Dealing With Missing Continuous Biomedical Data: a Data Recovery Method for Machine Learning Purposes

Abstract

There are different approaches to dealing with missing data. A common one is by deleting observations containing such data, but it is not applicable when the volume of the data is limited. In this case, a number of methods can be applied, such as Last Observation Carried Forward and the like. But these methods are not suitable when all data for a certain parameter are missing. This paper describes a possibility of addressing this issue in the case of time series of biomedical data. Behind the method is the idea of the human body as a complex system in which various parameters are correlated and missing data can be inferred from the available data using the estimated correlation. For this, machine learning-based linear regression models are built and used to recover data describing the sepsis state. Finally, recovered data are used to create a sepsis prediction system.