Personalized prediction of live birth prior to the first in vitro fertilization treatment: a machine learning method

Abstract

Background: Infertility has become a global health issue with the number of couples seeking in vitro fertilization (IVF) worldwide continuing to rise. Some couples remain childless after several IVF cycles. Women undergoing IVF face greater risks and financial burden. A prediction model to predict the live birth chance prior to the first IVF treatment is needed in clinical practice for patients counselling and shaping expectations.

Methods: Clinical data of 7188 women who underwent their first IVF treatment at the Reproductive Medical Center of Shengjing Hospital of China Medical University during 2014-2018 were retrospectively collected. Machine-learning based models were developed on 70% of the dataset using pre-treatment variables, and prediction performances were evaluated on the remaining 30% using receiver operating characteristic (ROC) analysis and calibration plot. Nested cross-validation was used to make an unbiased estimate of the generalization performance of the machine learning algorithms.

Results: The XGBoost model achieved an area under the ROC curve of 0.73 on the validation dataset and showed the best calibration compared with other machine learning algorithms. Nested cross-validation resulted in an average accuracy score of 0.70 ± 0.003 for the XGBoost model.

Conclusions: A prediction model based on XGBoost was developed using age, AMH, BMI, duration of infertility, previous live birth, previous miscarriage, previous abortion and type of infertility as predictors. This study might be a promising step to provide personalized estimates of the cumulative live birth chance of the first complete IVF cycle before treatment.

Keywords: Cumulative live birth; IVF/ICSI; Machine learning; Prediction model.

Fig. 1

Cross-validated model performance of four machine learning algorithms on the training dataset. a Receiver operating characteristic curve plot. b Calibration plot. AUC indicates area under the curve. Shaded areas depict the standard deviation across different folds in a five-fold cross-validation

Fig. 2

Final model performances of four machine learning algorithms on the validation dataset. a Receiver operating characteristic curve plot. b Calibration plot. AUC indicates area under the curve

Fig. 3

Nested cross-validation. Nested cross-validation outcomes for 11 times

Fig. 4

Live birth prediction tool (https://lbprediction.herokuapp.com)