O-183 Forecast of Total and Mature Oocyte Number for Three Different Trigger Days using Machine Learning Prediction Models
A. Hourvitz,E. Maman,R. Hourvitz,M. Baum,S. Reuvenny,M. Youngster,A. Luz
DOI: https://doi.org/10.1093/humrep/dead093.224
IF: 6.1
2023-06-01
Human Reproduction
Abstract:Can an Artificial Intelligence (AI) algorithm accurately predict the number of total\mature (MII) oocytes retrieved during an antagonist protocol cycle for different trigger days.
AI algorithm can accurately predict the number of total and MII oocytes for three different trigger days in an antagonist protocol.
Two recent studies used machine learning prediction models for determining the number of MII oocytes retrieved during ovarian stimulation. These studies predicted MII with mean absolute error (MAE) 2.87-3.11 and R2 0.62-0.64. However, currently no studies show results for predicting total and MII oocytes two days in advance.
The data used for developing this algorithm consists of 9,622 antagonist protocol cycles performed in a large center serving over 50 physicians, between August 2017 and November 2022. This dataset contains a subset of 5,517 cycles with information about the maturity of each retrieved oocyte representing ICSI cycles and egg freezing (MII dataset).
Two sets of prediction models were developed using XGBoost, to estimate the number of total and MII oocytes, each with three individual predictors for different possible trigger days: today, tomorrow, in two days. Parameters used include estradiol and LH levels, follicles’ size, stimulation days, gonadotropin dosage, and patients’ characteristics. Accuracy of the model was evaluated by comparing the predictions to the actual values on 20% of the data that was reserved as a test set.
The accuracy of the models for the total number of oocytes evaluated using the entire test set, while the models for the number of MII oocytes evaluated using the MII dataset test.
The performance of the prediction models for the number of oocytes: (1) trigger today, R2=0.72, MAE=2.59 (2) trigger tomorrow, R2=0.66, MAE=2.95 (3) trigger in two days, R2=0.64, MAE=3.12.
Of the prediction models for the number of MII oocytes: (1) trigger today, R2=0.68, MAE=2.24 (2) trigger tomorrow, R2=0.62, MAE=2.53 (3) trigger in two days, R2=0.59, MAE=2.54.
Further evaluation was done on a refined subset, excluding cycles with a very large difference between the number of follicles in the last ultrasound and the number of oocytes retrieved (∼10%).
The performance of the prediction models for the number of oocytes: (1) trigger today, R2=0.81, MAE=2.14 (2) trigger tomorrow, R2=0.73, MAE=2.42 (3) trigger in two days, R2=0.69, MAE=2.77.
Of the prediction models for the number of MII oocytes: (1) trigger today, R2=0.74, MAE=2.03 (2) trigger tomorrow, R2=0.66, MAE=2.25 (3) trigger in two days, R2=0.62, MAE=2.38.
To ensure accurate predictions, it is important to only provide predictions on days when a trigger can be performed. Thus, an outlier detection mechanism is required to identify predictable days. This requirement is a limitation when using the prediction models for the number of total or MII oocytes in real-time.
Estimating the number of oocytes retrieved for different trigger days may assist in selecting the optimal trigger day and potentially improving outcomes. Additionally, if the estimated number of oocytes retrieved for different trigger days are similar, it may provide more flexibility in choosing the trigger day.
HMC-0011-22
Computer Science,Medicine