Towards Learning with High Causal Fidelity from Longitudinal Event Dat

Longitudinal event data (LED) are irregularly time-stamped multi-type event sequences collected from heterogeneous subjects throughout different time scales. In this dissertation, we are interested in developing machine learning models and algorithms to identify potential causal relationships among various event types from LED so as to provide actionable insights for better decision-making.;As a concrete example of LED, we consider the use of electronic health records (EHRs). By viewing the occurrences of different drug prescriptions, condition diagnoses, and physical measurements as different event types, we are interested in identifying potential causal relationships regarding how different drugs could influence the occurrences of various conditions and the values of different physical measurements. This problem leads to two pivotal health applications: computational drug repositioning (CDR) and adverse drug reaction (ADR) discovery.;To deliver better CDR and ADR discovery, we focus on developing machine learning models and algorithms with high causal fidelity. Causal fidelity is concerned with whether a method can effectively identify signals residing in the data that indicate potential causality. By confronting various theoretical, methodological, and empirical issues stemming from the intricacies of LED, our models and algorithms strive to deliver high causal fidelity via the identification of signals in LED that are reflective of potential causal relationships among various event types. This leads to the title of the dissertation, Towards Learning with High Causal Fidelity from Longitudinal Event Data.;The primary content of the dissertation is hence to present how high causal fidelity can be achieved in CDR, ADR discovery, and beyond. Our solution is to identify and address three fundamental challenges constitutional to the intrinsic nature of LED---inhomogeneity, irregularity, and interplay---summarized as the 3-I challenge. We demonstrate that by a careful treatment of the 3-I challenge, it is possible to develop machine learning models and algorithms with high causal fidelity, as shown by the improved performance of CDR and ADR discovery exhibited in this dissertation.