Novel Deep Learning Tools for Clinical Decision Support in Postoperative Pain Management

Abstract
Postoperative pain (POP) burdens millions of Americans, and it costs hundreds of billions dollars to the US
healthcare system annually. Poorly managed acute POP often leads to increased morbidity, mortality, and many
other complications, such as chronic POP and opioid overuse. Accurate prediction of POP outcomes and in-depth
understandings of causal mechanisms of POP is critical to develop effective POP management. Also, many POP
studies indicate heterogeneity of responses to anesthesia methods and postoperative substance use, suggesting
a critical need for effective methods to accurately identify patient subgroups for more effective POP management
tailored to the individual patient's needs. However, achieving these goals is challenging due to the complex
POP mechanisms and limited data from ideal large randomized controlled trials. On the other hand, abundant
observational POP data found in surgery patients' electronic health records (EHRs) are readily available, and
they can serve as a cost-effective alternative to address the critical challenges in POP management.
However, the etiology of POP is intricate, i.e. many factors may interweave with each other and impact POP
outcomes non-linearly and non-additively, introducing daunting modeling challenges. Furthermore, confounding,
a major concern associated with observational data, represents a particular challenge for conducting causal
analysis on POP data. Also, POP outcomes such as POP intensity scores are often irregularly and repeatedly
measured, and distributed non-normally with two distinct data processes, requiring more advanced analysis
methods. This proposal aims to overcome these analytic and modeling challenges with state-of-the-art deep
learning methods to improve POP management. Specifically, we will 1) establish robust deep learning models for
more accurate predictions of both acute and chronic POP to achieve timely POP control and care; 2) develop valid
deep learning based semi-parametric methods to identify true causal factors and mechanisms of POP to design
more effective POP management interventions; and 3) build powerful models to conduct robust hidden subgroup
analysis to develop the optimal POP management tailored to the individual patient's needs. Methods developed
in Aims 1 3 are motivated and will be tested by two unique data: a large EHR data from the University of North
Carolina at Chapel Hill's Carolina Data Warehouse for Health (CDW-H), and a high-quality cohort data from NIH-
funded TEMporal PostOperative Pain Signatures study, which complements the CDW-H in scale and scope. The
project will elucidate the scientific underpinnings of POP mechanisms and provide improved POP management.