Methods for outcome prediction based on microbiome time series or can your microbiome dynamics predict when and will your future state happen?

Microbiomes have been associated with many different outcomes, including diseases, weights and even psychological aspects. Such relations were often used in machine learning contexts to predict future states. However, these machine learning (ML) applications were mostly static. The prediction was from the microbiome at a given time point to a predicted current/future state.

We here discuss our recent advances in such ML algorithms and propose ways to adapt them to a dynamic approach to predict future states from the microbiome dynamics. We often want to estimate the time to event (TTE), such as death or recurrence of the event of interest. This leads to a special type of learning task termed survival analysis or TTE analysis. The main challenge in performing such predictions is censored subjects and missing values.

We propose a general Recurrent Neural Network model containing two novel elements measuring the similarity between samples, and the time distribution of events. We show using this approach that censored data produces better predictors, and can be crucial, especially when the fraction of censored subjects and missing values is high.