Cross-National Analysis of IVF Registry Data

by Yuxiao Luo
Doctoral thesis
Date of Examination:2025-06-03
Date of issue:2025-05-08
Advisor:Prof. Dr. Wolfram-Hubertus Zimmermann
Referee:Prof. Dr. Wolfram-Hubertus Zimmermann
Referee:Dr. Gerd-Johannes Bauerschmitz
Persistent Address: http://dx.doi.org/10.53846/goediss-11254

 

 

Files in this item

Name:Embargo.pdf.pdf
Size:826.Kb
Format:PDF
Name:Luo_Yuxiao_Dissertation.pdf
Size:3.25Mb
Format:PDF

This file will be freely accessible after 2026-06-02.

Abstract

English

Approximately 15% (1 in 6) of couples worldwide face infertility issues, and assisted reproductive technology (ART), which is now adopted globally, has become a key method to address these challenges. Various in-vitro fertilization (IVF) registries across different countries and regions have provided various valuable data for research and clinical practice, which might also offer important insights for the development of artificial intelligence (AI) models. However, variations in terminology and definitions present difficulties and challenges to cross-country data comparison and the future application of AI. In addition, given that IVF clinical treatment is divided into 4 steps, different registries may focus on different steps, such as medical history and demographic information (step 1), while others place more emphasis on laboratory procedures related to ART treatment (step 2). These differences, both in quantity and content, further increase the complexity of unifying the data. This raises the questions: How can these terminology and definitions be unified, generating a standardized IVF glossary? What are the specific differences among different registries across countries and regions? And is there sufficient data or information available to support AI model development? We selected 7 IVF registries encompassing 4 continents: the Deutsches IVF-Register (DIR) in Germany, the Society for Assisted Reproductive Technology (SART) and the Centers for Disease Control and Prevention (CDC) in the United States, the Assisted Reproductive Technology Database (ANZARD) in Australia and New Zealand, the Chinese Society of Reproductive Medicine (CSRM) in China, the European IVF Monitoring Consortium (EIM), and the International Committee for Monitoring Assisted Reproductive Technologies (ICMART). We then systematically extracted and standardized the terminology (variables) and definitions included in each registry to ensure comparability. These variables were categorized into 4 steps according to the clinical treatment process: step 1 (patient properties before stimulation), step 2 (stimulation protocol and monitoring), step 3 (laboratory procedures), and step 4 (ART outcomes). We applied one-way ANOVA to compare the number and proportion of variables across the different steps to assess whether significant differences exist between them. We also used statistical models to fit the data and identify underlying patterns. Besides, we studied sub-grouping and combinations of variables, and, used network diagrams to visualize the relationships between variables to reveal the frequency and patterns of variable usage across the different registries. A standardized IVF glossary containing a total of 196 terms as variables and their definitions was generated from 7 selected IVF registries. The estimated maximum number of variables, derived through model fitting, was approximately 330, with DIR containing the largest number of variables, accounting for less than 1/3 of the estimated total (97 variables), and ICMART containing the fewest, less than 10% (24 variables). Summary 59 The 3 most frequently used variables were age, ET (embryo transfer), and CP (clinical pregnancy). The average proportion of variables in each step was as follows: step 1 (36%), step 2 (2%), step 3 (40%), and step 4 (21%). The proportion in step 2 differed significantly from the other steps (P < 0.05, one-way ANOVA). The average sub-group count for all variables was 6.91 while EIM with a value of 16 indicates a significant difference (z- score >2). The distribution of sub-grouped variables followed a generalized gamma distribution, however, not belong to any recognized distribution sub-family. And step 2 showed a significant difference from step 1 (P<0.05, one way ANOVA) in terms of the average occurrence of the variables. In the analysis of variable combinations, the range of combined correlated variables varied from 2 to 8. DIR had the most combinations (471) and significant difference (P < 0.05, one-way ANOVA), while ICMART had the fewest (43). Among these, combinations involving 3 and 4 variables were the most common, with 231 and 272 combinations, respectively. Additionally, we found that the frequency of variable usage, the overall connections between variables, and the number of unique connections for each variable followed Weibull distributions. Finally, we analyzed the most common variable pairs across the 7 registries, and the most frequent pairs were year- prospectivity (DIR), age-autologous (SART), age-autologous (CDC), autologous-fresh cycle (ANZARD), year-ET (CSRM), country-delivery (EIM), and autologous-delivery (ICMART). This study underscores the importance of standardizing IVF data, which is critical for enhancing cross-national comparability and enabling more comprehensive analyses. Significant differences in the quantity of variables and data collection across registries were identified, particularly in step 2 (involving stimulation protocols and monitoring), where the data are notably sparse and, in some cases, missing. This inadequacy fails to sufficiently reflect the complexity of clinical decisions and interventions in this stage. Furthermore, the number of variables within each registry, and across all registries, falls far short of the estimated maximum variable count of 330, derived through model fitting. Although the standardized data are not yet ready for direct application in AI models, they provide a valuable foundation for AI model development. We also propose the inclusion of key variables at different stages of treatment, such as lifestyle factors, medication details and response, embryo grading, and transfer procedures, to improve the comprehensiveness of the data, which also could link to the weights in AI model development. Our analysis further revealed that variable distributions conform to some interesting statistical distribution sub-families, highlighting the complexity and inherent randomness in the data and the network diagrams illustrate structural differences in data usage across registries, supporting future efforts in data integration and optimization. Finally, we propose a Generative Adversarial Network-based AI model to generate personalized stimulation protocols, facilitating the optimization of assisted reproductive treatments.
Keywords: IVF; IVF Registry Analysis; D.I.R.; SART; CDC; ANZARD; EIM; ICMART; CSRM; AI for IVF
Schlagwörter: IVF; IVF Registry Analysis; D.I.R.; SART; CDC; ANZARD; CSRM; EIM; ICMART; AI for IVF