Imputation and forecasting in medical multivariate time series using probabilistic generative machine learning

In the medical field, time series data are routinely collected on a large scale for a variety of purposes, such as patient care or clinical research. The data collected in everyday clinical practice are often subject to interruptions due to technical or human errors that cannot be prevented in these not fully standardised setups. Such disturbances include patient movement, sensor calibration phases, or electromagnetic interference. Not only intermediate interruptions but also a lack of estimates of future developments hinders personalised patient care, early intervention, and efficient planning in the medical domain. For these reasons, and because of the variability of medical time series, generic data-driven multivariate imputation and forecasting methods are needed. They help to achieve the data quality required for sophisticated analysis and decision support based on biomedical time series data. Due to the variance inherent in the problems described, the focus of this thesis is on probabilistic methods. These do not produce a single deterministic imputation or forecast, but allow a statistical view of the two tasks. To this end, a system based on recurrent encoder-decoder networks and attention mechanisms has been designed that can represent data distributions using the generative adversarial training method. The aim has been to develop a probabilistic generative method that can generate realisations of the missing or incorrect step values, interpreted as random variables. These random variables should be distributed according to the unknown data distribution conditioned on the non-missing contexts and measurements of other synchronously tracked variables. Different building blocks, training procedures and benchmark models have been systematically evaluated. Two sources of real-world data from different clinical contexts have been used for the evaluation.