Data-driven models for mortality assessment at the Intensive Care Unit

The Intensive Care Unit (ICU) is known as the department with the highest mortality numbers in any hospital. Patients at the ICU require extensive assessment by hospital staff. This is partially achieved through the use of state-oftheart monitoring devices that provide measurements and trends about each patient. This means that large amounts of useful data is available for research. With the increasing number of patients, maintaining a high standard of care at the ICU becomes time consuming and brings high operational costs with it. In this paper we develop data-driven models that can be used by hospital staff to assess the physical state of patients at the ICU. To be more precise, these models provide hospital staff with predictions about the risk of mortality of patients at the ICU. Two models were developed and compared based on accuracy and scalability. The first approach follows a pipeline that prepares the data for predictive modelling with logistic regression. The second approach allows the use of the instance-based learning model K-nearest neighbor (KNN) with Dynamic Time Warping (DTW). We show that logistic regression (AUC of 0.84) significantly outperforms KNN (AUC of 0.68) by conventional criteria. This paper provides insights on the performance of both algorithms and could be used as inspiration for further research.