PREDICTING NUMBER OF MATURE OOCYTES AT RETRIEVAL: A MACHINE-LEARNING MODEL FOR PATIENT COUNSELING

Abstract

To develop a prediction algorithm utilizing both clinical baseline characteristics and cycle-specific parameters to accurately predict (on the day of trigger shot administration) the number of mature oocytes (MII) that will be obtained at oocyte retrieval. This was a model-prediction study based on a retrospective cohort of patients who underwent their first fresh autologous in vitro fertilization (IVF) cycle between the years 2016-2019. Only cycles that resulted in oocyte retrieval were included. A dataset consisting of 1009 first IVF cycles was randomly partitioned into a training group (60% of cycles, N= 606) and a testing group (N= 403). Patient- and cycle-specific characteristics (ex: patient age, anti-Mullerian hormone level, FSH dosage, etc.) were evaluated as model inputs. Three different prediction models were used for cross-validation training with a K-fold of 10, repeated 20 times. The best-performing model (random forest) was validated on the test dataset. Root mean square error (RMSE) was calculated to ascertain model accuracy. RStudio® was used for all statistical analyses. Baseline and cycle characteristics of the study population are shown in the table. The final model used patient age, day 3 FSH, AMH, FSH dosage, number of follicles >14mm on the day of trigger, and peak serum estradiol level. RMSE of the model was 5.0 oocytes, indicating that the model is able to predict the number of mature oocytes at retrieval ± 2.5 oocytes. The model was most accurate for patients with <14 follicles 14mm or more in diameter on day of trigger (RMSE of 3.8; prediction accuracy: ± 1.9 oocytes). These models are more accurate than our current method of estimation, which uses only the number of follicles >14mm on the day of trigger to predict number of MIIs (RMSE 6.0 for the entire cohort and 4.3 for patients with fewer than 14 follicles). We have developed a machine-learning model to predict the number of MIIs at oocyte retrieval on the day of trigger shot. The model has an accuracy of ± 1.9 oocytes when used for patients with fewer than 14 visible follicles >14mm on transvaginal ultrasound. Validation will be carried out on a prospective cohort of patients undergoing IVF.Tabled 1N1009Age (y)36.1 (4.3)Age Group< 35373 (37%)35-37222 (22%)38-40244 (24%)41-42113 (11%)> 4257 (6%)BMI (kg/m2)*26.1 (5.8)FSH (mIU/mLU)*7.6 (2.9)AMH (ng/ml)*3.1 (3.1)Infertility Diagnosis (%)Not infertile396 (39)DOR213 (21)Unexplained151 (15)PCOS103 (10)Tubal Factor84 (8)Other/Uterine43 (4)Endometriosis19 (2)Days of stimulation*10 (2)FSH Dosage (IU)*4000 (1700)Number of follicles >14mm on day of trigger*10 (4)Peak Estradiol (pg/mL)*2489 (1176)No. oocytes retrieved*13 (8)No. MII oocytes retrieved*10 (6)*mean (SD) Open table in a new tab