A Multi-modal Graph Convolutional Network for Predicting Human Breast Cancer Prognosis

Abstract

Breast cancer is one of the most often found malignancies in women. For more focused treatment and disease management, a better prognosis for breast cancer is crucial. If breast cancer prognosis predictions were correct, a substantial number of people may be spared from unnecessary adjuvant systemic treatment and the enormous medical costs. Several studies have already been conducted to accomplish this. But, most studies employ specific gene expression data to create a predictive model. However, multi-modal cancer data sets have become accessible recently (gene expression, copy number alteration, and clinical). The introduction of multi-modal data presents possibilities for a more thorough investigation of the molecular aspects of breast cancer and, consequently, can enhance diagnosis. To incorporate multi-modal cancer data sets and to create a computational model for the prognosis of breast cancer, we proposed a novel classification model in this study, that is based on multi-modal graph convolutional networks (MGCN). To extract features, we first build a graph convolutional network (GCN) for individual modalities. And then, we feed the concatenated features generated by GCN into the stack-based ensemble model. The GCN model explores the underlying non-regular structural information from the data and learns the nodes’ (or samples’) hidden representation based on its properties and those of its surrounding nodes. This model outperforms currently used methods, according to the predictive performance assessed using various performance indicators. The precision, balanced accuracy, and Matthew’s correlation coefficient values produced by this model are 0.869, 0.740, and 0.498, respectively.