bleeding during surgery (18, 24, 25). It may also be because lipid-rich AT has lower CT values and requires more attention during surgery to prevent breaking the capsule, which prolongs the operation time (9, 16, 20). “DifferenceEntropy” in “GLCM Features” measures the randomness or complexity of differences between pixel intensity values. It was included because malignant tumors, such as metastases, exhibit more randomness or complexity between pixel intensity values, while their removal is more challenging than benign tumors (18).

Some clinical characteristics of patients also affect the difficulty of RPLA. Patients with diabetes mellitus are more likely to have perirenal fat adhesions, which affect surgical difficulty (26). Studies by Chen (17) and Takeda (27) have also shown that diabetes mellitus significantly affects it. Some studies also suggest that a history of hypertension and coronary heart disease affects surgical difficulty (28, 29). BMI is used to assess the degree of obesity and also affects it. However, it mainly reflects the overall body fat composition, while the distribution of visceral fat, especially perirenal fat, may differ (9, 16). Therefore, there is still controversy over BMI prediction of surgical difficulty. Some studies believe that measuring visceral fat would be more accurate (10, 25, 29).

Hb reflects the patient’s blood reserve and blood oxygen reserve situation (30). If it is too low, it will affect the surgery. Age affects almost all tumor surgeries and prognoses because older patients often have poorer nutrition and tolerance. Moreover, diseases tend to be more malignant in older patients (31, 32). In addition, some researchers believe that males may have more dangerous lifestyles (such as smoking), and there are differences in hormone levels between men and women, which may lead to poorer physical conditions and more incredible surgical difficulty in male patients (33, 34).

This study established ML models for predicting the difficulty of RPLA based on preoperative radiomics and clinical characteristics. It was validated internally and prospectively to prove that the ML models can significantly improve patients’ net benefit rate.

There needs to be more accurate prediction models for the difficulty of RPLA. The innovation of this study lies in combining ML with radiomics to analyze the risk factors for the difficulty of RPLA and establish prediction models for it, then conduct internal validation and prospective validation to make the model more meaningful. Moreover, this study is currently one of the largest cohorts using radiomics to predict the difficulty of RPLA.

The prospects of this study include: external validation to confirm its stability and accuracy further; using radiomics to analyze the tumor’s surrounding environment while analyzing AT and optimizing the model through more ML algorithms. Some studies have proposed that magnetic resonance imaging has multiple weighted sequences, which may have better effects when applied to radiomics than CT. Although the accuracy of this study’s models is high, the time cost of drawing ROI is high. If further promotion or clinical transformation is needed, combining deep learning to train artificial intelligence to draw ROI is necessary. Some studies have successfully trained artificial intelligence to draw ROIs for pancreatic duct tumors and predicted lymph node metastasis and prognosis based on them. Its sensitivity and specificity are superior to clinical and radiomics models (35).

In conclusion, Rad-score, Hb, age, BMI, gender, and diabetes mellitus affect RPLA surgical difficulty. The ML prediction model established based on patient clinical characteristics and Rad-score using RF and xGBoost has good predictive performance. Through the above model, surgeons can effectively evaluate the difficulty of RPLA, thereby reducing surgical risks and improving patient benefits.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving humans were approved by Ethics Committee of the Shanxi Bethune Hospital. The studies were conducted in accordance with the local legislation and institutional requirements. Written informed consent for participation was not required from the participants or the participants’ legal guardians/next of kin in accordance with the national legislation and institutional requirements.

SS: Conceptualization, Data curation, Methodology, Project administration, Validation, Visualization, Writing – original draft. WY: Data curation, Validation, Visualization, Writing – original draft. YW: Data curation, Validation, Writing – original draft. PY: Data curation, Writing – original draft. FG: Data curation, Writing – original draft. XD: Data curation, Writing – original draft. YZ: Conceptualization, Methodology, Project administration, Supervision, Writing – review & editing.