A machine-learning approach to survival time-event predicting: Initial analyses using stomach cancer data
Abstract
The problem of predicting whether an event of interest does occur, and if so, when it occurs for given values of independent variables, is one of the basic tasks in survival analysis. Cox proportional hazard model and its numerous variants are usually used to handle this task; however, they are limited by strict statistical assumptions.
In this study, we - rather than estimate the event's hazard function - prefer to make direct predictions by decomposition of the two-dimensional dependent time-event variable, depicting the event occurrence and time-to-event, into these two components, overcoming the statistical limitations. While the first part for the event occurrence is considered as a classification task, the second part for the time-to-event estimation is assumed to be a regression task.
The variable's parts are treated as the classification and regression tasks, therefore built using machine-learning algorithms such as logistic and linear regression, naïve Bayes classifiers, classification and regression trees, support vector machines and neural networks, and applied together with the Cox's model to stomach cancer data. The machine-learning modeling of both the decomposed survival time-event variable's parts seems to be a promising alternative to predictions based on the Cox's hazard modeling.