P-428 Identifying key factors for achieving biochemical pregnancy in first autologous IVF cycles: an artificial intelligence-assisted approach Free

What key factors determine a positive biochemical pregnancy outcome in first autologous IVF cycles, as identified by machine learning (ML) algorithms?

Maternal age, AMH, BMI, number of day 3 embryos and total embryos transferred emerged as top predictors for biochemical pregnancy across several ML methods.

ML methods are powerful tools that discern patterns from multiple features to predict outcomes. Beyond their clear predictive capabilities, ML algorithms also offer valuable clinical insights by uncovering underlying data trends. These insights come from understanding the algorithms’ decision-making processes, revealing the rationale behind their predictions. While maternal age is known to influence IVF success, other pre-treatment and in-cycle factors also play a significant role. Understanding these factors can enhance patient management and guide clinical advice and interventions.

This multicenter retrospective study included data from 9,945 cycles performed between January 2010 and July 2021. Only first autologous cycles using fresh oocytes and donor or patient sperm (fresh or frozen) were included. Preimplantation genetic testing (PGT) cycles were excluded. Both pre-treatment and in-cycle variables were considered, with treatment outcome expressed as cumulative biochemical pregnancy. Cases and variables deemed to have poor data quality were excluded. The final dataset comprised 80 variables and 7350 cycles.

Five ML classifier algorithms were applied to predict cumulative biochemical pregnancy using pre-treatment and in-cycle variables. The models tested included logistic regression, support vector machine, random forests, gradient boosting and multilayer perceptron. Shapley values analysis was performed to explain the decision-making structures of all the algorithms tested, quantifying the impact of each included feature on outcome prediction.

Within our cohort, mean maternal age was 37.3±4.7 years, mean serum AMH was 2.4±2.1 ng/ml, mean BMI was 23.7±4.0 and mean antral follicle count (AFC) was 11.6± 7.6. Notably, 40.1% of cycles used donor sperm, while 33.8% of male patients were normozoospermic. The majority of cycles (90.8%) used antagonist stimulation protocols, yielding an average of 7.8±5.5 oocytes per patient. Fertilization rate was 70.6%, with a mean of 2.7±1.8 usable embryos obtained per cycle. Fresh embryo transfers accounted for 73% of cases, with an average of 1.5±0.9 transfers performed per cycle. The embryo utilization rate was 64.6% and 48.1% of cycles resulted in a biochemical pregnancy. Among the evaluated algorithms, the gradient boosting model demonstrated superior performance, with an accuracy of 70.6±1.4% and AUC of 77.3±1.7%. Shapley values across all five algorithms consistently revealed the negative correlation between maternal age and biochemical pregnancy rates. Among pre-treatment variables, AMH levels had a positive influence on outcomes, followed by BMI, AFC, and parity status. Regarding in-cycle variables, the number of embryos available at day 3 and the total number of embryos transferred, positively affected the chances of pregnancy. The number of mature oocytes was also an important in-cycle predictive feature.

his retrospective study spans an extended timeframe, suggesting potential variations in clinical and laboratory practices over the years. Therefore, caution is advised in generalizing the results to other patient populations or to contemporary data gathered using newer and enhanced clinical protocols.

The consistent identification of maternal age, ovarian reserve markers, and embryo quantity as primary predictors of pregnancy by the five AI models corroborates existing clinical understanding. This consistency emphasizes the strength and dependability of AI to reinforce and complement clinical knowledge with data-driven evidence.

not applicable