347 Modeling long-term environmental effects on discrete events using shapelets: An application for stillbirth

Objectives/Goals: To develop an informatics framework that will allow study of environmental effects on stillbirth at large scale (i.e., US-level) and leverage recent advances in machine learning and artificial intelligence to produce reproducible results that can be compared across multiple institutional settings. Methods/Study Population: Experimental exposure data are often available in “absolute time,” where a clinical event can be anchored using a timeline transformation. We associate each stillbirth event with a set of {ti…ti+1}L shapelets [1] associated with a location, L, and time intervals for the entire dataset. These shapeless are aggregated using a state-of-the-art shapelet classifier [2]. An autoencoder is used to reduce the dimensionality of the stillbirth classification and to cluster stillbirth events according to their corresponding exposure patterns. The stillbirth cluster can be analyzed for other nonexposure (i.e., genetic, SDoH, and demographics) factors, which may be enriched and/or depleted. Results/Anticipated Results: The framework we are developing leverages a shapelet-based approach to produce clusters of stillbirth events according to their corresponding exposure patterns. These clusters can be analyzed for depletion or enrichment of nonenvironmental factors. This analysis will inform how to formulate (or not) class models of exposure that can be more informative and have better predictive power than overall population models. Moreover, the finding of depletion and enrichment of physiological properties of the individuals may lead to novel physiological hypotheses to better understand the injury mechanisms that the environmental exposure profile produces. Discussion/Significance of Impact: Nearly 20,000 babies are stillborn in the USA each year [3]. Environmental exposures, usually studied as time averages over certain periods of time, have produced mixed results for stillbirth risk [4]. However, temporal profiles matter [1], and we argue that they can be assessed using shapelet technology.