Development and validation of machine learning-based models for identifying synchronous distant metastasis at initial diagnosis in cutaneous melanoma using the SEER database

Accurate identification of synchronous distant metastasis at initial diagnosis is critical for prognosis and treatment guidance in cutaneous melanoma, yet current staging systems have limited accuracy for individualized risk prediction. This study aimed to develop a machine learning (ML)-based prediction model to enhance the accuracy of distant metastasis risk assessment, thereby improving risk stratification and enabling early intervention strategies.

A total of 76,514 patients with cutaneous melanoma were identified from the Surveillance, Epidemiology, and End Results (SEER) database. Patients with unknown data on key variables or with survival less than 1 month were excluded according to predefined inclusion/exclusion criteria. The study outcome was defined as distant metastasis (M1 stage) at diagnosis, as recorded in SEER. Candidate predictors encompassed demographic, clinicopathological, and treatment variables. The cohort was randomly split into training (70%) and validation (30%) sets. Univariate and multivariate logistic regression analyses, combined with least absolute shrinkage and selection operator (LASSO) regression, were employed to identify six key predictive variables. Ten ML algorithms were constructed and evaluated, including logistic regression, random forest (RF), support vector machine (SVM), gradient boosting machine (GBM), extreme gradient boosting (XGBoost), light GBM (LightGBM), categorical boosting (CatBoost), K-nearest neighbors (KNN), naive Bayes (NB), and neural network (NN). Model interpretability was achieved using SHapley Additive exPlanations (SHAP) analysis. The optimal model was subsequently utilized to develop a web-based calculator for identifying distant metastasis at initial presentation in cutaneous melanoma patients.

Distant metastasis was present in 3,256 patients (4.3%). Among 10 ML models, GBM achieved the best performance, achieving an area under the receiver operating characteristic (ROC) curve (AUC) of 0.977 [95% confidence interval (CI): 0.974-0.980] in the training set and 0.970 (95% CI: 0.964-0.976) in the test set. Key predictors included histologic type, T stage, N stage, surgical treatment, radiation therapy, and chemotherapy. External validation confirmed the robustness of the model with an AUC of 0.972. Based on the GBM model, a user-friendly web-based calculator was developed to facilitate clinical application.

This study successfully developed and validated a GBM-based predictive model with high accuracy for assessing distant metastasis risk in cutaneous melanoma, offering a robust tool for individualized risk stratification. The accompanying web-based calculator facilitates clinical translation, enabling timely identification of high-risk patients with synchronous metastatic disease to guide appropriate initial staging workup and treatment planning. Although the model performed well in internal validation, further multi-center prospective studies are warranted to confirm its generalizability before widespread clinical adoption.