Critically ill patients require treatment in specialized settings of intensive care units (ICU) to enable continuous recording of organ function parameters for early recognition of abnormal values, consecutive rapid patient assessment and appropriate interventions. At any given point, a clinician will try to assess if a patient’s condition is improving or deteriorating. The ICU of a hospital is therefore a data-rich environment with many parameters and many longitudinal observations. Machine Learning techniques excel in domains where one needs to describe and control a complex phenomenon and has many observations available. This motivates the use of Machine Learning to support physicians in better and faster understanding the condition of a patient and to support decision making.
Our research is structured into three broad aims. In Aim 1 we will make data accessible for computational research. This may seem trivial at first sight, but it is a crucial first step that needs to be completed carefully to make use of data from medical facilities that is stored but so far unavailable for research. In particular, we will assess different ways to encode, store and access large intensive care datasets. Also, we intend to develop a strategy to transform databases recorded from different sources into a standard format wherever feasible. Aim 2 is to develop the necessary methodology and software to predict organ dysfunction defined by specific medical endpoints. We will then develop novel statistical and machine learning approaches to to deal with missing data, multiple time scales, many outliers and other challenges. The availability of observational medical data from intensive care units opens the door for retrospective analyses with Machine Learning techniques with the aim to develop decision support systems. Thus, Aim 3 is farther reaching, and intends to advance the state of Machine Learning research in healthcare. This aim considers retrospective uses of large observational datasets, like the one collected at the University Hospital Bern ICU. This research aims to assist the practice of healthcare with evidence-based support systems powered by Machine Learning, which draw on large quantities and specifics of healthcare data not easily processable by clinicians.