Predicting BRAF Mutations in Cutaneous Melanoma Patients Using Neural Network Analysis

Abstract

Point mutations at codon 600 of the BRAF oncogene are the most common alterations in cutaneous melanoma (CM). Assessment of BRAF status allows to personalize patient management, though the affordability of molecular testing is limited in some countries. This study aimed to develop a model for predicting alteration in BRAF based on routinely available clinical and histological data. Methods: For identifying the key factors associated with point mutations in BRAF, 2041 patients with CM were recruited in the study. The presence of BRAF mutations was an endpoint. The variables included demographic data (gender and age), anatomic location, stage, histological subtype, number of mitosis, and also such features as ulceration, Clark level, Breslow thickness, infiltration by lymphocytes, invasiveness, regression, microsatellites, and association with nevi. Results: A relatively high rate of BRAF mutation was revealed in the Ukrainian cohort of patients with CM. BRAF-mutant melanoma was associated with younger age and location of nonsun-exposed skin. Besides, sex-specific differences were found between CM of various anatomic distributions and the frequency of distinct BRAF mutation subtypes. A minimal set of variables linked to BRAF mutations, defined by the genetic input selection algorithm, included patient age, primary tumor location, histological type, lymphovascular invasion, ulceration, and association with nevi. To encounter nonlinear links, neural network modeling was applied resulting in a multilayer perceptron (MLP) with one hidden layer. Its architecture included four neurons with a logistic activation function. The AUROCMLP6 of the MLP model comprised 0.79 (95% CІ: 0.74-0.84). Under the optimal threshold, the model demonstrated the following parameters: sensitivity: 89.4% (95% CІ: 84.5%-93.1%), specificity: 50.7% (95% CІ: 42.2%-59.1%), positive predictive value: 73.1% (95% CІ: 69.6%-76.3%), and negative predictive value: 76.0% (95% CІ: 67.6%-82.8%). The developed MLP model enables the prediction of the mutation in BRAF oncogene in CM, alleviating decisions on personalized management of patients with CM. In conclusion, the developed MLP model, which relies on the assessment of 6 variables, can predict the BRAF mutation status in patients with CM, supporting decisions on patient management.

Keywords: BRAF mutation; cutaneous melanoma; multilayered perceptron; predictive model.

Figure 1

The performance characteristics of the 6-factorial MLP model for predicting BRAF mutations in CM samples compared to those of the logistic regression model. Characteristics of the MLP model compared to those of the logistic regression model with a difference between areas of Δ = 0.11 (95% CI: 0.04–0.17), p=0.001. The sensitivity of the MLP model was 89.4% (95% CІ: 84.5%–93.1%), the specificity was 50.7% (95% CІ: 42.2%–59.1%), the positive predictive value (PPV) was 73.1% (95% CІ: 69.6%–76.3%), and the NPV was 76.0% (95% CІ: 67.6%–82.8%).