Targeted Learning of The Probability of Success of An In Vitro Fertilization Program Controlling for Time-dependent Confounders

Antoine Chambaz, Sherri Rose, Jean Bouyer, Mark J. van der LaanCollection 2012

Introduction

Infertility is a global public health challenge, with In Vitro Fertilization (IVF) serving as a primary treatment. However, assessing the cumulative success of an IVF program—typically consisting of multiple cycles—is statistically complex. Patients often drop out between cycles for various reasons, and clinical factors change over time. This paper introduces a double-robust semiparametric approach using Targeted Minimum Loss-based Estimation (TMLE) to provide a more accurate estimate of IVF success.

The Challenge of Time-Dependent Confounding

Standard statistical methods often fail to capture the nuances of multi-cycle IVF. Naive ratios (births divided by total participants) underestimate success by ignoring dropouts, while Kaplan-Meier analysis assumes dropouts are representative of the remaining population, which is rarely true.

The core difficulty lies in time-dependent confounding. In IVF, variables like the number of embryos available for transfer change at each cycle. These factors influence both the probability of success in subsequent cycles and the likelihood of a patient continuing the treatment. Traditional models that only look at baseline data (like age at the start) miss this dynamic "moving target," leading to biased results.

Methodology: TMLE and Super Learner

To address these biases, the authors employ Targeted Minimum Loss-based Estimation (TMLE). TMLE is a "double-robust" framework, meaning the final estimate remains valid if either the outcome model or the treatment/censoring model is correctly specified.

The process involves two key stages:

  1. Super Learner: Instead of relying on a single algorithm, the researchers use "Super Learner," an ensemble method that combines multiple machine learning models (e.g., Random Forests, Generalized Additive Models). It uses cross-validation to determine the optimal weight for each algorithm, ensuring the best possible initial prediction.
  2. Targeted Bias Reduction: Standard machine learning is optimized for overall prediction, not for estimating a specific parameter like "probability of success." TMLE "nudges" the initial ML estimate to minimize bias specifically for the parameter of interest, providing a more statistically sound result.

Data and Findings

The study utilized the DAIFI study dataset, comprising 3,001 women in France. Because the French national health system reimbursed the first four cycles during the study period, financial dropout was minimized, providing a clearer look at clinical outcomes.

Key Results: * 50% Success Rate: The TMLE analysis estimated a 50% probability of delivery (95% CI [0.48, 0.53]) for a program of up to four IVF cycles. * Methodological Superiority: In simulation studies with a known "ground truth," TMLE consistently outperformed standard estimators. The 50% finding is more conservative than Kaplan-Meier (52%) but significantly more optimistic than naive calculations (37%).

Conclusion

By integrating machine learning with causal inference frameworks, this research provides a robust tool for clinical assessment. Accounting for time-dependent confounders allows for a more realistic expectation of IVF success, suggesting that approximately half of the participants starting a four-cycle program can expect a successful delivery. Future work may focus on incorporating even more granular psychological or physiological variables to further refine these predictions.

Summary updated

Abstract

Infertility is a global public health issue and various treatments are available. In vitro fertilization (IVF) is an increasingly common treatment method, but accurately assessing the success of IVF programs has proven challenging since they consist of multiple cycles. We present a double robust semiparametric method that incorporates machine learning to estimate the probability of success (i.e., delivery resulting from embryo transfer) of a program of at most four IVF cycles in the French Devenir Apr`es Interruption de la FIV (DAIFI) study and several simulation studies, controlling for time-dependent confounders. We find that the probability of success in the DAIFI study is 50% (95% confidence interval [0.48, 0.53]), therefore approximately half of future participants in a program of at most four IVF cycles can expect a delivery resulting from embryo transfer.
BibTeX
@inproceedings{chambaz2012targeted,
  author = {Antoine Chambaz and Sherri Rose and Jean Bouyer and Mark J. van der Laan},
  title = {{Targeted Learning of The Probability of Success of An In Vitro Fertilization Program Controlling for Time-dependent Confounders}},
  booktitle = {Collection},
  year = {2012},
}
Figure