CAREER: Developing Actionable Methods for Observational Health Data

Ohio State University, The, Columbus OH

Investigators

Abstract

This award is funded in whole or in part under the American Rescue Plan Act of 2021 (Public Law 117-2). Observational health data contain large amounts of clinical information (e.g., comorbidities, prescriptions, and laboratory results) about heterogeneous patients and their responses to treatments in real-world settings. Many existing machine learning works focus on making predictions on observational data (e.g., chance of death or risk of heart attack) instead of providing actionable suggestions to physicians (e.g., when to use which drug for a specific patient). Developing actionable models for observational data is challenging in three aspects: 1) complex confounding factors that infect both treatment assignments and disease progression outcomes; 2) interpretability of treatment recommendation for actionable decision support; and 3) transferability of well-trained models to different environments. To address these challenges, the project will integrate deep learning algorithms and causal inference techniques to develop actionable methods for observational health data. The research team will closely collaborate with medical researchers and physicians for model validation on various clinical problems, and will actively seek technology transfer opportunities. The project will provide graduate and undergraduate students with new programs, courses, research, and internship opportunities on machine learning for healthcare applications. The project will also actively include students and outreach to high schools and the general public. The project will integrate deep learning algorithms and causal inference techniques for modeling longitudinal observational data and adjusting confounding factors, and develop actionable methods for observational data with two complementary tasks: individual treatment effects (ITEs), which estimate improvement in the outcome of taking a particular action to a particular target; and dynamic treatment regimes (DTRs), which derive a sequence of decision rules, one per stage of intervention, based on evolving treatment and covariate history. In the first thrust, the research team will model time-varying and hidden confounders by recurrently modeling historical information in observational health data for estimating ITEs; the researchers will generate a personalized treatment timing recommendation with an uncertainty quantification that achieves optimal causal effects; they will provide interpretability of treatment recommendations through both variable and global perspectives. In the second thrust, the research team will remove the confounding bias in observational health data via patient resampling and balancing weights; the researchers will develop a deconfounding reinforcement learning model for DTR learning, which simultaneously considers short-term and long-term rewards; they will introduce a policy adaptation method to the proposed model to transfer the learned DTR policies to new-source datasets. The project will result in the dissemination of new methods and software for irregularly spaced time series to the broader machine learning and healthcare communities. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.