Multimodal AI and tumour microenvironment integration predicts metastasis in cutaneous melanoma
Tom W. Andrew, Marc Combalia, Carlos Hernández, Sydney R. Grant, György Paragh, Susana Puig, Grant A. McArthur, G. O. Richardson, Phil Sloan, Sophia Z. Shalhout, Ruth Plummer, Penny E. Lovat
Abstract
Accurate prognostication is essential to guide clinical management in localised cutaneous melanoma (CM), the form of skin cancer with the highest mortality. While the tumour microenvironment (TME) plays a key role in disease progression, current staging systems rely on limited tumour features and exclude key clinicopathological prognostic features. Here we show that MelanoMAP, a multimodal AI model integrating TME-derived digital biomarkers and clinicopathological features from over 3,500 histology slides, improves prognostication of localised CM. MelanoMAP achieved a C-index of 0.82, a 24% improvement over traditional AJCC staging (0.66) and consistently outperformed clinicopathological-only models across six international patient cohorts. SHAP analysis identified TME-derived digital biomarkers, alongside traditional clinicopathological factors including age, mitotic count, and Breslow depth, were critical determinants of metastatic risk. MelanoMAP establishes a potential foundation for precision oncology in CM, demonstrating how AI-driven digital biomarkers can advance personalised prognostication and inform clinical-decision making.