# The Influence of Pre-IVF Day 2 TSH Levels on Treatment Success and Obstetric Outcomes: A Retrospective Single-Center Analysis with Machine Learning-Based Data Evaluation

## TL;DR

This study examines how pre-IVF TSH levels affect IVF success and uses machine learning to predict outcomes like pregnancy and live birth.

## Contribution

The study evaluates TSH's impact on IVF outcomes within normal ranges and applies machine learning to identify key predictors.

## Key findings

- TSH levels within the normal range (0.3–4.0 mIU/L) did not significantly affect IVF success.
- Machine learning models like RF and XGBoost showed moderate predictive accuracy for pregnancy and live birth outcomes.
- Key predictors included embryo score, maternal age, BMI, and hormone levels, with male factors also playing a role.

## Abstract

Background: Thyroid disorders, particularly thyroid autoimmunity, are increasingly prevalent among women of reproductive age and have been linked to fertility outcomes. While current endocrinology guidelines define distinct thyroid-stimulating hormone (TSH) target values for women undergoing assisted reproductive technology (ART), the optimal preconception TSH range for in vitro fertilization (IVF) success remains a topic of debate. Objectives: This study aimed to assess the impact of baseline TSH levels within the recommended normal range on IVF outcomes, specifically clinical pregnancy and live birth rates. Additionally, we assessed the predictive value of procedural and preprocedural factors, including maternal body mass index (BMI) and TSH, using machine learning models. Methods: We conducted a retrospective, single-center cohort study at the Institute of Reproductive Medicine, University of Szeged, involving 996 women who underwent IVF, with or without intracytoplasmic sperm injection. Biometric, medical history, laboratory, and procedural factors were analyzed. Pregnancy and live birth predictions were modeled using support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGBoost) algorithms. The significance of features in the RF and XGBoost models was assessed. Results: SVM models achieved a mean accuracy of 72.26% in predicting pregnancy but were less effective for live birth classification. RF and XGBoost models demonstrated an area under the receiver operating characteristic curve of 0.76 and 0.74 for pregnancy and 0.67 and 0.61, respectively, for live birth. Key predictors included embryo score, maternal age, BMI, and specific hormone levels. Notably, male factors also contributed to outcome prediction. Analysis suggested that variations in maternal TSH within the normal range (0.3–4.0 mIU/L) had no significant impact on IVF success. Conclusions: Our study suggests that preconception TSH levels within the reference range do not significantly influence IVF success, which indirectly supports the validity of the current recommendations on this matter. While machine learning models demonstrated promising predictive performance, larger prospective studies are needed to refine thyroid function targets in ART, with a separate analysis of women with thyroid autoimmunity.