Risk Prediction Based on Clinicopathologic Features in Korean Melanoma Patients by Machine Learning

Background/Objectives The prognosis of adult malignant melanoma differs significantly between Caucasian and non-Caucasian populations due to differences in histologic subtype prevalence. Accurate risk stratification at the time of diagnosis is essential to make treatment decisions and predict clinical outcomes. However, due to the low incidence of melanoma in Korea, pre-cise risk stratification remains challenging. Methods The study involved a retrospective cohort of patients with malignant melanoma at Sever-ance Hospital between January 2006 and October 2023. Patients diagnosed with cancers other than melanoma and with insufficient clinical information for analysis were ex-cluded. Only clinical and pathological variables available at the time of diagnosis were included as features. Machine learning models (decision tree, random forest, bagging, AdaBoost, GradientBoost, and XGBoost) were applied to predict 2-year time point dis-ease-free survival (DFS) status. Model performance was evaluated using accuracy, preci-sion, recall, and F1 score. Results The study included 1,657 patients with 12 clinical features, showing median DFS of 63.93 months (95% CI 52.93-85.33) and a 2-year survival rate of 67.68% (95% CI: 65.16-70.00). Among the models evaluated, XGBoost demonstrated the highest F1 score of 0.761. Pri-mary cancer site, mitotic rate, and body mass index (BMI) were identified as important features predicting survival status. Kaplan-Meier curves and Cox proportional hazards analysis confirmed the clinical relevance of these features. This model suggests that these features are closely related to the DFS of Korean malignant melanoma patients. Conclusion The XGBoost model was the most effective for 2-year DFS probability prediction in Korean melanoma patients, and primary cancer site, mitotic rate, and BMI were identified as sig-nificant features influencing survival.