SCH: Synergizing Topological Deep Learning and Spatio-Temporal Causal Inference to Unpack Health

Harvard University D/B/A Harvard School Of Public Health, Boston MA

Investigators

Abstract

This research aims to provide novel statistical and deep learning (DL) tools and theory to address open methodological challenges needed for understanding and improving health outcomes in the United States. It focuses especially on un- covering patterns on large-scale health datasets combining individual and aggregated spatiotemporal data to uncover and measure the effects of environmental exposures such as air pollution and their interaction with extreme weather events. To enable such analysis, this multi-disciplinary research brings together computer scientists and biostatisticians to promote principled and synergistic advancements in Topological Deep Learning (TDL) and spatial Causal Inference (CI) to address open technical challenges in the fields of TDL, CI, and Spatial Statistics. The project is centered around three thrusts. Thrust 1: Develop new TDL methods to process multi-resolution irregular areal data; Thrust 2: Leverage TDL in spatiotemporal causal inference; Thrust 3: Establish a methodological framework to jointly leverage TDL-based spatial and individual-level representations. Innovations include new TDL methods to address key gaps in DL for areal spatial data, the first theoretically grounded framework leveraging TDL for learning from aggregated outcomes under spatial heterogeneity in areal data, and new TDL methods for causal inference using spatiotemporal data. This project will result in the first multi-modal scalable framework for learning joint representations of geospatial data and individual-level health records. Large-scale case studies utilizing Medicare data from 2000 to 2020 and a vast amount of spatial and longitudinal data will allow the translation of research results into actionable information to inform policies and interventions aimed at improving health outcomes. The successful completion of this research will address scientific policy-relevant questions such as: Can joint modeling of individual clinical histories and time-varying environ- mental exposures predict hospitalizations or disease progression reliably across geographic regions? What patterns of multimorbidity emerge when embedding individual health trajectories in spatial contexts, and how do environmental exposures (e.g. extreme heat events, air quality fluctuations) contribute to these trajectories?