McLaverty, Brian
(2023)
Unifying Data-Driven Modeling with Machine Learning to Improve Personalized Treatment of Critical Care Patients.
Doctoral Dissertation, University of Pittsburgh.
(Unpublished)
Abstract
Sepsis is a life-threatening organ dysfunction resulting from a dysregulated host response to infection. Early identification and appropriate management of sepsis has been identified as a global health priority. The development and validation of patient-specific models of acute inflammation offer the opportunity to expand current knowledge of underlying interpatient differences in the fight against infection. Early prediction into identified sepsis subtypes exhibiting differing health outcomes could be used as a guide for clinicians to develop a personalized therapeutic regimen.
A dynamical model of inflammation response in sepsis was derived, leveraging cytokine mediators that play a key role in sepsis pathology. The model was validated against a selected cohort of sepsis patients from the Protocolized Care for Early Septic Shock (ProCESS) trial to create patient-specific models of inflammation. Using the fitted model parameters and unsupervised machine learning, four subtypes of septic patients emerged with distinct inflammation responses and clinical outcomes. A clinical tool was proposed that accurately predicts patient membership into subtypes exhibiting high or low mortality risk using cytokine levels measured within six hours of hospital admission.
Dialysis patients frequently experience intradialytic hypotension (IDH), which is an independent predictor of mortality. Development of a personalized treatment support system for hemodialysis that reduces risk of IDH could potentially improve dialysis patient outcomes. A predictive model of future risk of IDH was developed by training and testing a random forest model using electronic health record data from hemodialysis treatments performed at UPMC acute care facilities. The model forecasted future IDH several hours ahead of occurrence and produced dynamic risk evolution toward instability. An early-warning tool was developed that accurately classifies a patient as high or low risk for IDH using model-derived risk scores available within minutes of dialysis initiation. A risk-based reinforcement learning algorithm was then developed that recommends preemptive, personalized treatment. The agent-suggested treatment strategy resulted in decreased incidence of hypotension and increased accomplishment of individualized fluid goals in silico at the cost of increased intervention. Clinicians could use the tools presented in this work to provide timely, individualized treatment to patients and potentially improve sepsis and hemodialysis patient outcomes.
Share
| Citation/Export: |
|
| Social Networking: |
|
Details
| Item Type: |
University of Pittsburgh ETD
|
| Status: |
Unpublished |
| Creators/Authors: |
|
| ETD Committee: |
|
| Date: |
13 June 2023 |
| Date Type: |
Publication |
| Defense Date: |
19 January 2023 |
| Approval Date: |
13 June 2023 |
| Submission Date: |
25 March 2023 |
| Access Restriction: |
No restriction; Release the ETD for access worldwide immediately. |
| Number of Pages: |
119 |
| Institution: |
University of Pittsburgh |
| Schools and Programs: |
Swanson School of Engineering > Chemical and Petroleum Engineering |
| Degree: |
PhD - Doctor of Philosophy |
| Thesis Type: |
Doctoral Dissertation |
| Refereed: |
Yes |
| Uncontrolled Keywords: |
Dynamic Modeling, Reinforcement Learning, Machine Learning, Risk Modeling, Hemodialysis, Sepsis |
| Date Deposited: |
13 Jun 2023 14:07 |
| Last Modified: |
13 Jun 2023 14:07 |
| URI: |
http://d-scholarship.pitt.edu/id/eprint/44328 |
Metrics
Monthly Views for the past 3 years
Plum Analytics
Actions (login required)
 |
View Item |
![[feed]](/images/feed-icon-32x32.png){kind=icon}