Date of Graduation
5-2026
Document Type
Thesis
Degree Name
Bachelor of Science in Industrial Engineering and Operations Analytics
Degree Level
Undergraduate
Department
Industrial Engineering
Advisor/Mentor
Shengfan Zhang
Committee Member
Brandon Crisel
Abstract
Sepsis remains a leading cause of morbidity and mortality among young children, with infants experiencing the highest risk of adverse outcomes, yet they are often aggregated with older pediatric populations in research and clinical risk models. Using nationally representative inpatient data from the 2018–2020 Healthcare Cost and Utilization Project National Inpatient Sample, this study characterizes sepsis hospitalizations among infants aged 0–3 years, examines comorbidity structure using network-based methods, and evaluates mortality prediction under severe class imbalance. Comorbidity networks constructed from ICD-10 diagnosis co-occurrence revealed dense diagnostic relationships, with renal, respiratory, metabolic, and neurologic conditions occupying structurally central positions independent of sepsis itself. Infants who died during hospitalization exhibited substantially higher diagnostic burden and longer lengths of stay than survivors. Random forest mortality prediction models demonstrated marked trade-offs between accuracy and sensitivity: a baseline model achieved high overall accuracy (0.935) but low sensitivity (0.350), while a model trained with class-balanced downsampling improved sensitivity to 0.700 at the cost of increased false positives; both models achieved strong and comparable ROC-AUC values (0.862 and 0.854) using administrative variables alone. Post-hoc SHAP analysis confirmed that diagnostic burden was the dominant predictor across both models, with the rank ordering of features stable regardless of sampling strategy, providing convergent evidence across descriptive, network, and predictive analyses. These findings highlight the value of network-informed representations of comorbidity and imbalance-aware modeling for identifying high-risk infants, and demonstrate a scalable, interpretable framework for studying infant sepsis risk using administrative data.
Keywords
infant sepsis; comorbidity network analysis; class imbalance; random forest; administrative data; HCUP National Inpatient Sample
Citation
Jones, L. S. (2026). A Comorbidity Network–Based Machine Learning Approach to Mortality Risk in Infant Sepsis. Industrial Engineering Undergraduate Honors Theses Retrieved from https://scholarworks.uark.edu/ineguht/105
