P-656 Machine learning algorithm automatically manages and accurately predicts ovulation in natural frozen-thawed embryo transfer cycles.

Abstract

Abstract Study question Can an Artificial Intelligence (AI) algorithm automatically manage frozen-thawed embryo transfer (NC-FET) treatment cycles and give an accurate prediction of ovulation day. Summary answer An AI algorithm automatically managed and predicted the ovulation of NC-FET treatment cycles with 94.8% accuracy using an average of 3.01 test days. What is known already Today the preferred method for frozen embryo transfer is natural cycle based on ovulation detection. Currently, there is no software capable of managing the treatment cycle automatically and identifying the time of ovulation to support doctor decisions. The aim of this study is to develop a physician support AI software for determining ovulation time reliably with high accuracy. Study design, size, duration 2083 NC-FET cycles from September 2018 to June 2021 were used to develop the ovulation detection and treatment management algorithms. Each cycle had data from at least 2 visits including: hormonal levels (Estrogen/Progesterone/LH) and follicle sizes. The dataset was divided into a train set and two test sets. In the 1st test set ovulation was determined by experts’ opinions and the 2nd test set included cycles in which follicle rupture was documented in consecutive ultrasounds. Participants/materials, setting, methods Two algorithms were developed, an ovulation prediction algorithm based on an NGBoost model and a treatment management algorithm that used the model to determine if and when to call for a new test or declare the ovulation day. Both algorithms were jointly tuned to reach the highest success rate, defined as providing the correct day of ovulation using the available cycle data, with as few test days as possible. Main results and the role of chance On the first test set, which consisted of 176 cycles in which ovulation was determined through the majority decision of 2 independent experts and the attending physician, the treatment management algorithm required on average 3.01 tests to reach a prediction and successfully predicted the ovulation day in 94.8% of cycles. In the second test set, which consisted of 29 cycles in which ovulation was determined through the follicular rupture in two consecutive ultrasounds, only the ovulation prediction model was tested. To ensure that the model provides a reliable answer and does not rely solely on the follicular disappearances, examined cycles were tested twice: Once using the ovulation day without the day prior to it, and again using only the day prior to ovulation without the ovulation day itself. The algorithm accurately predicted ovulation in 28 out of 29 instances (96.6%) using the day of ovulation and in 28 out of 29 instances (96.6%) using the day before ovulation. Limitations, reasons for caution The main drawback is this being a retrospective study: while the algorithm was trained to maximize accuracy when it selects the test days, the dataset test days were selected by the attending physicians. Statistical methods were used to overcome this, however further prospective trials are needed to validate the results. Wider implications of the findings This is the first AI algorithm designed to automatically manage NC-FET IVF treatment cycles and predict ovulation. The high accuracy and low average tests count might improve treatment outcomes, reduce the patients’ life disruption, and allow physicians to spend less time monitoring their patients’ treatments. Trial registration number not applicable