Neural Distribution Learning for generalized time-to-event prediction
Predicting the time to the next event is an important task in various domains. However, due to censoring and irregularly sampled sequences, time-to-event prediction has resulted in limited success only for particular tasks, architectures and data. Using recent advances in probabilistic programming and density networks, we make the case for a generalized parametric survival approach, sequentially predicting a distribution over the time to the next event. Unlike previous work, the proposed method can use asynchronously sampled features for censored, discrete, and multivariate data. Furthermore, it achieves good performance and near perfect calibration for probabilistic predictions without using rigid network-architectures, multitask approaches, complex learning schemes or non-trivial adaptations of cox-models. We firmly establish that this can be achieved in the standard neural network framework by simply switching out the output layer and loss function.
PDF Abstract
