A comprehensive model for predicting the probability of live birth prior to the start of progesterone during artificial FET cycles

Objective: To develop a comprehensive model for predicting the probability of live birth prior to the start of progesterone during artificial frozen embryo transfer (FET) cycles that can identify at-risk cycles and set expectations.

Materials and Methods: Historical, de-identified EMR data was collected from a single IVF clinic in the United States. Records were filtered for autologous cryo-synthetic frozen embryo transfers (FETs) resulting in 5,813 cycles from 4,133 patients between 2014-2021. For endometrial response we extracted measurements from the end of the proliferative phase, a common decision point. Parameters with high variance inflation factor were dropped for accurate interpretation of regression results. We developed a mixed effects logistic regression model using parameters from the embryo, patient, and endometrial response for the primary outcome of live birth. RESULTS: 13 parameters were found to be significant (P<0.01) with respect to live birth outcomes. The most important parameters positively associated with live birth were higher endometrial thickness (OR 4.22) and embryo ICM grade A (OR 2.00). Parameters negatively associated with live birth included BMI (OR 0.51) and number of previous failed FETs (OR 0.29). Calibration curves showed predicted probabilities closely matched observed live birth rates, with an expected calibration error of 0.028 on a 25% hold-out test set.

Results: 13 parameters were found to be significant (P<0.01) with respect to live birth outcomes. The most important parameters positively associated with live birth were higher endometrial thickness (OR 4.22) and embryo ICM grade A (OR 2.00). Parameters negatively associated with live birth included BMI (OR 0.51) and number of previous failed FETs (OR 0.29). Calibration curves showed predicted probabilities closely matched observed live birth rates, with an expected calibration error of 0.028 on a 25% hold-out test set.

Conclusions: We developed a comprehensive and well-calibrated model to predict live birth probabilities prior to the start of progesterone during artificial FET cycles. Future work will expand the dataset and develop additional tools for clinical decision support.

Impact Statement: Successful outcomes in artificial FET cycles are dependent upon parameters related to the patient, embryo, and endometrial response, and modeling all parameters together enables accurate predictions of live birth rate for identifying at-risk cycles and setting expectations.