Link to the University of Pittsburgh Homepage
Link to the University Library System Homepage Link to the Contact Us Form

Safe Reinforcement Learning for Sepsis Treatment

Lu, Liling (2022) Safe Reinforcement Learning for Sepsis Treatment. Master's Thesis, University of Pittsburgh. (Unpublished)

Download (1MB) | Preview


Sepsis, defined as an overactive immune system response to infection followed by acute life-threatening organ failure, kills eight million people annually. Mortality of acute sepsis is up to 50%, and significantly higher in low-income countries. The correction of the absolute hypovolemia with intravenous fluids and vasopressors is the most difficult aspect of sepsis treatment. There were promising Reinforcement Learning (RL) approaches to learn the optimal administration of vasopressor and intravenous fluids to treat septic patients. However, the existing RL approaches did not take some safety constraints into consideration. Firstly, they only captured end-point outcome and ignored patients’ intermediate outcomes, which are also very important to patients. Secondly, they did not consider the dose change of vasopressor within a short amount of time. This is not in accordance with clinical safety protocol, which states that the dose change of vasopressor should be gradual, while a dramatic major change of vasopressor dose is unsafe to patients. In this project, we extended an existing model-free Q-learning algorithm by addressing its two safety concerns. We learned a more robust and safer AI agent which takes intermediate outcomes into consideration by incorporating SOFA score and lactate level as intermediate health status. Additionally, we developed another safer and more competitive AI agent to address the sudden major change in vasopressor dose use by adding vasopressor penalty. The two learned AI agents are more adherent to current clinical practices and knowledge.


Social Networking:
Share |


Item Type: University of Pittsburgh ETD
Status: Unpublished
CreatorsEmailPitt UsernameORCID
ETD Committee:
TitleMemberEmail AddressPitt UsernameORCID
Committee ChairTang, Lulutang@pitt.edulutang
Committee MemberChang, Chung-Chou H.changj@pitt.educhangj
Committee MemberTalisa, Victor Brodzikvit13@pitt.eduvit13
Date: 12 May 2022
Date Type: Publication
Defense Date: 12 April 2022
Approval Date: 12 May 2022
Submission Date: 29 April 2022
Access Restriction: No restriction; Release the ETD for access worldwide immediately.
Number of Pages: 68
Institution: University of Pittsburgh
Schools and Programs: School of Public Health > Biostatistics
Degree: MS - Master of Science
Thesis Type: Master's Thesis
Refereed: Yes
Uncontrolled Keywords: Sepsis, Reinforcement Learning, Q-learning
Related URLs:
Date Deposited: 12 May 2022 13:28
Last Modified: 12 May 2022 13:28


Monthly Views for the past 3 years

Plum Analytics

Actions (login required)

View Item View Item