Live-Birth Prediction of Natural-Cycle

Background: Natural-cycle Design and Methods: The dataset contained 57,558 anonymized register patient records undergoing NC-IVF cycles from 2005 to 2016 filtered from 7bsp;60,732 records in the Human Fertilisation and Embryology Authority (HFEA) data. We selected matching records and features through data filtering and feature selection methods. Two groups of twelve machine learning models were trained and tested. Eight metrics, e.g., F1 score, Matthews correlation coefficient (MCC), the area under the receiver operating characteristic curve (AUC), etc., were computed to evaluate the performance of each model. Results: Two groups of twelve models were trained and tested. The artificial neural network (ANN) model performed the best in the machine learning group (F1 score, 70.87%; MCC, 50.37%; and AUC score, 0.7939). The LogitBoost model obtained the best scores in the ensemble learning group (F1 score, 70.57%; MCC, 50.75%; and AUC score, 0.7907). After the comparison between the two models, the LogitBoost model was recognized as an optimal one. Conclusion: In this study, NC-IVF-related datasets were extracted from the HFEA data, and a machine learning-based prediction model was successfully constructed through this largest NC-IVF dataset currently. This model is universal and stable, which can help clinicians predict the live-birth success rate of NC-IVF in advance before developing IVF treatment strategies and then choose the best benefit treatment strategy according to the patients' wishes. As "use less stimulation and back to natural condition" becomes more and more popular, this model is more meaningful in the decision-making assistance system for IVF.