Machine learning approach to identify HDL as a prognostic factor for melanoma immunotherapy.

Abstract

e21531 Background: Melanoma immunotherapy faces challenges due to heterogeneous patient responses. Current predictive biomarkers are not available in clinical practice. Serum lipids, crucial in tumor biology, show prognostic potential but lack consistent validation. To develop and validate a machine learning (ML) prognostic model integrating systemic lipid metabolism factors for predicting overall survival in melanoma patients receiving immunotherapy. Methods: A retrospective study of 381 melanoma patients was conducted. LASSO and Cox regression analyses identified prognostic features. Eight ML algorithms were trained and validated on a 7:3 split dataset to build the model. The best model was deployed as a publicly accessible application. We conducted SHAP analysis at 1-year, 3-year, and 5-year survival timepoints to generate variable importance rankings for each time node. To further investigate the key populations benefiting from the protective factor and explore whether this protection might be mediated through lipid regulation, we performed subgroup and interaction analyses using previously established thresholds. Results: The LASSO-Cox model demonstrated superior performance. Multivariable analysis identified mucosal subtype, advanced AJCC stage, high Lactate Dehydrogenase, and Ki67 index as risk factors, while high high-density lipoprotein cholesterol (HDL-C) was a protective factor. SHAP analysis ranked HDL-C as the most important predictive feature. Subgroup analysis revealed a more pronounced protective effect of high HDL-C in patients with hyperlipidemia. Finally, the model was deployed as a web (https://melanomaslnmodel.shinyapps.io/Shiny/) application to facilitate its potential clinical utility. Conclusions: HDL-C is a significant prognostic factor in melanoma immunotherapy. The web-based model provides an accurate risk prediction tool, highlighting the importance of monitoring lipid profiles in patient management.