Causally-Aware Intraoperative Imputation for Overall Survival Time Prediction

Figure 1. Our idea illustration. Patient A and B have different OS time (109 months and 16 months, respectively) but similar pre-operative image-based patterns and indexes; thus, preoperative based model fails to distinguish A and B (their predicted OS time are both larger than 9 years). However, their intra-operative attributes, which can describe the severity of disease in a more informative way, are very different from each other (e.g., G-score, S-score, Clinicopathologic hepatocirrhosis). Therefore, it can help the model discriminate A and B. In the right image, we show that our CAWIM -that leverages intra-operative indexes in the training stage -can largely improve the prediction power than other methods. Our method managed to correctly classify these two cases, and the overall performance of our CAWIM surpasses other baseline models by approximately 10 points on 4-category classification task.