Machine-Learning Prediction of Pregnancy Outcomes After Assisted Reproductive Treatments

Infertility affects more than 15% of the global population, and ART are increasingly used to achieve pregnancy. The field continues to evolve rapidly with ongoing technological and clinical advances. Although these advances have improved ART, success rates continue to vary, contributing to significant psychological distress among patients due to uncertainty regarding outcomes.

The most intense stress typically occurs after the procedure, especially following a positive pregnancy test and before ultrasound confirmation. This period, which lasts at least 3-4 weeks, is characterized by considerable uncertainty, as neither patients nor healthcare providers have reliable tools to determine pregnancy viability, assess early development, or accurately predict pregnancy progression. An individualized assessment of the probability of live birth after a positive pregnancy test could provide clinically meaningful information during this period and potentially improve counseling and management of patient expectations.

Serum beta-hCG is one of the earliest measurable biomarkers following embryo implantation and has long been recognized as an important prognostic marker of early pregnancy outcome. Both the absolute concentration of beta-HCG and its subsequent rate of increase have been associated with pregnancy viability and live birth following embryo transfer.

Investigators plan to use a machine-learning system to collect and analyze retrospective clinical data. Beta-HCG values from this dataset were collected at very early stages of pregnancy (around aspiration days +14 and 16, and post-FET days 9 and 11).

Patients were advised to have blood drawn for beta-HCG measurement 14 days after oocyte aspiration or insemination, followed by a second measurement two days later. For FET, measurements were taken 9 days after blastocyst transfer (or the corresponding day for cleavage-stage embryos) and two days afterward. Due to scheduling constraints, such as weekends, actual measurements may have occurred one day earlier or later, or several days apart. The dataset included the exact dates of each blood draw, allowing precise calculation of the interval between measurements for analysis.

ART outcomes were primarily collected through direct patient reporting of events, including adverse pregnancy outcomes or live birth; for patients who did not report the outcomes, clinic staff conducted follow-up calls to ensure complete follow-up.

This study is an observational retrospective-prospective cohort study for development and validation of an AI (artificial intelligence) - based prognostic model predicting live birth after ART.

The retrospective phase will use existing clinical data from patients who underwent ART procedures at the BetaPlus Center between January 2017 and 31.12.2025.

The prospective phase will include consecutive patients undergoing ART after model development in both BetaPlus Center and Clinical University Hospital in Rijeka. These patients will be followed to assess pregnancy outcomes, enabling external validation of the model.

The study is observational in nature. No interventions, treatment modifications, or randomization will be performed as part of this research. Clinical management will follow standard operating procedures (SOP) independent of model predictions.