High throughput biomedical data has evolved into an integral technology for biomedical discoveries. Recent biotechnological advances have introduced new spatial resolutions into single-cell experiments, allowing for a deeper knowledge of cellular processes and responses. Spatial transcriptomics quantifies gene expression at distinct spatial locations in tissues and can detect tumour heterogeneity, which is relevant to patient survival. However, the cost of such assays prevents its adoption into routine clinical use at this stage. Recently there is a number of deep-learning based approaches that use matched histology and spatial transcriptomics to build a model for predicting gene expression. In this study, we investigated the value of incorporating spatial gene expression with histological criteria that is traditionally used in survival prediction, and we compared the relative prediction performance between spatial transcriptomics and bulk RNA-seq. We conducted this work in two steps: (1) we investigated recent gene expression prediction networks (e.g., ST-Net) trained on spatial transcriptomics datasets to obtain spatial gene expression; and (2) we developed a novel deep-learning based model that extracts and fuses features from histology and gene expression data to predict survival. We compared the survival prediction performance using cancer datasets from The Cancer Genome Atlas, we investigated the survival prediction performance using the same main architecture of the survival prediction model with two sets of inputs derived (i) histology and bulk RNA-seq; and (ii) histology and the predicted spatial gene expression.