Factor enhanced DeepSurv: A deep learning approach for predicting survival probabilities in cirrhosis data

Abstract

Background

Over the years, various models, including both traditional and machine learning models, have been employed to predict survival probabilities for diverse survival datasets. The objective is to obtain models that provide more accurate estimates of survival probabilities. Certain datasets exhibit complex nonlinear effects and interactions between variables that may necessitate the application of deep learning algorithms to comprehend the underlying data generation process.

Method

In this paper, we introduced Factor Enhanced DeepSurv (FE-DeepSurv), a novel deep neural network designed to study complex structures and excels at filtering noise within predictors, thereby enhancing precision of survival probability estimates. FE-DeepSurv incorporates factor analysis to reduce predictor dimensionality, applies a transformation technique to account for data censoring, and employs a deep neural network to predict conditional failure probabilities for each time interval. These predictions are subsequently utilized to estimate survival probabilities for each subject. We applied our proposed model to study cirrhosis survival data, a secondary data from Mayo Clinic trial focused on primary biliary cirrhosis (PBC) of the liver and compared its performance with the Cox proportional hazard model (Cox model), random survival forest (RSF), DeepHit, and DeepSurv, using the concordance index (C-index), brier score (BS), and integrated brier score (IBS).

Results

The results show that FE-DeepSurv outperforms many existing survival models. FE-DeepSurv's accurate predictions of survival probabilities and hazard rates can drive improvements in clinical practice, healthcare management, insurance risk assessment, and various other domains.

Conclusions

By adopting FE-DeepSurv, institutions can harness the power of advanced analytics to make more informed decisions, ultimately leading to better outcomes across multiple sectors.