Can methods of artificial intelligence aid in optimizing patient selection in patients undergoing intrauterine inseminations?

• Published in: Journal of assisted reproduction and genetics Vol. 38; no. 7; pp. 1665 - 1673
• Main Authors: Kozar, Nejc, Kovač, Vilma, Reljič, Milan
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.07.2021; Springer Nature B.V
• Subjects: Adult; Artificial Intelligence; Assisted Reproduction Technologies; Citric acid; Clomiphene; Clomiphene - therapeutic use; Decision Making, Computer-Assisted; Female; Fertility Agents, Female - therapeutic use; Gonadotropins; Gonadotropins - therapeutic use; Gynecology; Human Genetics; Humans; Insemination, Artificial - methods; Learning algorithms; Letrozole - therapeutic use; Machine learning; Male; Medicine; Medicine & Public Health; Neural Networks, Computer; Patient Selection; Patients; Pituitary (anterior); Prediction models; Pregnancy; Reproductive Medicine; Spermatozoa - cytology; Spermatozoa - physiology; Support Vector Machine; Random forest; Neural network; Intrauterine insemination; Artificial intelligence; Partial least squares
• DOI: 10.1007/s10815-021-02224-y
• ISSN: 1058-0468

Purpose AI and its machine learning algorithms have proven useful in several fields of medicine, including medically assisted reproduction. The purpose of the study was to construct several predictive models based on clinical data and select the best models to predict IUI procedure outcomes. Methods Clinical data (patient baseline characteristics, sperm quality, hormonal status, and cycle data) from 1029 IUI procedures performed in 413 couples stimulated by clomiphene citrate, letrozole, or gonadotropins were used to build several models to predict clinical pregnancy. The models included ANN, random forest, PLS, SVM, and linear models using the caret package in R. The models were evaluated using ROC analysis by means of random CV on test data. Results Out of the best performing models, the random forest model achieved an AUC of 0.66, a sensitivity of 0.432, and a specificity of 0.756. This performance was followed by the PLS model, which achieved a sensitivity of 0.459 and specificity of 0.734. The other models achieved significantly lower AUCs. When adjusting the predictive cutoff value, confusion matrices show that clinical pregnancy is twice as likely in the case of positive prediction. Conclusion Among the compared methods, the random forest and PLS models demonstrated superior performance in predicting the clinical outcome of IUI. With additional research and clinical validation, AI methods may be successfully used in improving patient selection and consequently lead to better clinical results.