Data in various domains are two-dimensional: there are two data types and the values of interest are the interactions between elements of each type. 

The typical machine learning setting when learning from interactions assumes that these interactions are binary or numeric. In many real-life cases, however, these interactions denote the time until an event occurs. Time-to-event analysis, also called survival analysis, is a well-studies area in statistics. The main issue that separates it from a regression task is that some data instances may not have an observed event, they are so-called censored observations. 

In this project, we will combine learning from interactions and time-to-event analysis. We will consider prediction, clustering and recommendation tasks. The developed methodology will be applied to a high impact domain. We will analyse data from clinical alarms in an intensive care unit, proposing algorithms to make these alarms smarter, and thereby contributing to preventive healthcare. Given the current overload of alarms, this project will also reduce alarm fatigue in clinical staff and alarm anxiety are sleeping disturbances in patients. Our techniques will lead to a clinical decision support system that employs explainable AI techniques to support the physician, and ultimately lead to better patient care.