Improving Safety and Trustworthiness in Data-Driven Decision Learning for Sepsis

University Of Pittsburgh At Pittsburgh, Pittsburgh PA

Investigators

Abstract

The goal of this project is to develop robust machine learning approaches that improve the safety and trustworthiness of existing data-driven decision-learning algorithms, and to ultimately enable the translation of algorithm-based decision learning into real-world medical practice. The proposal is motivated by the treatment of sepsis at the University of Pittsburgh Medical Center (UPMC). Sepsis is a heterogeneous group of conditions that could respond differently to individual treatments. Challenges remain in deciding the best strategy on a case-by-case basis despite existing general guidelines. UPMC encounters 50,000 sepsis patients each year and collects a rich set of data information, making it ideal for uncovering treatment effects on a personal level, and hence allowing the development of precision health strategies. While existing approaches could be directly applied to enable learning of decision-making algorithms to guide sepsis treatments in various settings, there are still major gaps between the idealized situations assumed by existing approaches and the reality, which can be summarized into the following: (i) a gap between data used for training and deployment, (ii) a gap between in silico simulation and bedside operation, and (iii) a gap between the reality and the assumed scenarios in causal discovery. Without properly accounting for these shortfalls, direct application of existing algorithms may lead to risky and unintuitive decisions that will eventually fail and lead to heavy consequences. We propose a Safety and Trustworthiness-Enabled decision Recommendation (STEER) system that combines a suite of novel approaches in optimization and causal inference to enhance the safety and trustworthiness of decisions being made by algorithms. Specifically, this project focuses on methodological improvements respectively in two aims: 1. To develop a decision-learning algorithm that reduces adverse outcomes in the presence of incomplete prospective data; 2.To develop a decision-learning algorithm that reduces suboptimal actions in the presence of unmeasured confounding. The effectiveness of the methods will be rigorously examined and compared with existing algorithms using UPMC data as primary examples and public data as secondary examples. Upon completion of this project, user-friendly software will be developed, and its applicability will be demonstrated using public data examples. This framework is general and can be readily extended to medical decision-making scenarios beyond sepsis, such as care at the emergency room and long-term patient management.