PROJECT TITLE :

Data pre-processing using Neural Processes for Modelling Personalised Vital-Sign Time-Series Data

ABSTRACT:

In order to better manage available resources, clinical time-series data are frequently retrieved from electronic medical records and put to use in the construction of predictive models of adverse events. It can be difficult to use many common Machine Learning methods when dealing with this kind of data because the data is frequently sparse and sampled in an irregular manner. Interpolation can be used to fill in missing data by either using linear regression or "carrying forward" the most recent value. Imputation is another application of Gaussian process (GP) regression, as is the frequent resampling of time series at regular intervals. GP regression can be used to perform both of these tasks. When generalized projections are used, it may be necessary to conduct a comprehensive and most likely ad hoc investigation in order to determine the structure of the model, which may include a suitable covariance function. This can be difficult to do with multivariate real-world clinical data, particularly when time-series variables exhibit different dynamics to one another than other variables. In this study, we construct generative models to estimate missing values in clinical time-series data by using a neural latent variable model that is more commonly referred to as a Neural Process Model (NP). In order to learn about the global uncertainty in the data, the NP model makes use of a conditional prior distribution in the latent space. This is done by modeling variations on a more local level. This prior is not predetermined like in traditional generative modeling; rather, it is learned along with the rest of the model during the training process. As a result, the NP model offers the adaptability necessary to accommodate the changing nature of the clinical data that is available. We propose a variant of the NP framework for the purpose of efficiently modeling the mutual information that exists between the latent space and the input space, while simultaneously ensuring that meaningful learned priors are generated. Experiments conducted with the MIMIC III dataset demonstrate the effectiveness of the proposed approach in comparison to the methods that have traditionally been used.