Date of Award

January 2026

Document Type

Open Access Thesis

Degree Name

Master of Public Health (MPH)

Department

School of Public Health

First Advisor

Heping Zhang

Abstract

Fresh versus frozen embryo transfer does not have a single answer across all IVF populations, which raises the possibility of individualized treatment selection. This thesis examined whether baseline patient characteristics could be used to develop individualized treatment rules for fresh versus frozen embryo transfer using pooled multicenter randomized trial data.

The analysis used individual-level data from four randomized controlled IVF trials conducted in China, with live birth as the primary outcome. Two primary models were evaluated: a prespecified clinically informed logistic regression model with selected spline terms and treatment interactions, and a neural-network S-learner. Performance was assessed using internal-external cross-validation, holding out one trial at a time. Individualized treatment recommendations were compared with two benchmark strategies, assigning all patients to fresh transfer or all patients to frozen transfer, using policy value under a prespecified reward function as the primary metric.

Neither individualized model consistently outperformed the benchmark strategies in the real data. The neural-network S-learner achieved the highest policy value in only one held-out trial, while the clinically informed logistic model did not achieve the highest policy value in any held-out trial. Arm-specific discrimination was modest for both models. Follow-up analyses showed meaningful differences across trials in observed live-birth rates and baseline covariate distributions, and trial membership was highly predictable from baseline covariates, indicating substantial cross-trial structural heterogeneity.

To investigate why individualized rule learning was unstable, simulation studies calibrated to the IVF dataset were conducted. Under idealized homogeneous conditions, individualized rules could outperform treat-all strategies once treatment-effect heterogeneity was strong enough. However, performance became less stable as between-trial heterogeneity was introduced. Trial-level baseline-risk shifts and covariate distribution shifts had only modest effects, whereas trial-level treatment-effect shifts caused a much larger decline in performance, especially for the neural-network learner. Across heterogeneous settings, the more structured logistic model was consistently more stable than the neural network.

Overall, these results suggest that the main barrier in this multicenter IVF setting was not simply limited model flexibility, but weak and non-transportable treatment-effect signal across trials. Between-trial treatment-effect heterogeneity appeared to be a more plausible explanation for model failure than baseline-risk or covariate shift alone.

Comments

This is an Open Access Thesis.

Download
Open Access

This Article is Open Access

Share

COinS