View analytic
Tuesday, January 24 • 4:30pm - 6:00pm
Poster: Using Domain Knowledge to Construct Causal Models from Clinical Observational Data

Sign up or log in to save this to your schedule and see who's attending!

Causal discovery methods seek to ascertain causal attribution from observational data. Although their use has been established in cancer and epidemiological research, surprisingly little work has been done with such methods in the area of the detection of causal drug/adverse drug event (ADE) relationships in clinical observational data derived from Electronic Health Records (EHR). Since these data were originally created for other purposes, they are inaccurate and incomplete. We reason that by integrating constraints from domain knowledge, causal methods may compensate for issues that limit the accuracy of purely statistical approaches. To evaluate this hypothesis, we used a publicly available reference data set with 4 ADEs and 399 drug-ADE pairs. Mining the literature, we identified covariates that fell within the orbit of the respective drug-ADE pairs using discovery patterns (relationship constraints based on normalized predicates). We calculated baseline scores using standard disproportionality metrics, with and without the identified covariates. Where drug pharmacological class strongly indicated a causal association with the ADE (e.g., Ibuprofen ∈ NSAIDs -> gastrointestinal bleeding), directed edges were included as prior knowledge. We then constructed causal graphs from the clinical data. We attained significant predictive improvements of ~0.05-0.3 AUC over traditional statistical methods, with ~0.7-0.9 overall AUC.

Tuesday January 24, 2017 4:30pm - 6:00pm
BioScience Research Collaborative Event Hall 6500 Main Street, Houston, TX 77030-1402

Attendees (1)