Penile carcinoma (PC) is a rare malignant tumor of the genitourinary system that exhibits a predilection for the glans and inner prepuce (1). The overall incidence of PC in men in developed countries such as Europe and the United States is less than 1 in 100,000, and it accounts for approximately 0.24% of all new cancer cases in men (2, 3). In contrast, PC accounts for 2% of malignancies in men in less developed countries or regions such as Africa and South America, and in some places like Uganda, the incidence is even higher than 8/100,000 (4, 5). This phenomenon could be linked to risk factors such as poor economic conditions, a high prevalence of HPV infection, and phimosis (6, 7). Squamous cell carcinoma accounts for more than 95% of penile cancers and is the most common histologic type (8). The risk of developing squamous cell carcinoma of the penis (SCCP) increases with age, with the peak age range being 50 to 70 years (9). For the surgical treatment of SCCP, the aim of primary tumor treatment is to completely remove the tumor and preserve as many organs as possible without compromising control of the tumor. However, for advanced invasive tumors, partial or radical penectomy is unavoidable (10).

Lymph node invasion of SCCP conforms to the anatomy, and it most commonly metastasizes to the inguinal lymph nodes, followed by the pelvic lymph nodes. Superficial and deep inguinal lymph nodes are the first regional lymph nodes to be invaded, most commonly in the superior medial region, either unilaterally or bilaterally (11). Previous studies have shown that the occurrence of lymph node metastases in SCCP patients is significantly associated with poor prognosis (12). The survival rate of patients with SCCP decreases dramatically with increase in lymph node invasion. The five-year survival rate of patients with inguinal lymph node metastasis (ILNM) is only 50%–80%, but the five-year survival rate of patients who develop pelvic lymph node and peripheral lymph node invasion is less than 33% (13, 14). Due to limitations in imaging techniques, up to 25% of cases in patients with clinically negative lymph nodes cannot be detected despite the presence of micrometastases (15). Therefore, identifying independent risk factors for the development of ILNM at an early stage and establishing corresponding risk prediction models for identifying patients at high risk of ILNM can improve the survival prognosis of this group of SCCP patients through more frequent imaging and more comprehensive clinical treatment in the early stage of diagnosis.

Early studies had shown that stage and grade of the tumor as well as lymphatic and vascular embolization were associated with the risk of lymph node metastasis (16, 17). Ficarra et al. showed that pT stage, histologic grade, venous embolization, and lymphatic embolization were independent predictors of lymph node metastasis in SCCP (18). Peak et al. developed a predictive model based on the clinicopathological characteristics of patients recorded in the National Cancer Database (NCDB), which showed better efficacy in their internal validation (19). In previous studies, several prediction nomograms were established based on logistic regression. The traditional logistic regression algorithm has some limitations in terms of nonlinear complex computation, with area under the curve (AUC) both less than 0.8, so this leaves room for further optimization (20, 21). Machine learning (ML) is an advanced algorithmic model that automatically learns and improves performance by identifying complex nonlinear relationships in different patterns, and is considered superior to traditional algorithms (22). As one of the components of artificial intelligence, ML has been widely used in clinical practice, such as for epidemic prediction (23, 24), and survival analysis (25, 26). ML models have also been proposed for the prediction of lymph node metastasis in a variety of malignancies (27, 28). Based on these promising clinical applications of ML models, the present study aimed to develop and validate a novel ML-based model for predicting the risk of early-stage ILNM in patients with SCCP.

The occurrence of ILNM in SCCP is associated with poor patient prognosis. ILND is the most important procedure for the prevention and treatment of ILNM (29). However, surgeons must consider the balance between the survival benefit to the patient and the high incidence of postoperative complications (30). The use of dynamic anterior node biopsy has been advocated to avoid unnecessary ILND, but the risk of false-negative results remains unavoidable (31), thus threatening the postoperative quality of patient survival. Therefore, early and accurate prediction of LNM is imminent. Previous studies on the SEER database suggest that age at diagnosis, tumor grade, tumor size, T stage, primary site, and LVI may be predictors of ILNM. However, these studies did not take into account the postoperative adjuvant therapy received by the patients. The models based on these variables had AUC of 0.776 and 0.795, respectively (20, 21). The advent of ML algorithms has made it possible to improve model performance.

In this study, we included SCCP patients no distant metastases and non-N3 stage, and developed an accurate predictive model for predicting the risk of early-stage ILNM based on multiple clinical and pathological indicators. Our aim is to screen the SCCP population with a high risk of ILNM for more careful perioperative management of this group of patients, with the ultimate goal of improving patient prognosis. To the best of our knowledge, this is the first model based on a variety of ML algorithms to predict early-stage ILNM with big data and performing external validation. This study included 1,056 cases of SCCP from the SEER database for model establishment and 74 cases from an independent institution in China for external validation. We adopted five ML methods: LR, XGB, RF, SVM, and KNN. By integrating the findings related to the effectiveness and stability of the models in the training and external validation groups, we finally identified XGB as the best prediction model algorithm for early-stage ILNM risk prediction. In the training cohort, the XGB algorithm model (AUC = 0.922, 95% CI = 0.900–0.945) showed superior performance and also performed well in the external validation. The model can be embedded as a web calculator or applied by way of a mobile terminal application. By entering the values or classification results of each indicator, the individual risk of progression can be calculated and, for patients at high risk, more aggressive physical examination, imaging, adjuvant therapy or even surgery should be performed postoperatively.

In both the baseline analysis and the univariate regression, a lower proportion of married than unmarried individuals had early-stage ILNM. This might partly be attributable to the fact that married people tend to have better economic conditions and personal hygiene. In addition, having physical intimacy with a partner increases the likelihood of detecting masses in the groin area and visiting a hospital at an earlier stage of the disease. In our study, the site of the primary tumor did not show a significant association with ILNM risk, but this is not consistent with the findings of previous studies (21). Because delayed intervention can adversely affect the survival of patients with PC diagnosed with lymph node involvement, treatment guidelines recommend lymph node dissection. Therefore, management of regional lymph nodes is very important for patient survival (32). Our findings indicate the importance of ILND for the detection of early-stage ILNM. LVI is defined as infiltration of tumor cells in the lymphatic or hematologic system. Previous studies have shown that LVI is significantly associated with lymph node metastasis and distant metastasis (33). Recently, there is evidence that the presence of LVI is an important risk factor for occult micrometastases in patients with penile cancer and affects the overall survival of patients (34). The correlation of risk with T stage and nuclear grade observed in the present study is in line with previous literature (21, 34). Meanwhile, patients with indications for radiation and chemotherapy, have a higher risk of developing early ILNM, based on clinician judgment.

Our study has certain limitations. First, the number of patients used for external validation at our center is small, and we hope to further improve it at a later stage through multi-center collaboration. Second, ML algorithms are inconvenient for use in clinical practice, and we hope to develop an online calculator app of our XGB model for clinicians in the future. In addition, there are also some limitations to the data in the SEER database. The lack of data on immunohistochemistry, genomics, patient physical indicators, underlying diseases, and hematological indicators reduces the possibility of improving the accuracy of the model, and we hope to remedy this through the establishment of a multicenter database. In addition, the practical value of the model was determined in a predominantly Caucasian database, so its applicability in other regions (including China) is unclear due to the inevitable differences in ethnicity and treatment levels in different countries/regions. Finally, the indications for ILND or adjuvant therapy vary from one medical center to another, so there may be some errors in its practical application. Nevertheless, our study yielded encouraging results that ML algorithms appear to have greater efficacy potential for early ILNM risk prediction in patients with SCCP compared to traditional logistic regression analysis.

We have built a precise, big data-based ML model for predicting early-stage ILNM in patients with SCCP. External validation proved that our novel model has excellent predictive accuracy and clinical utility. Therefore, in the future, it may guide clinicians' decisions and improve the long-term prognosis of patients.