Machine Learning-Based Modeling of Ovarian Response and the Quantitative Evaluation of Comprehensive Impact Features

Abstract

Appropriate ovarian responses to the controlled ovarian stimulation strategy is the premise for a good outcome of the in vitro fertilization cycle. With the booming of artificial intelligence, machine learning is becoming a popular and promising approach for tailoring a controlled ovarian stimulation strategy. Nowadays, most machine learning-based tailoring strategies aim to generally classify the controlled ovarian stimulation outcome, lacking the capacity to precisely predict the outcome and evaluate the impact features. Based on a clinical cohort composed of 1365 women and two machine learning methods of artificial neural network and supporting vector regression, a regression prediction model of the number of oocytes retrieved is trained, validated, and selected. Given the proposed model, an index called the normalized mean impact value is defined and calculated to reflect the importance of each impact feature. The proposed models can estimate the number of oocytes retrieved with high precision, with the regression coefficient being 0.882% and 89.84% of the instances having the prediction number ≤ 5. Among the impact features, the antral follicle count has the highest importance, followed by the E₂ level on the human chorionic gonadotropin day, the age, and the Anti-Müllerian hormone, with their normalized mean impact value > 0.3. Based on the proposed model, the prognostic results for ovarian response can be predicted, which enables scientific clinical decision support for the customized controlled ovarian stimulation strategies for women, and eventually helps yield better in vitro fertilization outcomes.

Keywords: clinical decision support; controlled ovarian stimulation; dosage of Gn; machine learning; number of oocytes retrieved.

Figure 1

The overall framework of our study. (A) Data-processing toward the research cohort with 1365 sets of clinical data; Pearson correlation analysis is conducted to identify the impact features with significant correlation value. (B) Based on the two types of methods, ANN and supporting vector regression (SVR), two prediction models of the number of oocytes retrieved are built; the model selected is based on the training and prediction results. (C) Quantitative evaluation of the impact features, and ranking the importance of them. (D) Clinical application of the proposed model.

Figure 2

The structure of the proposed ANN.

Figure 3

Prediction performance of the proposed models. (A) Regression coefficient R for the ANN-based PMORN; (B) regression coefficient R for the SVR-based PMORN. The horizontal and vertical axes, respectively, represent the actual value and the predicted value, which are denoted by the symbols Y and T, respectively.

Figure 4

Distribution of the prediction error for the proposed models. (A) prediction error for ANN-based PMORN; (B) prediction error for SVR-based PMORN. The horizontal axis and vertical axis are the prediction error and the corresponding ratio for all the instances, respectively.

Figure 5

ANN-based PMORN: ranking of the features according to the NMIV, and the comparison of NMIV with correlation coefficient.