Combining machine learning with the immunohistochemical expression of AMBRA1 and loricrin to identify non-ulcerated AJCC stage I/II melanomas at high-risk of metastasis.

9570 Background: Precision-based personalised biomarkers able to identify both low-risk and high-risk patient subpopulations with localised cutaneous melanoma are urgently needed to guide clinical follow up and treatment stratification. The combined immunohistochemical expression of AMBRA1 and Loricrin (AMBLor) in the epidermis overlying non-ulcerated AJCC stage I/II melanomas as prognostic biomarker able to accurately identify genuinely low-risk patient subpopulations (NPV >96%, clinical sensitivity >95%, Ewen et al Brit J Dermatol. 2024). To further identify distinct subsets of patients at high risk of metastasis, the present study aimed to develop a machine learning (ML) risk-prediction model combining AMBLor ‘at-risk’ status with six specific patient clinical and tumour pathological features. Methods: Using common and widely used ML models a Naïve Bayes and a Generalized Linear Model with adaBoost, ML algorithms were trained and tested using three geographically distinct retrospective-prospective cohorts of AMBLor at-risk non-ulcerated AJCC stage I/II melanomas from Australia, USA and Spain (n=552), with validation studies performed in a 4 th independent retrospective-prospective cohort of 120 AMBLor at-risk non-ulcerated localised melanomas derived from the UK. Results: Based on a training: test data split of 50:50, 20% of patients were defined as high-risk, with a 5-year recurrence-free survival (RFS) probability of 56% (Log-rank [Mantel-Cox) P < 0.0001, HR 6.88, 95% CI 3.03-15.63, clinical specificity 87.2%, PPV 44.4%). Further validation of the ML algorithms in the UK validation cohort identified 24% patients as high-risk, with a 5-year RFS of 56.3% (Log-rank [Mantel-Cox) P < 0.0001, HR 7.59, 95% CI 2.94-19.6, clinical specificity 82.1%, PPV 50%). Conclusions: Through the proven negative predictive power of AMBLor with the cumulative power of prognostic clinical and pathological features these data provide a novel and improved risk- prediction model to stratify patients with non-ulcerated localised melanomas at low or high risk of tumour recurrence thereby aiding optimal personalised patient management and treatment stratification.