Causal Discovery of Adverse Drug Events in Observational Data

Download or read book Causal Discovery of Adverse Drug Events in Observational Data written by Aubrey Barnard and published by . This book was released on 2019 with total page 0 pages. Available in PDF, EPUB and Kindle. Book excerpt: Automatic causal discovery without experiments offers to accelerate scientific investigation and knowledge acquisition, for example, by searching databases of electronic health records to discover the unknown effects of drugs. However, effective causal discovery requires methods that control for confounders and that scale to large data sets which have the power to support or refute causal hypotheses. Accordingly, this dissertation first introduces a method for efficiently learning formal structural causal models of medical histories via parameter learning in log-linear temporal Markov networks. Such models work well when all of the effects of interest are already defined and measured, but it might not be the case that all possible effects are suspected beforehand, especially when considering the adverse effects of drugs. Therefore, this dissertation next develops machine learning methods for causal discovery, including differential classification and temporal inverse probability weighting, that hypothesize likely causal effects while analyzing controlled observational studies. Applying all of these methods to causal modeling and finding adverse drug effects in synthetic and real-world electronic health records demonstrates their ability to accurately discover causal effects despite the irregularity, noise, and sparsity of such data. This dissertation thus establishes (1) that scalable, causal methods discover causal effects more accurately than methods that ignore causality, do not scale to large databases, or are not robust to the messiness of medical data, and (2) that methods that hypothesize effects improve genuine causal discovery by avoiding the limitations of human bias. In summary, the methods herein distinguish themselves by bridging machine learning and epidemiology: they bring causal inference and observational studies to machine learning, and they apply learning techniques and formal causal models to tasks in epidemiology. By integrating multiple approaches to causality, these methods achieve a wider perspective that overcomes the limitations of the individual perspectives, and leads to new methods for automatic causal discovery from observational data.