Machine learning predicts live-birth occurrence before in-vitro fertilization treatment

Area of Science:

• Reproductive Medicine
• Biomedical Data Science
• Artificial Intelligence in Healthcare

Background:

• In-vitro fertilization (IVF) addresses various infertility causes but lacks guaranteed success.
• Predicting IVF outcomes is challenging due to numerous complex factors.
• High costs and uncertain results make IVF a significant burden for patients.

Purpose of the Study:

• To employ Artificial Intelligence (AI) for predicting live-birth occurrence in IVF.
• To compare the efficacy of different AI algorithms for IVF success prediction.
• To focus on predictions using embryos from couples, excluding donor contributions.

Main Methods:

• Utilized a publicly available dataset from the Human Fertilisation and Embryology Authority (HFEA).
• Compared classical Machine Learning, deep learning, and ensemble AI algorithms.
• Trained models with and without feature selection, evaluating metrics like F1-score, precision, recall, and ROC AUC.

Main Results:

• The Random Forest model achieved the highest F1-score of 76.49% without feature selection.
• This model also demonstrated strong performance with 77% precision, 76% recall, and 84.60% ROC AUC.
• AI models successfully predicted pregnancy based on clinically relevant IVF parameters.

Conclusions:

• AI offers a promising tool to support decision-making in IVF.
• AI can aid in diagnosis, prognosis, and treatment planning for fertility.
• Accurate prediction of live-birth occurrence can help manage patient expectations and optimize treatment strategies.