Multimodal physiological time series analysis for outcome prediction in the intensive care unit.

The intensive care unit (ICU) is a hospital department where critical patients are monitored and healed: the sensors used to check their vital conditions produce a huge amount of data that can be exploited to predict outcome about their future conditions. In this work mortality prediction is performed using data recorded from more than 9000 patients in different hospitals of the Capital Region of Denmark from 2009 and 2016. Since the high number of physiological variables acquired in the ICU a study of which variables subset provides more information is carried on, taking into account the number of patients who have that subset of physiological variables monitored. Both signal processing and statistic methods are used to extract information from the first 24 hours after the admission in the ICU; then a long short term memory (LSTM) model is used for classification and regression tasks. Finally, SHapley Additive exPlanations approach is used to increase the interpretability of the model in order to understand which variables require more attention and to help doctors and technicians in the choice of the best treatment pathway. Regardless of which task the model was designed to, age of admission is the most important feature; vitals signs generally account the most with respect to arterial blood gas (ABG) measurements and no significant improvements are present when features from empirical mode decomposition (EMD) are added to the feature space. Moreover the model has comparable performances either the training includes all the patients in the dataset or only the patients with measurements for every variable in the feature space.