Predicting patient status dependent on their treatment using a clustering model with SAX

Abstract

The goal of this research is to try to predict the condition of a patient in the future, given the current condition and conditional on the treatment using data of the University Medical Center Utrecht. There are two separate use cases: the Pediatric Intensive Care Unit (PICU) and the operating room (OR). In both cases, the haemodynamic parameters are predicted. The prediction is aided by lab measurements and patient information and made conditional on the intervention. The intervention in the PICU dataset consists of inotropes and in the OR dataset it is a combination of inotropes and anesthetics. A model is developed that uses K-Means combined with Symbolic Aggregate ApproXimation (SAX) to cluster the patient windows and uses these clusters and the interventions to build a probability matrix. This probability matrix can be used to predict new cases. The model performs significantly better than a model predicting no change. The model performs equally well as a clustering method using only K-Means, but is better able to consistently cluster the patient status into meaningful categories. The influence of the interventions cannot be isolated as they are too highly correlated with the patient status.