Regularizing the Deepsurv Network Using Projection Loss for Medical Risk Assessment

Abstract

State-of-the-art deep survival prediction approaches expand network parameters to accommodate performance over a fine discretization of output time. For medical applications where data are limited, the regression-based Deepsurv approach is more advantageous because its continuous output design limits unnecessary network parameters. Despite the practical advantage, the typical network lacks control over the feature distribution causing the network to be more prone to noisy information and occasional poor prediction performance. We propose a novel projection loss as a regularizing objective to improve the timeto-event Deepsurv model. The loss formulation maximizes the lower bound of the multiple-correlation coefficient between the network's features and the desired hazard value. Reducing the loss also theoretically lowers the upper bound on the likelihood of discordant pair and improves C-index performance. We observe superior performances and robustness of regularized Deepsurv over many state-of-the-art approaches in our experiments with five public medical datasets and two cross-cohort validation tasks.

MRI-based radiomic prognostic signature for locally advanced oral cavity squamous cell carcinoma: development, testing and comparison with genomic prognostic signatures

“…Parallel to the radiomic approach proposed herein, end-to-end radiomic-based deep learning models have been proposed (e.g., [ 71 , 72 ]) and used to predict OS in HNSCC patients [ 73 ]. However, differently from the deep features, radiomic features can be linked to tissue properties, as the shape, size and texture, thus allowing for an easier interpretability of the model.…”

Improving Adjuvant Liver-Directed Treatment Recommendations for Unresectable Hepatocellular Carcinoma: An Artificial Intelligence–Based Decision-Making Tool

“…Pairwise comparison groups included the following: no additional treatment versus SBRT, no additional treatment versus RFA, and RFA versus SBRT. Each comparison group uses average negative log-likelihood with L2 regularization as the loss function for the model to output an estimated hazard risk score 12,14 independent of treatment modality (given that modality is not included as a covariate within individual models and the DeepSurv architecture enables model prediction without a priori specification of potential treatment covariate interactions). An optimal strategy was chosen on the basis of an assessment of the greatest improvement in estimated hazard risk score by treatment modality for each patient across all models.…”