Multimodal approach predicts relapse upon cessation of immune checkpoint inhibitors in advanced melanoma

Abstract

Purpose: Treatment with immune checkpoint inhibitors (ICI) in advanced melanoma can result in durable responses, yet an algorithm to decide which patients can safely discontinue ICI is still lacking. Experimental Design: We used a multimodal approach combining clinical data, AI-based analysis of H&E-stained whole-slide images of melanoma before ICI start, and gene expression signatures to identify biomarkers for relapse after discontinuing ICI in the absence of treatment progression. Results: Univariable Cox regression analysis identified best overall response, mRNA expression of six genes, tumor cell density (TCD), and the lymphocyte to plasma cell ratio (LYM/PC) as factors predictive of relapse upon cessation of ICI. Multivariable Cox regression analysis showed that both TGFBR1 expression and the integral digital pathology parameter-based prognostic system were independently associated with relapse after ICI discontinuation. Training a Multivariate Adaptive Regression Spline (MARS) model achieved the highest overall predictive accuracy of 84.6% for relapse after ICI discontinuation. Conclusions: The identified prognostic markers are fully explainable and easily implementable in routine practice and facilitate risk stratification upon cessation of ICI therapy.