Olivier Bouaziz - A comprehensive framework for time to event predictions with right censored data

Predicting the time to an event of interest, based on patient attributes, is of great interest when analysing medical data. However, time to event data usually suffer from right-censoring which makes it difficult both to propose a prediction algorithm and to assess the quality of the predictions since the true times are not all observed. In this talk we will propose a comprehensive framework for evaluating the quality of predictions made by a prediction model. We present a new Mean Squared Error (MSE) criterion based on Inverse Probability Censoring Weighting (IPCW), a new method to construct prediction intervals based on the split conformal approach and a new statistical test to assess the importance of each variable. All those methods are designed to handle right-censoring. If time allows we will also discuss how to construct machine learning algorithms by combining pseudo-observations with the super learner, an ensemble method.

This talk is based on the following paper:

https://hal.science/hal-04143419v1/document

It is a joint work with Ariane Cwiling (MAP5) and Vittorio Perduca (MAP5).