DeepPseudo: Pseudo Value Based Deep Learning Models for Competing Risk Analysis

Competing Risk Analysis (CRA) aims at the correct estimation of marginal probability occurrence an event in presence competing events. Many statistical approaches developed for CRA are limited by strong assumptions about underlying stochastic processes. To overcome these issues and to handle censoring, machine learning have designed specialized cost functions. However, not generalizable, computationally expensive. This paper formulates as a cause-specific regression problem proposes DeepPseudo models, which use simple effective feed-forward deep neural networks, predict cumulative incidence function (CIF) using Aalen-Johansen estimator-based pseudo values. models capture time-varying covariate effect on CIF while handling censored observations. We show how can address co-variate dependent censoring modified Experiments real synthetic datasets demonstrate that our proposed obtain promising statistically significant results compared state-of-the-art approaches. Furthermore, we explainable methods such Layer-wise Relevance Propagation be used interpret predictions models.