Treatment Prediction in the ICU Using a Partitioned, Sequential, Deep Time Series Analysis

We have developed a time-oriented machine-learning tool to predict the binary decision of administering a medication and the quantitative decision regarding the specific dose. We evaluated our tool on the MIMIC-IV ICU database, for three common medical scenarios. We use an LSTM based neural network, and considerably extend its use by introducing several new concepts. We partition the common 12-hour prediction horizon into three sub-windows. Partitioning models the treatment dynamics better, and allows the use of previous sub-windows' data as additional training data with improved performance. We also introduce a sequential prediction process, composed of a binary treatment-decision model, followed, when relevant, by a quantitative dose-decision model, with improved accuracy. Finally, we examined two methods for including non-temporal features, such as age, within the temporal network. Our results provide additional treatment-prediction tools, and thus another step towards a reliable and trustworthy decision-support system that reduces the clinicians' cognitive load.