A Recurrent Network Approach to G-Computation for Sepsis Outcome Prediction Under Dynamic Treatment Regimes

Abstract

Sepsis is a life-threatening condition that occurs when the body’s normal response to an infection is out of balance. A key part of managing sepsis involves the administration of intravenous fluids and vasopressors, but prescribing the correct balance of interventions is challenging since both under- and over-resuscitation can lead to adverse outcomes. While many retrospective studies have attempted to understand the relationship between sepsis treatment, fluid overload, mortality, and other outcomes, most are correlation-based and cannot actually estimate the causal effects of intervention. Prospective randomized clinical trials allow researchers to test the effects of alternative therapies more directly, but these types of studies tend to span multiple years and recent results regarding optimal regimes have been conflicting.

In this thesis, we use methods from causal inference to predict outcomes in sepsis patients under different fluid and vasopressor strategies. Specifically, we explore a recurrent neural network approach to g-computation, a technique that allows us to estimate effects under treatments that are dynamic and time-varying. Our work builds on a previous sequential deep learning implementation known as G-Net. We evaluate G-Net using synthetic physiological data and show that it outperforms traditional linear regression models in predicting patient trajectories under alternative interventions. We then adapt and apply the improved architecture for analyzing outcomes under counterfactual treatment strategies in a real-world cohort of sepsis patients, using observational data collected from the intensive care unit. Our results demonstrate that G-Net is able to generate reasonable counterfactual estimates under alternative regimes.