Comparison of predictive models built with different machine learning techniques using the example of predicting the outcome of assisted reproductive technologies

Advancements in machine learning (ML) have resulted in the development of numerous supporting software products for reproductive medicine. Predicting the performance of assisted reproductive technology (ART) using ML can be accomplished using different algorithms, depending on the type of data and specific task at hand.

Objective: This study aimed to compare the predictive ability of logistic regression, decision tree algorithm, and Random Forest in relation to the likelihood of pregnancy based on the clinical, anamnestic, and embryologic data of patients undergoing ART.

Materials and methods: This retrospective study included 854 married couples and analyzed clinical and laboratory data as well as parameters of the stimulated cycle in relation to the effectiveness of the ART program using three ML algorithms: logistic regression, decision tree, and Random Forest.

Results: The most accurate algorithm for predicting pregnancy rates in the ART program was the Random Forest model, which identified the significance of the following predictors: embryonic arrest, triggering of final oocyte maturation, number of embryos of excellent and average quality, duration of stimulation, infertility factor, body mass index, FSH, and AMH levels. The model confirmed the significance of the predictors determined in the previous stages of the study using a decision tree algorithm, including the presence/absence of a history of previous pregnancies, parameters of the stimulated cycle (number of MII oocytes), spermogram indicators on the day of the puncture, number of embryos of excellent and good quality, and quality of the embryo according to morphological evaluation criteria.

Conclusion: To enhance the prediction of ART effectiveness, this study suggests the need for better mathematical models with an integrated approach to solve the problem using a large sample of patients with various input data presented in a balanced volume. Additionally, this study suggests the inclusion of additional markers that determine ART effectiveness, thereby improving the accuracy of the software product.