Clinical validation of an AI-based prognostic assay for stage I-II cutaneous melanoma.

Abstract

9578 Background: In early-stage cutaneous melanoma (CM), AJCC staging demonstrates limited ability to identify patients at elevated recurrence risk who may benefit from adjuvant therapy or intensified surveillance. We evaluated DiaSurv, an AI-based prognostic assay analyzing standard H&E whole-slide images (WSI) to refine risk stratification in stage I-II melanoma. Methods: DiaSurv was validated on a retrospective cohort of 456 stage I-II primary CM patients with 5-year follow-up at Gustave Roussy. The algorithm assigns high- or low-risk scores based solely on primary tumor WSI analysis. Kaplan-Meier analysis estimated 5-year recurrence-free survival (RFS), distant metastasis-free survival (DMFS), and overall survival (OS). Log-rank tests compared survival between risk groups. Multivariable Cox regression assessed independent prognostic value after adjusting for clinicopathological factors (Breslow thickness, ulceration, age, sex, tumor location, histological subtype, mitotic count). Results: Among 456 stage I-II patients (109 recurrences, 67 distant metastases, 57 deaths), DiaSurv classified 285 (62.5%) as low-risk and 171 (37.5%) as high-risk. Low-risk patients had significantly better 5-year outcomes compared to high-risk patients: RFS 92% vs 38% (HR = 7.71, 95%CI 4.86-12.23, p < 0.001), DMFS 94% vs 55% (HR = 10.39, 95%CI 5.31-20.36, p < 0.001), and OS 96% vs 71% (HR = 8.12, 95%CI 4.10-16.07, p < 0.001). In multivariable analysis, DiaSurv was independently associated with all endpoints (RFS: HR = 2.43, p = 0.006; DMFS: HR = 3.71, p = 0.002; OS: HR = 3.78, p = 0.003). In the clinically relevant stage I-IIA subgroup (n = 368), DiaSurv identified 86 patients (23.4%) with high-risk profiles despite favorable AJCC staging. These patients had markedly inferior outcomes compared to the 282 low-risk patients: 5-year RFS 71% vs 92% (HR = 4.85, p < 0.001), DMFS 77% vs 96% (HR = 5.48, p < 0.001), and OS 84% vs 96% (HR = 4.39, p < 0.001). The NPV of low-risk classification at 5 years in stage I-IIA was 92.2% for RFS, 96.5% for DMFS, and 96.5% for OS which is in line or superior to some commercially available gene-expression profiling assays. Conclusions: In early-stage melanoma (I-IIA), where adjuvant therapy is not currently standard, DiaSurv identifies nearly one-quarter of patients with a high-risk profile and significantly elevated recurrence rates. This AI-based assay provides actionable prognostic information beyond AJCC staging to guide decisions regarding adjuvant therapy consideration, sentinel lymph node biopsy, and surveillance intensity.