Machine Learning Techniques in Predicting BRAF Mutation Status in Cutaneous Melanoma From Clinical and Histopathologic Features

Melanoma is the cutaneous neoplasm responsible for more patient deaths in all countries. BRAF mutations are the most common driver mutation and with the development of molecular targeted therapy, the precise knowledge of BRAF status has become increasingly important. Evaluation of BRAF mutation status has routinely been performed by polymerase chain reaction, a time consuming and expensive technique. Immunohistochemistry has been suggested as a cheaper alternative, but it has not gained general acceptance. A retrospective observational study in a cohort of 106 patients with invasive melanoma was conducted in order to develop and evaluate a machine learning approach to predict BRAF status using clinical and histologic variables. We compared the performance of different common machine learning algorithms and use SHapley Additive exPlanations (SHAP) to explain individual predictions and extract medical insights to define a heuristic model to estimate BRAF mutation probability. The Extreme Gradient Boosting algorithms obtained the best performance. Interpretability of models shows that the most important variables to estimate BRAF mutation probability are: age, Breslow thickness, and Breslow density. Based in this interpretation and medical knowledge, a simplify heuristic model is proposed to predict BRAF status using only 7 variables and obtain a performance of 0.878 of area under the curve. We propose a heuristic model that could be used by clinicians to obtain a good estimator of BRAF mutation probability.