Semester of Graduation

Summer 2026

Degree

Master of Science (MS)

Department

Department of Biological and Agricultural Engineering

Document Type

Thesis

Abstract

Aortic stenosis is the most prevalent valvular heart disease in North America. Its definitive treatments—surgical aortic valve replacement (SAVR) and transcatheter aortic valve replacement (TAVR)—carry a substantial burden of unplanned hospital readmission, which worsens patient outcomes, strains hospital capacity, and creates financial risk under value-based payment programs. Prior predictive work has largely focused on binary classification of 30-day readmission after TAVR. Such studies typically report predictor effects as odds or hazard ratios without validating patient-level probabilities, and neglect both the timing of readmission and the resources it consumes. To address these gaps, this thesis develops and empirically validates a unified, multi-horizon predictive framework for both SAVR and TAVR. Using one processing pipeline applied to the 2022 Nationwide Readmissions Database (NRD), it models readmission occurrence, readmission timing, post-discharge mortality, and resource utilization. Ridge-regularized logistic regression incorporating index-admission predictors after automatic removal of linearly dependent columns was fit for 1-, 3-, and 6-month readmission, achieving validation AUCs of approximately 0.63-0.68—near the upper range reported for administrative-data models—and producing calibrated individual-level probabilities with wider risk stratification than a replicated seven-predictor benchmark. Readmission timing was then modeled as a time-to-event process using a Weibull proportional hazards formulation; predicted cumulative readmission curves agreed with empirical estimates across the 1st through 99th percentiles of readmission hazard. Because the NRD does not record out-of-hospital death, long-horizon readmission probabilities are biased upward. To correct this, readmission and post-discharge mortality were modeled as competing Weibull processes anchored to external 1- and 5-year mortality estimates, which reduced this bias and reproduced published five-year readmission trends. Finally, regularized lognormal and negative binomial regressions of readmission cost and length of stay, conditional upon readmission, showed weak individual-level association with index-admission predictors (R² of 0.12-0.16 and 0.03-0.11, respectively), reflecting the limited clinical granularity of claims data; combined with the joint survival model, the cost model nonetheless yielded validated, mortality-adjusted forecasts of cumulative expected readmission cost. Collectively, these models translate routinely collected discharge data into validated patient-level risk trajectories and aggregate resource projections supporting transitional care targeting, capacity planning, and episode-based financial management.

Date

7-29-2026

Committee Chair

Rego, Bruno

LSU Acknowledgement

1

LSU Accessibility Acknowledgment

1

Download

Share

COinS