Integrative prognostic modeling for breast cancer: Unveiling optimal multimodal combinations using graph convolutional networks and calibrated random forest

The most crucial step in the clinical decision-making process for patients with breast cancer is the accurate prediction of prognosis and survival length. Correct prognosis prediction aids in making the best treatment decision and may even lessen the effects of cancer. To achieve this, we have developed a novel predictive model using multimodal graph convolutional networks with calibrated random forest (MGCN-CalRF) for forecasting the prognosis of breast cancer by combining multiple sources of information or modalities. We have considered six different modalities, namely, mRNASeq, DNA methylation, Copy number variation, miRSeq, Clinical, and Whole Slide Image data, which have been retrieved from TCGA Database. We have applied a Graph Convolutional Network for the feature extraction from individual modalities to grasp the structural relationships between data. Further, we concatenated all the extracted features concerning various combinations of modalities aiming to find the optimal combination of available modalities. The concatenated features from the Graph Convolutional Network are further fed to the Calibrated Classifier Model using Random Forest for the final prediction. It has been observed that the Graph Convolutional model which has been trained with the combination of three modalities, namely Clinical, miRSeq, and Whole slide Image data outperforms not only the combination of other modalities but also the other state-of-the-art models. This model attained accuracy, F1-score, and AUC of 0.771, 0.867, and 0.729, respectively.