Benchmarking Multi-instance Learning for Multivariate Time Series Analysis

Successful incorporation of Electronic Health Records to the data mining tools created new frontiers in digital clinical data analysis. One of the well-known applications of clinical data analysis is the mortality prediction of patients in intensive care units (ICUs). One important aspect of mortality prediction is the analysis of multivariate time series of observations after 24 or 48 h of ICU admission. Recent mortality prediction models for ICU patients are based on either recurrent neural networks or traditional machine learning algorithms using statistical summaries of timestamped observations. Instead of using complex neural network architectures and statistical summaries, we transform multivariate time series into multi-instance representation by keeping the expressiveness of the original observations. We then perform mortality prediction using multi-instance machine learning algorithms. Our empirical study shows that multi-instance representation achieves comparable or better (in some configurations) performance in various experiments.