Development of a Machine Learning Model for Distant Metastasis Risk Stratification in Acral Melanoma.

backgroundAcral melanoma (AM) is a distinct melanoma subtype associated with delayed diagnosis, aggressive progression, and poor prognosis once distant metastasis occurs. However, prediction models specifically designed for distant metastasis risk stratification in AM remain limited.

aimsThis study aimed to develop and internally evaluate a machine learning-based model for individualized distant metastasis risk stratification in patients with AM. METHODS AND

resultsClinical data of 1822 patients with AM diagnosed between 2000 and 2021 were extracted from the SEER database. Patients were divided into training and internal test sets at a ratio of 7:3 using stratified sampling. Logistic regression analyses were performed to identify factors associated with distant metastasis, and six machine learning algorithms were developed and compared. SMOTE was applied only to the training set to address class imbalance. Multivariate logistic regression identified sentinel lymph node biopsy as an independent protective factor, whereas higher N stage and lower median household income were independent risk factors. Among the evaluated models, LightGBM showed relatively balanced overall performance and was selected as the optimal model. SHAP analysis identified N stage, sentinel lymph node biopsy record, and median household income as the most important predictors.

conclusionThe LightGBM model demonstrated moderate predictive performance for distant metastasis risk stratification in patients with AM. This model may serve as a research-oriented tool for individualized risk assessment, although external validation using independent real-world cohorts is required before clinical application.