Machine learning models demonstrate that clinicopathologic variables are comparable to gene expression prognostic signature in predicting survival in uveal melanoma

Purpose: Since molecular assays are not accessible to all uveal melanoma patients, we aim to identify cost-effective prognostic tool in risk stratification using machine learning models based on routine histologic and clinical variables. Experimental design: We identified important prognostic parameters in a discovery cohort of 164 enucleated primary uveal melanomas from 164 patients without prior therapies. We then validated the prognostic prediction of top important parameters identified in the discovery cohort using 80 uveal melanomas from the Tumor Cancer Genome Atlas database with avail-able gene expression prognostic signature (GEPS). The performance of three different survival analysis models (Cox proportional hazards (CPH), random survival forest (RSF), and survival gradient boosting (SGB)) was compared against GEPS using receiver operating curves (ROC). Results: In all three selection methods, BAP1 status, nucleoli size, age, mitotic rate per 1 mm2, and ciliary body infiltration were identified as significant overall survival (OS) predictors; and BAP1 status, nucleoli size, largest basal tumor diameter, tumor-infiltrating lymphocyte den-sity, and tumor-associated macrophage density were identified as significant progression -free survival (PFS) predictors. ROC plots for the median survival time point showed that sig-nificant parameters in SGB studied model can predict OS better than GEPS. For PFS, SGB model performed similarly to GEPS. The time-dependent area under the curve (AUC) showed SGB model performing better than GEPS in predicting OS and metastatic risk. Conclusions: Our study shows that routine histologic and clinical variables are adequate for patient risk stratification in comparison with not readily accessible GEPS. 2022 Elsevier Ltd. All rights reserved.

Automated summary

The study aims to identify a cost-effective prognostic tool for risk stratification in uveal melanoma patients using machine learning models based on routine histologic and clinical variables. The results show that routine histologic and clinical variables are sufficient for patient risk stratification, without the need for molecular assays.