Synergies Between Case-Based Reasoning and Deep Learning for Survival Analysis in Oncology

Survival analysis is a field of statistics specialized in making predictions about the survival length of patients, even though it can be applied to the prediction of any future event. It is routinely used in medical research to stratify patients in groups based on risk, such as high-risk groups and low-risk groups, and has paramount important in patient stratification and treatment. Recently, deep neural networks (DNNs) have raised considerable attention for survival analysis because of their non-linear nature and their excellent ability to predict survival, in comparison to statistical methods. In this domain, case-based survival methods have started to by applied as well, with some success. It is therefore interesting to study how to synergistically combine the two for improved performance for several reasons. From the case-based reasoning standpoint, the deep neural network can detect deep similarity between cases with a time-to-event structure and from the DNN standpoint, case-based reasoning can provide the glass-box approach that remedies the "black box" label attached to them. In this study, we propose a synergy between case-based reasoning and Long Short-Term Memory (LSTM) model for survival prediction in oncology. In this deep survival model network, the total loss function combines four different factors and uses an adaptive weights approach to combine the four loss terms. The network learns a prototype layer during training which naturally comes with an explanation for each prediction. This study employs cross-validation and the concordance index for assessing the survival prediction performance and demonstrate on two cancer methylation data sets that the developed approach is effective.