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

We developed a neural network architecture to evaluate the patient's state using temporal data, patient's demographics and comorbidities. We examined the model's ability to predict both a binary medication-treatment decision and its specific dose in three common scenarios: hypokalemia, hypoglycemia and hypotension. We partition the common 12-hours horizon window into three sub-windows, examining how patterns of treatment evolve following a key clinical event or state. This partitioned analysis also helps in alleviating the problem of small data sets, by utilizing previous sub-windows' data as additional training data. We also propose a solution to the problem of the relative inability of dose-prediction models to output a "no treatment" classification, through the use of sequential prediction.