Predicting single-cycle cumulative live birth rate in POSEIDON Group 2 Patients: a prediction model based on machine learning

Abstract

Abstract Background Outcomes in patients with poor ovarian response (POR) have been less favorable and there is a need for improvement. The patient-oriented strategy encompassing individualized oocyte number (POSEIDON) criteria, proposed in 2016, are now widely accepted and used in clinical practice. POSEIDON Group 2 is considered as “Unexpected low response”, which is a challenge for clinicians. Currently, multiple reviews have retrospectively analysed the ART outcomes in the hyporesponsive populations of the POSEIDON Groups. However, no study has systematically examined the influencing factors specifically associated with the single-cycle cumulative live birth rate in POSEIDON Group 2. A prediction model was developed to predict the cumulative single-cycle live birth rate in POSEIDON Group 2 Patients. Methods A total of 565 assisted reproductive cycles from the low-response population of POSEIDON Group 2 were retrospectively analyzed from January 2018 to December 2021 at the center for Reproductive Medicine, Wuhan Union Hospital, Tongji Medical College. Cases were randomized 7:3 into two groups. Baseline levels were compared among the total, training and validation groups. A total of 26 variables were included and analyzed using the Least Absolute Shrinkage and Selection Operator (LASSO) regression with "lambda.min" as the screening criterion. To construct a predictive model of cumulative live birth rate, the selected variables were subjected to multivariate logistic regression. The predictive performance of the model was validated in the validation group. Results After randomization, 392 cases were assigned to the training group and 173 cases to the validation group. There were no statistical differences in baseline characteristics among the three groups. Seven variables were screened out by LASSO regression, including female age, assisted reproduction cycles, type of infertility, normal fertilization rate, blastocyst formation rate, number of frozen embryos, and whether fresh embryos were transferred. Furthermore, logistic regression was performed on these seven variables to construct a regression model, which had a ROC (Receiver Operating Characteristic) curve of 0.818 in the training group and 0.7971 in the validation group, with good predictive power and goodness-of-fit tests > 0.05 in both the training and validation groups. The model had an area under the ROC curve of 0.818 in the training group and 0.7971 in the validation group. The prediction efficiency was good, and the Goodness of fit test in both the training group and the validation group was > 0.05. Conclusions In this study, the prediction model constructed had good predictive performance with female age, normal fertilization rate, blastocyst formation rate, number of frozen embryos, and fresh embryo transfer. These factors work as independent predictors of single cycle cumulative live birth rate in patients with POSEIDON Group 2. Trial registration: This is a retrospective study, and the study was ethically approved by Tongji Medical College, Huazhong University of Science and Technology (HUST), Wuhan, China.