Comparison of an AI‐based prognostic assay with gene expression profiling in early‐stage melanoma

Abstract

We thank Magnaterra et al.1 for their constructive remarks and their interest in our work. Their comments highlight key aspects of interpretability, positioning and clinical implications, which are central to the adoption of AI tools in melanoma prognostication. We would like to expand on these points with some comments and additional data from a recent validation study. We recognize that reclassifying subsets of early-stage (I-II) melanomas is challenging but crucial to propose the most appropriate strategy of adjuvant treatment (with potentially serious side effects) or follow-up. In our previous study,2 Kaplan–Meier curves based on DiaSurv Melanoma assay demonstrated a significant difference in 5-year overall survival between low-risk (92%) and high-risk (62%) patients. Similar results were confirmed in a larger external independent retrospective cohort,3 supporting the robustness and generalizability of the DiaSurv assay in prognostic stratification of early-stage melanoma patients, independently of pre-analytic WSI processing. These results confirm that early-stage patients represent a more heterogeneous group than conventional pTNM staging delineates. Concerning explainability, Magnaterra et al. suggest a combined morphology-immunohistochemical score, including Ki67 marker, that could enhance interpretability and facilitate clinical adoption. Indeed, Ki67 expression was recently shown to stratify acral melanoma patients into different prognostic groups for DFS and OS, and was an independent prognostic factor in multivariate analysis.4 Our approach was entirely unsupervised and based only on HE-stained WSI to propose a simple, rapid and more precise prognostic score requiring no further technique. We identified morphological patterns associated with risk predictions that make biological sense. These aspects need to be confirmed in ongoing and future studies, and we are not sure that Ki67 expression alone could explain the underlying biological prognostic contextures of tumours. This point of explainability is often raised for AI to justify caution in clinical use. However, molecular signatures that are regularly used by clinicians could benefit from the same criticisms in terms of interpretability and “black box” phenomenon. In our view, true biological explainability requires a deeper multimodal approach based on multi-omics data. Beyond these important considerations, another interesting issue concerns the comparison of performances between these two different approaches, that is AI-based pathology vs. molecular assays, in terms of prognostic stratification. We evaluated the performances of our algorithm, DiaSurv Melanoma assay, against a commercially available gene-expression profiling (GEP) assay on a retrospective cohort of 50 early-stage melanomas (Table 1). The two approaches showed 74% overall concordance in risk stratification. For overall survival (OS), deaths occurred predominantly in the high-risk group defined by DiaSurv (5/7; 72%) versus the high-risk group defined by the GEP assay (2/7; 29%); corresponding PPV/NPV were 26%/94% for DiaSurv and 20%/88% for GEP. For relapse-free survival (RFS), DiaSurv assigned all relapses (4/4) to the high-risk group, whereas the GEP assay assigned 2/4 to high risk; PPV/NPV were 21%/100% for DiaSurv and 20%/95% for GEP. Kaplan–Meier curves (Figure 1) based on DiaSurv differentiates more significantly the two groups of patients, with a significant difference for RFS (log-rank p < 0.05), while differences were non-significant for the GEP assay. Despite the limited size of the dataset, these results are very promising and need to be confirmed and validated on larger cohorts and with different GEP assays, as suggested by Magnaterra et al. We believe that AI-based tool applied to WSI can offer robust, rapid and clinically meaningful prognostic insights while improving interpretability to discover potential new biological targeted pathways, a prerequisite for responsible integration into melanoma care pathways. None. C.B., Y.S. S.S. and J.J.C. hold shares in DiaDeep. The remaining authors declare that they have no conflicts of interest related to this study. The RicMel database (Clinical Trials n°. NCT03315468) gathers data from 49 participating centres in different French regions. It received ethics committee approval on 9 February 2012 (no. 12.108) from the Independent Ethics Committee in Paris and received authorization from the French Data Protection Agency (CNIL, DR-2012-259, 28 May 2012). Data on patients with melanoma were obtained from the RicMel database, a French national multicentric database dedicated to the follow-up of MM patients. All patients provided informed consent for their data to be included in the registry and used for research purposes. The data that support the findings of this study are available from the corresponding author upon reasonable request.

Ähnliche Arbeiten