Towards Causal Discovery on EHR data : Evaluation of current Causal Discovery methods on the MIMIC-IV data set

Sammanfattning: Causal discovery is the problem of learning causal relationships between variables from a set of data. One interesting area of use for causal discovery is the health care domain, where application could help facilitate a better understanding of disease and treatment mechanisms. The health care domain has recently undergone a major digitization, making available a large amount of data for use in learning algorithms, available in formats such as medical images or electronic health records. This thesis aims to explore the application of causal discovery on electronic health record data. We provide an overview of the field of causal discovery and identify 3 contemporary methods for causal discovery on time-series data which we apply on a preprocessed version of the MIMIC-IV data set. Each causal discovery method is run on time-series comprising of electronic health record data related to hospital stays for patients with sepsis. We provide an empiric report of the overlap between the learned graphs from different hospital stays as a heuristic evaluation measure. We find that it is possible to identify common themes in the learned graphs between different causal discovery methods, indicating potential practical value of causal discovery on electronic health record data. We also identify important considerations for future application and evaluation, such as incorporating extensive domain knowledge, and provide suggestions for future work.