Off-campus WSU users: To download campus access dissertations, please use the following link to log into our proxy server with your WSU access ID and password, then click the "Off-campus Download" button below.
Non-WSU users: Please talk to your librarian about requesting this dissertation through interlibrary loan.
Access Type
WSU Access
Date of Award
January 2022
Degree Type
Dissertation
Degree Name
Ph.D.
Department
Electrical and Computer Engineering
First Advisor
Hao Ying
Abstract
Sepsis is a systematic inflammatory state and life-threatening condition caused by the body’s response to an infection. Early identification of sepsis patients and prompt antibiotics is critical to reduce fatality and improve outcomes. In the literature, there are an abundance of sepsis clinical decision support systems developed for in-hospital or intensive care units. Only a few sepsis alerts were designed for the initial hours of patient admission into the emergency department (ED). This dissertation aims to provide a transparent and explainable model to identify sepsis patients in the ED during their first six hours of care. The detection performance emphasis is on sensitivity, which reflects real-world clinical priority. We analyzed the existing genetic-algorithm-optimized crisp-rule-based model and improved its performance by adding creatinine and its related rule. We also fuzzified the original and newly-added crisp rules and added a second stage to form a novel two-stage fuzzy-rule-based model. The first stage aimed to boost the sensitivity of the model; increase the correctly identified sepsis cases. A potential drawback of increasing sensitivity is the increase in false alerts. Excessive false positive alerts may create a dangerous condition of alert fatigue in which the care providers become insensitive to all alerts. Therefore, we added a decision tree model as the second stage to filter out potential false alerts and provided the probability of false alerts based on similar historical cases. For benchmarking, we trained 24 machine learning models, including decision tree, discriminant analysis, logistic regression, k-nearest neighbors, ensemble classifications, support vector machine, and neural network. The models in this dissertation used two datasets that were the electronic medical records for patients admitted to the ED of the Detroit Medical Center in Michigan, USA. The cases in the first dataset were individually adjudicated by an expert physician and were used in training. It consisted of 912 sepsis and 975 non-sepsis patients. The second dataset used for testing consisted of 1175 sepsis and 7800 non-sepsis patients, which was not adjudicated for practical reasons, and thus may contain misclassified cases and uncertainties. Using the first dataset, the performance of the previous genetic-algorithm-optimized crisp-rule-based-sepsis model was 90.9% sensitivity, 90.9% specificity, and 90.3% positive predictive value. After adding the creatinine rule, the new model increased sensitivity by 2.63% at the expense of decreased specificity by 0.23% and positive predictive value by 0.06%. Compared to the genetic-algorithm-optimized crisp rule-based model, our novel two-stage fuzzy-rule-based model increased sensitivity by 3.95%, specificity by 1.41%, and positive predictive value by 1.72%. It had a higher sensitivity than the genetic-algorithm-optimized crisp-rule-based model, the genetic-algorithm-optimized crisp-rule-based model with the creatinine rule, and the 24 machine learning models. The fine Gaussian support vector machine had the best overall specificity and positive predictive value at 94.78% and 95.39%, respectively, but it also had the lowest sensitivity at 80.41%. Using the second dataset, the performance of the previous genetic-algorithm-optimized crisp rule-based model was 70.81% sensitivity, 90.96% specificity, and 54.13% positive predictive value. Compared to the crisp model, the two-stage fuzzy-rule-based model was lower in sensitivity by 0.26%, specificity by 1.06%, and positive predictive value by 2.86%. The decrease of performance in the two-stage fuzzy-rule-based model may be attributed, at least partly, to the misclassification and uncertainties associated with the second dataset. In conclusion, adding creatinine rule to the genetic-algorithm-optimized crisp-rule-based model increased the sensitivity of the sepsis alert. The two-stage fuzzy-rule-based model deserves further exploration because it demonstrated the best sensitivity performance with the first dataset while providing complete transparency.
Recommended Citation
Mohamed, Ahmed, "Sepsis Detection In The Emergency Department: A Two-Stage Fuzzy-Rule-Based Approach" (2022). Wayne State University Dissertations. 3645.
https://digitalcommons.wayne.edu/oa_dissertations/3645