Deep Recurrent Neural Networks for Mortality Prediction in Intensive Care using Clinical Time Series at Multiple Resolutions.
Suparna Ghanvatkar, Vaibhav Rajan · Journal of the Association for Information Systems · 2019
Mortality models in Intensive Care Units (ICU) are important for clinical decision support tasks such as identifying high-risk patients and prioritizing their care. Previous mortality models have used predictive variables mainly from Electronic Medical Records (EMR) where each patient observation can be represented as a sparse multivariate time series. Bedside monitors are another common data source in ICUs containing high-resolution time series, which have not been explored in combination with EMR data for mortality modelling. We take the first step towards building such a model. Specialized techniques developed for sparse time series cannot be used to model multiple time series at different resolutions. To address this problem, we develop MTS-RNN, a new deep recurrent neural network architecture. Our preliminary experiments on real clinical data show that MTS-RNN outperforms state-of-the-art mortality models in predictive accuracy, highlighting the importance of using clinical time series at multiple resolutions for ICU mortality prediction.