A 3-Window Framework for the Discovery and Interpretation of Predictive Temporal Functional Dependencies

Clinical databases collect large volume of data. Relationships and patterns within these data could provide new medical knowledge. Temporal data mining has as major scope the discovery of potential hidden knowledge from large amounts of data, offering the possibility to identify different features less visible or hidden to common analysis techniques. In this work, we show how temporal data mining, precisely mining of functional dependencies, can be fruitfully exploited to improve clinical prediction. To develop an early prediction model, a window-based data aggregation approach could be a good starting point, therefore we introduce a new temporal framework based on three temporal windows designed to extract predictive information. In particular, we propose a methodology for deriving a new kind of predictive temporal patterns. We exploit the predictive aspect of the approximate temporal functional dependencies, formally introducing the concept of Predictive Functional Dependency (PFD), a new type of approximate temporal functional dependency. We discuss some first results we obtained by pre-processing and mining ICU data from the MIMIC III database, focusing on functional dependencies predictive of Acute kidney injury (AKI).