WMRNN: Weighted Modal Regression Neural Networks for Right Censored Data

In this paper, we propose a novel weighted modal regression neural network (WMRNN) for analyzing right-censored data. WMRNN provides a flexible nonlinear framework for modeling the conditional mode of the response variable based on covariates. This method excels in predicting the most likely outcomes and enhances predictive accuracy, effectively managing survival data even when it exhibits asymmetric, peaked, or heavy-tailed distributions. It functions well regardless of the independence of censoring variable from covariates. By harnessing the nonlinear capabilities of deep neural networks within the modal regression framework, WMRNN addresses random censoring using inverse probability of censoring weighting (IPCW) in its loss function. This approach eliminates the need for predetermined functional forms, allowing for greater adaptability. Extensive simulations and two real-world case studies demonstrate the strong performance of WMRNN in predicting outcomes from right-censored data, highlighting its potential as a valuable tool in survival analysis.