Deep-learning in survival analysis

Abstract

In this paper, hospitalisation duration is modelled using traditional survival model, machine-learning models and deep-learning models. Machine-learning and deep-learning algorithms typically assume that all event of interest are known at the time of modelling. However, hospitalisation duration is a time-to-event data with right-censoring (not all events are known at the time of modelling). Hence, specific techniques were employed to deal with this inconsistency. Subsequently, the various models are evaluated using the Concordance-index (C-index). It is a ranking evaluation metrics that can account for censored observation. The empirical results showed that the deep-learning model is best in predicting hospitalisation despite the small dataset. This paper can be further improved by incorporating geo-spatial data in the analysis.